Home

Land use and cover classification using airborne MASTER and

1. Repeat the previous two steps for all the vector layers Once completed go to the Image Display menu bar and click Overlay gt Region of Interest In the ROI Tool windows which opens automatically you should see all the vector layers exported as ROI Adjust the class names and colors to your preference Click on Select All Then on the menu bar File gt Save ROIs gt Select All Items and choose a suitable name and path Click OK With the files still selected click on File gt Output ROIs to ASCH In the Select Input File for ROI data highlight the MASTER image file name Make 32 You can use both the raw and the atmospherically corrected MASTER image Chapter 10 Appendices 75 sure that all 25 bands are selected in the Spectral subset and then click OK In the Output ROIs to ASCII Parameters window select the first item form the list and choose a suitable name and path to save the txt file Click on Edit Output ASCII Form In the windows that opens untick Geo Location Increase the precision of Band Values to 10 Check the other attributes and precision When satisfied click OK Click OK in the Output ROIs to ASCII Parameters window Repeat the last 3 steps for all ROI layers of the list Go to the folder where the new txt files has been saved Copy and paste them Rename the copies with csv extension Accept the pop up window warning for possible changes in the file Op
2. Chapter 4 Land use and cover LUC classes 15 Now that the terms land cover use became familiar and these two officially recognized classification schemes were briefly introduced it is possible to speculate that the FAO distinction between cultivated managed areas and natural vegetation or the IGPB definition of cropland do involve some sort of land use knowledge However far from debating here this choice in this thesis we will keep on referring to their schemes as LCCS We prefer anyway to include the term use for our classification scheme and legend Because of the objectives and smaller scale of this study the global scale FAO and IGBP LCCSs had to be reduced or slightly modified The principles to design the LUC classification scheme were adhere as much as possible to the FAO and IGBP LCCSs be hierarchical in order to be easily implemented in Decision Trees and larger areas We wanted the resulting legend to be applicable at the given spatial resolution meaningful to fulfill the objectives thus irrelevant classes excluded site adapted thus unobserved classes excluded A visual representation of the adopted classification scheme and legend are presented in Fig 6 and Table 4 The distinction between tree shrub herbaceous life forms follows a simplified version of Raunkizr s classification Appendix B The aim was to shape artificial class delineations in a way that better resembles
3. ASTER GDEM Validation Team METI ERSDAC NASA LPDAAC USGS EROS 2009 ASTER Global DEM Validation Summary report NASA and METI Atkinson P 1996 Optimal sampling strategies for raster based geographical information systems Global Ecology and Biogeography Letters 5 271 280 Berk A Bernstein L Anderson G Acharya P Robertson D Chetwynd J Adler Golden S 1998 MODTRAN cloud and multiple scattering upgrades with application to AVIRIS Remote Sensing of Environment 65 3 367 375 Blomme G Eden Green S Mustaffa M Nwauzoma B Thangavelu R 2011 Major diseases of banana In Banana breeding progress and challenges ed M Pillay and A Tenkouano 85 120 Boca Raton FL CRC Press Taylor amp Francis Group Breiman L 2001 Random forests Machine learning 45 1 5 32 Butterfield R and Mariano E 1995 Screening trial of 14 tropical hardwoods with an emphasis on species native to Costa Rica fourth year results New Forests 9 2 135 145 Calle M and Urrea V 2011 Letter to the Editor Stability of Random Forest importance measures Briefings in bioinformatics 12 1 86 Canty M 2010 Image analysis classification and change detection in remote sensing With algorithms for ENVI IDL Second ed CRC Press Taylor amp Francis Group Canty M and Nielsen A 2008 Automatic radiometric normalization of multitemporal satellite imagery with the iteratively re weighted MAD transformation Remote Sensing of Environment 11
4. SVM classification Generate now reference data for SVM analysis by following the steps in chapter 2 of the Classification tutorial van der Linden e a 2010b You can play with the sampling methodology for reference pixels Random Equalized Proportional or Disproportional Stratified random as well as with the number of pixels to consider Note and appreciate that the pixels used for validation testing areas are different from the training set Follow now chapter 3 of the Classification Tutorial to parametrize and execute the SVM algorithm Chapter 4 will guide through a fast accuracy assessment evaluation Repeat the steps by changing the parameters if necessary Once you are satisfied proceed with Chapter 5 Apply SVM model to image Chapter 10 Appendices 79 10 9 Appendix I Field manual INTRODUCTION The purpose of this study is to develop an easy and economical methodology for land use and cover LUC classification using remote sensed imagery with particular focus on identification of banana coffee agroforestry systems The expected outcome of this pilot study is to provide useful background material to a team of scientists working on the mapping of a fungi Fusarium oxysporum cubense FOC that is severely affecting one important banana cultivar in Central America This research plan is part of a larger project named Mejorando la producci n y mercadeo de bananos en cafetales con rboles de peque os productores
5. Unfortunately measurements of atmospheric parameters are rarely associated with remote sensed data It is therefore necessary to retrieve their imprint on multispectral data through the data themselves and some other known parameters Atmospheric correction models are able to retrieve indirectly these information and use them to estimate the surface reflectance The atmospheric correction model used in this study is FLAASH Fast Line of sight Atmospheric Analysis of Spectral Hypercubes by ITT ENVI FLAASH estimates directly also the reflectance value Its prominent features are TT ENVI Atmospheric correction module Mattew e al 2003 wavelength correction from the visible to the SWIR short wave infrared region up to 3 um Chapter 5 Methodology 27 incorporation of MODTRAN4 radiation transfer code this avoids using a pre calculated database of modeling results correction of spectral mixing adjustable spectral polishing parameters correction of images taken from both nadir or slant positions The ancillary information that need to be retrieved from the Header file are the exact day and time of flight plane and mean ground elevation coordinates of the scene centre pixel size and some climatic and geographic descriptions tropical rural area etc For a step by step technical guide of the followed procedure see Appendix F After performing atmospheric correction with FLAASH the noisy bands will be a
6. Chapter 10 Appendices 77 10 8 Appendix H Feature selection and SVM with EnMAP Toolbox EnMAP Toolbox is a software developed by Umbold University of Berlin and DLR for image processing and analysis Suess e al 2010 Van den Linden e al 2010 a amp b Though the overall executing potentials are limited when compared to ITT ENVI or ERDAS IMAGINE some operations are extensively developed One example is the Support Vector Machine SVM algorithm which in EnMAP Toolbox allows to easily evaluate and modify some SVM parameters A short description of these parameters is provided along this appendix However for further details and investigations please refer to the EnMAP Box tutorials and manuals Data preparation Before starting with SVM classification you need to format your remote sensed image and reference areas in a way suitable for automatic parametrization of the SVM For doing this 1 Open your image in ENVI and from the main menu bar File gt Save file as gt ENVI Standard In this way two files should appear in the chosen directory It corresponds to the so called 4oa_060722 image of the En MAP Alpine_foreland tutorials If you have saved the sampling polygons as ROIs overlay them to the MASTER image opened into the Display window and jump to step 6 Otherwise from ENVI main menu bar choose File gt Open Vector File and browse to the vector files with sampling polygons Open all of them and eventuall
7. The geological origin of Costa Rica is rather recent being until 3 millions years ago just a volcanic archipelago Coates e al 1992 Specifically the Turrialba valley was formed by the Rio Reventaz n which eroded the substrate formed by the volcanoes Turrialba and Iraz Kass er a 1995 The predominant soil types are Eutric Cambisol and Umbric Andosol from alluvial and volcanic origin respectively Both substrates are 2 Nomenclature according to the World Reference Base for Soil Resources FAO 2006 In soil taxonomy they are called Andic Eutropept and Acrudoxic Melanudand respectively Eutric Cambisol is very deep moderately well to well drained dark brown gravelly clay loam with larger stones in the subsoil Umbric Andosol is very deep well drained dark brown clay soil usually located on slopes gt 25 The thick A and BC layers are influenced by volcanic ashes and are highly acidic Chapter 2 Study site 6 fertile and can be cropped easily with small addition of P and N Lime is sometimes required for Andosols Kass e al 1995 The study site altitude lays between 475 msl and 1225 msl with an average of 768 msl data obtained from DEM chapter 3 4 Slopes are pronounced in some areas though the average inclination is just 11 20 2 4 Land use The area has a good range of land cover types which span from urban areas to forest Agriculture is the leading activity and this provides several examples of
8. uusssssesesesesnenenennnennenennnnnennen 71 10 7 Appendix G Feature selection with Ru 2a 74 10 8 Appendix H Feature selection with EnMAP Toolbox neeseeeenenne 77 10 9 Appendix I Field manual anne ek 79 TA E S A A EE E Salon E ESE NEEE Bee awe E SE 82 List of Figures Fip MASTER eyes ee eier 7 Fig 2 Channels and wavelength regions of MODIS ASTER and MASTER 8 Fig 3 Study site over georectified MASTER Ale nase ee 10 Big Ar LOCS used byth FAO ze a ss ee eI a au 14 Fig 5 Legend used in the IGBP UCCOS anuess 14 Fig 6 LUC classification scheme applied in this study een 16 Fig 7 Digitized sampling polygon sales er 23 Fig 8 Example of confusion man ne 33 Fig 9 MASTER and GeoEye 1 images overlaid with OSM essen 36 Fig 10 Radiance and reflectance MASTER images a 36 Fig 11 Feature selection using R randomForest package uunnesnnenenseenennennn 37 Fig 12 Feature selection using EnMAP Toolbox sarah ee 38 Fig 13 ML and SVM classification maps ee 40 Fig 14 LUC class spectra of MASTER radiance at sensor image 42 Fig 15 LUC class spectra of MASTER reflectance image 42 Fig 16 Confusion matrix of sieved amp clumped MASTER 44 Fig 17 Examples of rule images resulting from ML classification 45 Fig 18 Confusion matrix of combined class MASTER 46 B18 19 Samplneern en R A ARa R E A Aas 80 List of Tables Table 1 Agroforestry extend estimates unse 2 Table
9. was initiated in March 2009 The project coordinated by Bioversity International and funded by GIZ Deutsche Gesellschaft f r 1 Improving small farm production and marketing of bananas under trees Chapter 1 Introduction 4 Internationale Zusammenarbeit has study sites in Costa Rica Honduras Nicaragua and Peru In the specific case of Costa Rica the partners can also benefit from the facilities and support of CATIE Centro Agron mico Tropical de Investigaci n y Ense anza research centre Within this international project the Chair of Forest Inventory and Remote Sensing University of Gottingen was contacted with the goal of mapping FOC distribution in the 4 study sites As small contribution this thesis aims at delineating an efficient methodology to identify coffee banana agroforestry systems using available remotely sensed images Though this has to be considered just a pilot study we hope that the blazed trails will serve as backbone and inspiration for further study developments in land cover and FOC mapping In detail the objectives of this thesis are e Find an effective methodology for agroforestry and LUC classification for the study site around Turrialba with the given resources e test the suitability of a MASTER scene at 10 m spatial resolution and 25 visible near infrared VNIR and short wave infrared SWIR spectral bands to classify agroforestry and other land uses e develop a simple and c
10. GEORG AUGUST UNIVERSIT T G TTINGEN 7 2 57 Land use and cover classification using airborne MASTER and spaceborne GeoEye 1 sensors Focus on coffee banana agroforestry systems near Turrialba Costa Rica by Marina Martignoni A thesis submitted in partial fulfillment for the degree of Master of Science in Sustainable forest and Nature management atthe Fakult t f r Forstwissenschaften und Wald kologie Chair of Forest inventory and Remote sensing Georg August Universit t G ettingen Germany September 2011 Supervisor Prof Dr Christoph Kleinn Co supervisor Prof Dr Martin Worbes Acknowledgments None of the achievements would have been possible without the constant supportive and inspiring guidance of my supervisor Dr Hans Fuchs His attention hard work prompt replies and proactive nature have set an example I hope to match some day I am deeply indebted for what I could learn and appreciate during this experience and I shall never forget his contribution for my development For the completion of this thesis special thanks to Prof Dr Christoph Kleinn for the opportunity enlightened expertise and inspiring advice Prof Dr Martin Worbes for having being an excellent teacher and correcting this thesis Dr Charles Staver for the granted financial contribution and chance to work for Bioversity GIZ for the funding Dr Lutz Fehrmann and Dr Yang Haijun for the statistical support and gui
11. In addition it has to be mentioned that the latter carries small overlapping errors occurred during mosaicing Thus the best available spatial reference were the OSM layers 30 and 40 m spatial offset between the MASTER and respectively the OSM layer and GeoEye 1 mosaic indicate that there is room for geospatial improvements On the other hand for the purposes of this study it is possible to accept this misregistration as the only images that were actually compared were the georeferenced MASTER image and the deriving thematic maps which obviously match exactly In future studies if land use changes have to be detected over time it is important to keep in mind the positional accuracy and matching of the analyzed maps For ideas how to improve it see chapter 9 Outlooks When considering larger study areas cloud cover and haze might require tile mosaicing It is possible to conduct this operation with ENVI however care should be taken when analyzing and processing the resulting mosaic In fact different illumination conditions can affect further processing outputs Therefore in mountainous terrain with steep slopes topographic normalization should be applied Also to create a more homogeneous image it is recommended to atmospherically correct the mosaic for example with the ENVI extension IR MAD Canty and Nielsen 2008 or with FLAASH model in ENVI with one manual method like Empirical Line Calibration or Flat Field and or by mat
12. To save the data chose File gt Save To gt File and select the desired folder path Here you can also chose the file extension e g SHP KML etc e Link digital photos to GPS points in vector file Open both vector layers with the digitized land cover polygons and with the recorded waypoints in Quantum GIS Open the Attribute Table of the waypoint layer Click on Toggle editing mode and then on New column Name the new column for example picture Type Text Width e g 30 Click again on New column and name it height Type Whole number Width 3 Click again on New column and name it compass Type Whole number Width 3 For each GPS point type in the relative path and picture number e g pictures 1937 jpg under the column picture In the column height you can type in the camera height and in compass the bearing Close the Toggle editing mode and save your changes Make sure that the layer with the GPS waypoints is selected in the legend window Click on Plugins gt eVis gt eVis Event ID tool Click on one of the waypoints displayed on the map In the tab Options Attribute containing path to file gt chose picture Tick also Path is relative and Remember this In the same tab Compass bearing gt compass Tick Display compass bearing In the same tab Base path gt chose the full path where you saved your pictures Tick Remember this Click Save Now by cl
13. species and structure For example coffee low Erythrina tall Laurel DATA UPLOADING At this point data collected in the field can be transferred to digital format There are mainly 2 sets of data that need to be uploaded the land cover information of each polygon and the GPS waypoints with associated pictures A GIS software such as ArcGIS or Quantum GIS is required to complete this step See Appendix C for procedure details EQUIPMENT GPS device reserve batteries compass measurement tape digital camera tripod writing board clip with pencils reserves suitable pens to write on the maps one printed copy of the field sampling chapter one printed copy of the classification key several copies of the field form one printed map of each quadrant with digitized polygons in scale suitable for writing printed overview maps of the area with the quadrant in scale suitable for navigating to the target quadrant if possible bring some maps of the area a satellite navigator would obviously help Chapter 11 References 82 11 References Alavalapati J Shrestha R Stainback G Matta J 2004 Agroforestry development An environmental economic perspective Agroforestry systems 61 1 299 310 Arnold G Fitzgerald M Grant P Platnick S Tsay S Myers J King M Green R Remer L 1996 MODIS Airborne Simulator radiometric calibration Paper presented at the meeting of the Proceedings of SPIE
14. 1 The approach used by Zomer e al 2009 appears to be the most comprehensive attempt to estimate global agroforestry cover Using 1 km resolution data they conclude that agroforestry involves 46 of all agricultural land One highlighted problem is that agroforestry systems are difficult to classify through remote sensing because they are composed by different plant species and have different vegetation structures which in turn leads to very different spectral reflectance values In spite of this their spatial importance cannot be neglected Chapter 1 Introduction 2 Minha Region Notes Reference 8 South and South East Asia Homegardens Kumar 2006 45 India Indonesia Mali Niger Different agroforestry types IAASTD 2008 C America Spain Portugal 45 Europe Silvoarable land Reisner e al 2007 235 USA Alleycropping silvo pasture Nair and Nair 2003 windbreaks and riparian buffers 823 worldwide Silvopasture 516 ha Nair et al 2009 agroforestry 307 ha 585 1215 Africa Asia Americas Agrosilvopastoral agroforestry Dixon 1995 1000 worldwide Tree cover gt 10 agricultural land Zomer ez al 2009 Table 1 Agroforestry extend estimates found in the literature for different regions of the globe Due to the conceptual frame of this thesis versus the spectral complexity of agroforestry the focus of this research is limited to coffee banana mixed cropping systems The main objective is
15. 13 2653 2662 Stehman S 1992 Comparison of systematic and random sampling for estimating the accuracy of maps generated from remotely sensed data Photogrammetric Engineering and Remote Sensing 58 9 1343 1350 Stehman S and Czaplewski R 1998 Design and Analysis for Thematic Map Accuracy Assessment Fundamental Principles Remote Sensing of Environment 64 3 331 344 Strahler A 1980 The use of prior probabilities in maximum likelihood classification of remotely sensed data Remote Sensing of Environment 10 2 135 163 Suess S van der Linden S Leit o PJ Rabe A Wirth F Okujeni A Hostert P 2010 En MAP Box Tutorial SVM regression Humboldt Universitat zu Berlin Germany Thacher T Lee D Schelhas J 1996 Farmer participation in reforestation incentive programs in Costa Rica Agroforestry Systems 35 3 269 289 Thomlinson J Bolstad P Cohen W 1999 Coordinating methodologies for scaling landcover classifications from site specific to global Steps toward validating global map products Remote Sensing of Environment 70 1 16 28 Tornquist C Hons FM Feagley SE Haggar J 1999 Agroforestry system effects on soil characteristics of the Sarapiqui region of Costa Rica Agriculture ecosystems amp environment 73 1 19 28 Tso B and Mather P 2009 Classification methods for remotely sensed data Second Boca Raton FL CRC Press Taylor amp Francis Group Unruh J and Lefebvre P 1995 A spatial database
16. 2 Spectral characteristics of MASTER channels 0 0 0 eeeeseseeeeeeeeeeeseeeeeeeeees 8 Table 3 MASTER sensor characteristics nun sen ke 9 Table 4 Legend of land use and cover LUC classes applied in this study 17 Table 5 Overall accuracies and Kappa coefficients obtained from Gaussian Mixture elassificati Nee a a a r erae A 39 Table 6 Overall accuracies and Kappa coefficients from ML and SVM classification 41 Table 7 Overall accuracies and Kappa coefficients of 3 post classification processed Nager 43 List of Abbreviations 6S Second Simulation of a Satellite Signal in the Solar Spectrum ACORN Atmospheric CORrection Now ASTER Advanced Spaceborne Thermal Emission and Reflection Radiometer ATREM ATmospheric REMoval CARTA Costa Rican Airborne Research and Technology Application CATIE Centro Agron mico Tropical de Investigaci n y Ensefianza CENAT Centro Nacional de Alta Tecnologia DEM Digital Elevation Model DLR Deutsches Zentrum ftir Luft und Raumfahrt DN Digital Number ERSDAC Earth Remote Sensing Data Analysis Center FAO Food and Agriculture Organization FLAASH Fast Line of sight Atmospheric Analysis of Spectral Hypercubes FOC Fusarium oxysporum formae specialis f sp cubense GDEM Global Digital Elevation Model GIS Geographic Information System GIZ Gesellschaft f r Internationale Zusammenarbeit GSD Ground Sample Distance ICRAF International Centre for Resear
17. 6 3 identify the quadrant which lies on the 6 row and on the 3 column Go to the field navigate the area and find the selected quadrants with the use of the GPS device and the overview maps again remember to take a print with you Once reached the target quadrant make a rapid survey to verify that all drawn polygons correspond to the real boundaries If necessary change the shape of the polygons by manually drawing the new version on the quadrant map Chapter 10 Appendices 81 5 Assign a number to each land cover polygon The suggestion is to assign a decimal number where the whole part corresponds to the class key number and the decimal part to the sequence number For example assign the number 5 3 to the 3 shrub plantation class key 5 that you encounter in that quadrant Write the number on the quadrant map with digitized polygons This will be the record for later land cover data comparison 6 For all accessible or visible polygons take a picture Use a tripod to maintain the camera in straight position Automatic camera settings should be used for all photos Try to capture a good overview of the area you want to photograph but do not alter the camera inclination 7 Record the picture number camera height compass bearing and GPS point of the location where the photo is taken from on the field form Do not forget to write the box and the polygon number on each page If possible add some notes about the vegetation type the
18. Utilizaci n de los recursos salud de los suelos selecci n de cultivares y estrategias de mercado which runs in Costa Rica Honduras Nicaragua and Peru Due to time and budget restrictions the procedures proposed in this manual are minimized to an effective workload The chapter on SAMPLING DESIGN gives a brief overview on the methodology used for indirect land cover classification and preliminary work FIELD MESUREMENTS and DATA UPLOADING instead provide a complete description of the work to be carried out in the field and the successive data handling SAMPLING DESIGN This section summarizes the procedure followed to select and pre classify the land cover type within some sample quadrants inside our area of interest before the actual field sampling 1 Selection of an area 7 x 7 km around Turrialba CATIE and San Juan Norte This area of interest is part of Site 1 of the project mapping FOC 2 Drawing of square grid over the selected area gt primary quadrants of 200 x 200 m each 3 Systematic selection of 36 quadrants distance between selected quadrants 1200 m The first quadrant has been chosen randomly by the software the other 35 accordingly Fig 19 4 Preliminary classification of the land cover within these 36 quadrants With the help of the underlying GeoEye 1 mosaic and MASTER airborne image the borders of each LUC polygon are digitized with a suitable software Appendix C The result is a SHP vector layer of poly
19. accuracy When ENVI merges two classes it operates on the classified image not on the ROIs To calculate the confusion matrix of the image with combined classes is therefore necessary to merge the ROIs Note that a new ML classification based on the merged ROIs would output a different classification map from the combined class image not only in accuracy terms but also in pixel class attribution No unclassified pixels would result taking this approach since no sieving is contemplated Chapter 8 Conclusion 57 8 Conclusion The aim of this work was to test the potentials of LUC classification using MASTER airborne imagery Special attention was paid to the class coffee banana agroforestry systems The results are very promising with 77 overall accuracy and 86 producer s and 71 user s accuracy for the shade coffee agroforestry class As sampling frame for reference data collection a systematic square grid was applied all land within the selected sampling quadrants was manually digitized by visual interpretation using the MASTER and high spatial resolution GeoEye 1 image published in GoogleEarth The classification key discussed and described in the legend could be applied to all sampled plots However it was realized that defining the size of the areas to label is crucial The sample size for this pilot study is discussed and referred to further analysis The data collection methodology proposed is rather economical in terms of resour
20. approach for estimating areas suitable for agroforestry in subSaharan Africa aggregation and use of agroforestry case studies Agroforestry systems 32 1 81 96 Van Asten P Wairegi L Mukasa D Uringi N 2011 Agronomic and economic benefits of coffee banana intercropping in Uganda s smallholder farming systems Agricultural Systems 104 4 326 334 Van der Linden S Rabe A Wirth F Suess S Okujeni A Hostert P 2010a imageSVM Classification Manual for Application imageSVM 2 1 Humboldt Universitat zu Berlin Germany Van der Linden S Wirth F Leit o P Rabe A Suess S Okujeni A Hostert P 2010b En MAP Box Application Tutorial SVM Classification Humboldt Universitat zu Berlin Germany Chapter 11 References 91 Van der Meer F 1999 Geostatistical approaches for image classification and assessment of uncertainty in geologic processing In Advances in remote sensing and GIS analysis ed P Atkinson and N Tate 147 166 Chichester John Wiley and Sons Van Genderen J Lock B Vass P 1978 Remote sensing statistical testing of thematic map accuracy Remote Sensing of Environment 7 1 3 14 Vermote E Tanr D Deuze J Herman M Morcette J 1997 Second simulation of the satellite signal in the solar spectrum 6S An overview Geoscience and Remote Sensing IEEE Transactions on 35 3 675 686 Walter V 2004 Object based classification of remote sensing data for change detection ISPRS Journal of Photogrammetry and Re
21. cells in between represent the number of pixels assigned to each category An example is illustrated in Fig 8 Input classes from training set OA diagonal total 1 3 44 15 le oa es P 4 7 column i 8 1 82 U i i X rowi 41 42 f 0A P xU i 1 6 1 6 2 639 GA Output classes from classification 5 PXU i 1 where 7 is a class number Fig 8 Confusion matrix of classification at 5 classes Cells are indicated as row number column number P and U stand for Producer s and User s accuracy with the respective class number OA for Overall Accuracy K is the Kappa coefficient The overall accuracy is given by the diagonal sum correctly classified pixels divided by the sum of all pixels Other accuracy indexes can be retrieved from the confusion matrix such as the producer s and the user s accuracy per each class These are respectively the ratio of pixels correctly assigned to one class over the total number of pixels that should be assigned to it column sum and the ratio of pixels correctly assigned to one class over Chapter 5 Methodology 34 the total number of pixels that were assigned to it row sum Tso and Mather 2009 Last but not least from the confusion matrix it is possible to calculate the Kappa coefficient Contrary to the producer s and user s accuracy the Kappa coefficient K takes into consideration all the informa
22. correction of spectral imagery Evaluation of the FLAASH algorithm with AVIRIS data Paper presented at the meeting of the Proceedings of SPIE McRoberts RE and Meneguzzo DA 2007 Estimating tree species richness from forest inventory plot data Notes from the 2005 Proceedings of the Seventh Annual Forest Inventory and Analysis Symposium 275 279 Miller R and Nair P 2006 Indigenous agroforestry systems in Amazonia from prehistory to today Agroforestry systems 66 2 151 164 Montagnini F 2006 Homegardens of Mesoamerica biodiversity food security and nutrient management Springer Mooney P Corcoran P Winstanley A 2010 Towards quality metrics for OpenStreetMap Paper presented at the meeting of the Proceedings of the 18th SIGSPATIAL International Conference on Advances in Geographic Information Systems Nair P Kumar B Nair V 2009 Agroforestry as a strategy for carbon sequestration Journal of Plant Nutrition and Soil Science 172 1 10 23 Nair PKR 1993 An introduction to agroforestry Klgwer Academic Publishers Nair PKR 1985 Classification of agroforestry systems Agroforestry systems 3 2 97 128 Nair PKR and Nair VD 2003 Carbon Storage in North American Agroforestry Systems In Kimble J Heath LS Birdsey RA and Lal R eds The Potential of U S Forest Soils to Sequester Carbon and Mitigate the Greenhouse Effect Boca Raton FL CRC Press LLC NASA 2008 MASTER Jet Propulsion Laboratory http masterw
23. cover extensive areas The vegetation can be present in parton either salt brackish or fresh water Mainly 12 Cropland Lands covered with temporary crops followed by harvest and a bare soil aunty period e g single and multiple cropping systems Note that perennial non ve getate di woody crops will be classified as the appropriate forest or shrub land Artificial water cover type bodies snow 13 Urban and Built up Land covered by buildings and other man made structures Note that this amp ice class will not be mapped from the AVHRR imagery but will be developed Aquatic or ne T populated places layer that is part of the Digital Chart of the orld regularly flooded x 14 Cropland Natural Lands with a mosaic of croplands forest shrublands and grasslands in Natural water Vegetation Mosaics which no one component comprises more than 60 of the landscape bodics snow 1s Snow and Ice Lands under snow and or ice cover throughout the year amp ice 16 Barren Lands exposed soil sand rocks or snow and never has more than 10 vegetated cover during any time of the year 17 Water Bodies Oceans seas lakes reservoirs and rivers Can be either fresh or salt water bodies Fig 4 LCCS used by the FAO The scheme resembles Fig 5 Legend used in the IGBP LCCS The classes are the hierarchical order The resulting legend depends on Inon hierarchically defined FRA 2000a the accomplished detail level DiGregorio amp Jansen 1998
24. cropping land use for spectral detection Little to none original vegetation is left due to slash and burn activities practiced since 3000 years Maize cultivated under shifting rotation constituted the staple crop of indigenous populations wheat sugar cane plantains and cattle were introduced after the Spanish reached Cartago in 1563 In the 19 century coffee became the dominant crop and together with sugar cane is still at the top of Turrialba agricultural production Kass et al 1995 Coffee is usually grown under shade trees like laurel Cordia alliodora and or por Erythrina poeppigiana and mixed with Musa spp Laurel is native to Costa Rica and its wood is highly appreciated on the market Butterfield and Mariano 1995 Por6 has been introduced in the 19 20 century from South America and is mainly used to supplement nitrogen fixation in the soil Russo and Budowski 1986 These mixed agroforestry systems are reported in Costa Rica since the early 20 century Cook 1901 although we would like to stress the fact that agroforestry system as defined above are virtually impossible to date precisely back in time Miller and Nair 2006 Chapter 3 Material 3 Material 3 1 MASTER images 3 1 1 Instrument and data description The data used in this study to perform the LUC classification were acquired by the MASTER imaging sensor Fig 1 MASTER is the MODIS ASTER airborne simulator developed from the joint effort of
25. data Photogrammetric engineering and remote sensing USA 54 593 600 Congalton R and Green K 1999 Assessing the accuracy of remotely sensed data CRC Press Taylor amp Francis Group Cook OF 1901 Shade in coffee culture Issue 25 of bulletin U S Dept of Agriculture Division of Botany Curran P 1988 The semivariogram in remote sensing an introduction Remote Sensing of Environment 24 3 493 507 Dahlquist R Whelan M Winowiecki L Polidoro B Candela S Harvey C Wulfhorst J McDaniel P Bosque P rez N 2007 Incorporating livelihoods in biodiversity conservation a case study of cacao agroforestry systems in Talamanca Costa Rica Biodiversity and conservation 16 8 2311 2333 Di Gregorio A and Jansen LJM 1998 Land cover classification system LCCS classification concepts and user manual FAO Publications Dixon R 1995 Agroforestry systems sources of sinks of greenhouse gases Agroforestry systems 31 2 99 116 Chapter 11 References 84 Efron B and Tibshirani R 1986 Bootstrap methods for standard errors confidence intervals and other measures of statistical accuracy Statistical science 54 75 Ellenberg H Mueller Dombois D 1967 A key to Raunkiaer plant life forms with revised subdivisions Ber geobot Inst eidg tech Hochschule Rubel 37 56 73 Erbek S Ozkan C Taberner M 2004 Comparison of maximum likelihood classification method with supervised artificial neural network algorit
26. list of classes resulting from the classification of a specific area using a defined mapping scale and data set Di Gregorio and Jansen 1998 All classes should be mutually exclusive and total exhaustive Congalton 1991 Both the FAO and IGPB use discrete classes to arrange continuous variables such as vegetation cover in nominal scale Di Gregorio and Jansen 1998 which are suitable for discussion and comparison There are however also examples of legend based on continuous gradients such as the MODIS Vegetation Continuous Fields VCF or AVHRR CF datasets Schwarz and Zimmermann 2005 Zomer e al 2009 which are arguably closer to reality but can be 8 FAO Food and Agriculture Organization 9 IGBP International Geosphere Biosphere Programme Their land cover legend and map Loveland et al 2000 are one of the eight used for the Global Land Cover Characteristics GLCC database generated by the U S Geological Survey s USGS National Center for Earth Resources Observation and Science EROS the University of Nebraska Lincoln UNL and the Joint Research Center of the European Commission 10 Each pixel should be classified unambiguously as belonging either to one or to the other class 11 All pixels of the study area should fall into one of the classes Chapter 4 Land use and cover LUC classes 14 applied only to few LUC classes at a time The mapping legend used by the FAO Forestry department FRA 2010 is prese
27. name to each ASCH file created in a by typing gt name lt read csv full path and name of the csv file header TRUE sep e g gt classl lt read csv c Users Marina Documents Thesis MASTER_CSV_classe s class 1 header TRUE sep Create an internal new column by typing gt name column name lt as factor land use number e g gt classl class key lt as integer 1 Make sure you repeat the last step for all classes Combine all these internal files using gt new name lt rbind name filel name file 2 w Convert the newly created data frame into a ranked data frame by gt new name column name lt as factor new name column name e g bind all class key lt as factor bind all class_key c Create a plot which highlights the band importance in determining the land use class gt set seed 4543 Create a random forest gt new name rf lt randomForest column name columnl column2 data new name ntree 1000 keep forest FALSE importance TRUE e g gt bind_all rf lt randomForest class key Bl B2 B3 BA B5 B6 B7 B8 B9 B10 B11 B12 B13 B14 B15 B16 B17 B18 B19 B20 B21 B22 data bind all ntree 1000 keep forest FALSE importance TRUE Display two plots calculated with two different accuracy indices gt varImpPlot new _name rf e g gt varImpPlot bind_all rf
28. reflectance image The used parameters were initial number of classes 11 min threshold 0 max threshold 2 The output clusters were visually analyzed and confronted with the true color composite MASTER image and other thematic from supervised classification maps to facilitate interpretation a land use label was assigned to the clusters which in some cases were merged The results were then tested in a confusion matrix against the reference ROIs created in 5 1 3 Note that only the identified LUC classes could be tested in the confusion matrix this mean that each time a different number of test pixels was considered The overall accuracies of these 4 Gaussian Mixture classifications and the relative Kappa coefficients are listed in Table 5 Image No bands No classes No tested pixels OA Kappa coeff PCA 4 8 13339 46 4 0 3732 MNF 6 8 12951 54 2 0 4578 Radiance at sensor 5 7 12728 53 5 0 4111 Reflectance 5 9 13562 42 8 0 3210 Table 5 Overall accuracies OA and Kappa coefficients obtained from confusion matrices of 4 images classified with the Canty s modified version of Gaussian Mixture algorithm No classes indicates the total number of classes after merging and labeling tested in the confusion matrix No tested pixels indicates the total number of pixels from the reference ROIs which were used in the comparison Initial parameters number of classes 11 min threshold 0 max threshold 2 20 The algorithm can
29. taken into account In our case though the gap was 6 years the land use management had not changed much Only one site consisting of 2 digitized plots was converted from presumed pasture to tree plantation and was thus not possible to validate In all other instances it was mainly a seasonal change rather than a change in land use e g sugar cane fields were harvested when in the MASTER image they resulted still green or viceversa In such cases it is important to adhere to the image interpretation as it is the image which will be processed When there is the need of assessing the usefulness of data collection digitizing land use polygons before going to the field could provide a valuable basis for comparison with the field data Considering the proportion of changed land use class about 1 out of 5 and polygon shape 1 out of 12 it appears clear the need for field sampling in this study 7 3 Image processing It is very hard if not virtually impossible to achieve 100 geospatial accuracy in a map Even the used OpenStreetMap OSM layers which were taken as spatial reference might contain geometric inaccuracies These can be due to the GPS receiver positional error and or low density of recorded points However we assume these errors to be Chapter 7 Discussion 51 limited as compared to automatized georectification of large images such as the MASTER tile or manually georeferenced images as in the case of the GeoEye 1 mosaic
30. the NASA Ames Research Center the Jet Propulsion Laboratory and the EROS Data Center The main objectives of this cooperation were the elaboration of algorithms calibration and validation of data for MODIS and ASTER spaceborne sensors flying on the Terra platform NASA 2008 Fig 1 MASTER camera Image from http asapdata arc nasa gov Image htm A full description of the optical and electronic systems is provided by Hook e al 2001 MASTER scanning radiometer records radiance data in 50 spectral bands from the VNIR visible near infrared to the TIR thermal infrared regions of the electromagnetic spectrum Table 2 The channel position has been designed to simulate at best both ASTER and MODIS measurements Fig 2 The ground spatial resolution depends on the terrain surface and on the flying altitude of the aircraft For MASTER it normally lies between 5 and 50 m Table 3 3 MODIS Moderate Resolution Imaging Spectroradiometer 4 ASTER Advanced Spaceborne Thermal Emission and Reflection Radiometer 5 Stored at 16 bits resolution Chapter 3 Material Full width Full width Channel half maximum Channel center Channel peak Channel half maximum Channel center Channel peak 1 0 0433 0 4574 0 458 26 0 1559 3 1477 3 142 2 0 0426 0 4981 0 496 27 0 1459 3 2992 3 292 3 0 0427 0 54 0 538 28 0 1478 3 4538 3 452 4 0 0407 0 5807 0 58 29 0 1544 3 6088 3 607 5 0 0585 0 6599 0 652 30 0 1345 3 7507 3
31. the aircraft and retrieved from the MASTER Header File Chapter 3 Material 11 visual interpretation we do not expect major problems as rarely evapotranspiration gt precipitation for a period long enough for plants to dry out in a tropical rainforest region Eva e al 2004 some underestimations of vegetation cover should be accounted 3 2 GeoEye 1 images GeoEye 1 images were used to complement LUC visual interpretation and facilitate orientation in the field GeoEye 1 is a satellite launched in orbit on September 8 2008 by GeoEye It is designed and equipped with very advanced technology which allows the highest resolution among commercial satellites It records spectral data in one panchromatic band ground sampling distance GSD 0 41m one near infrared and three visual RGB bands GSD 1 65m The study site around Turrialba belongs to a scene acquired on February 5 2010 order identifier 20100205_16072001603031604257 This image is available for private and unpublished works without concession through the GoogleEarth platform as long as any reproduction carries the disclaimer attributes Alternatively the raw image in 5 bands can be purchased through the GeoEye website http www geoeye com GeoFuse gt Advanced search gt Upload file gt Open Permalink As the purposes of this thesis did not involve any GeoEye image analysis simple screen shots were taken from GoogleEarth Only at the very end a standard pr
32. the georeferenced scene you first need to overlay the vector file In the Available Band List right click on the Georeferenced image and choose Load True Color In the Display Image window menu bar select Overlay gt Vectors In the Vector Parameters Cursor Query window click on File gt Open Vector File In the Select Vector Filenames window browse to the vector file defining the study site Make sure that the search file type is on shp if you saved your vector file in such format Select the file and click OK In the Import Vector Files Parameters window evaluate the Native File Projection Parameters and click OK when you are satisfied Visually evaluate the overlaying vector file of the study area 29 SGL Super Geometry Lookup table Chapter 10 Appendices 70 Scene subsetting In the Vector Parameter window click on File gt Export Active Layer to ROT In the Export EVF Layers to ROI window choose Convert all records of an EVF layer to one ROI and click OK In the Spatial Subset via ROI Action window select the geoeferenced image produced in d In the Spatial Subset via ROI Parameters window select the ROI file name corresponding to the study area vector file Select Mask Pixels Outside ROI Enter the Subset Output file name and click OK 30 ROI Region Of Interest Chapter 10 Appendices 71 10 6 Appendix F Atmospherically correct MASTER image with FLAASH ENVI Module T
33. the right rule image of class 3 shade coffee at 2 strata Dark pixels indicate higher probability of class belonging The class matches with less than 74 8 overall accuracy value are tree crown cover and mixed tree grassland Tree crown cover is often confused with shade coffee at 23 strata 11 shade coffee at 2 strata 8 and mixed tree grassland 7 Because the aim is to distinguish these categories the class tree crown cover was left untouched Mixed tree grassland is misclassified with shade coffee at 2 strata 12 tree crown cover 8 sugar cane 8 and building 6 Since these classes have little management use in common they were preserved too To reduce the number of classes the two shade coffee classes were merged together into one class called shade coffee and the class paved road was merged with building into a new category called settlement The comparison between the combined class image and the merged ROIs gives overall accuracy of 76 7 and Kappa coefficient of 0 7127 For specific class accuracy in percentage refer to Fig 18 The class shade coffee was classified respectively with 86 24 See Discussion 7 4 for details Chapter 6 Results 46 and 71 producer s and user s accuracy As mentioned before some pixels 1 4 were left unclassified after the sieving step Unclassified tree crown co barren land settlement water sh coffee sun coffee sugar cane grassland mixed
34. to classify land use and cover LUC paying special attention to agroforestry We will use imagery from MASTER the airborne MODIS ASTER simulator developed by NASA for calibration of the spaceborne MODIS and ASTER satellite sensors Hook e al 2001 In addition a high spatial resolution GeoEye 1 satellite scene will serve as complement for calibration validation visual interpretation and orientation in the field As the MASTER images have a spatial resolution of approx 10m and 25 bands in the visible near and short wave infrared VNIR and SWIR spectral range of the electromagnetic solar spectrum a second objective is to test whether a medium spatial resolution complemented with high spectral resolution imaging sensor is capable to identify agroforestry systems In fact whereas high spatial resolution sensors are very powerful in identifying land cover and single objects on the ground low and medium sensors may potentially allow best inference on land use as they take into Chapter 1 Introduction 3 account spectral information from a larger surface Clevers e al 2004 Franklin and Wulder 2002 The pilot study site is located near Turrialba Costa Rica Coffee Coffea spp and bananas Musa spp are respectively the first and the third cash crops in the country for land use and production value FAO 2003 In addition bananas and plantains represent important stable food for locals and are commonly integrated in home garden
35. whole image note it can be that the final classification of some training pixels differs from the original assigned class e the confusion matrix is calculated by taking pixels from these training areas and compare them with the final classification Whereas Canty s variants e separate the pixels of all training areas into training and testing sets in proportion 2 1 e take the average band values of the training set to classify the whole image e the confusion matrix is calculated by comparing the testing pixels with the final classification Clearly the advantage of Canty s variant lays in the fact the testing pixels are completely independent from the training one therefore the assessment is less bias Canty s ENVI IDL extensions can be downloaded for free from his webpage http mcanty homepage t online de software html Chapter 5 Methodology 33 5 4 3 Accuracy assessment Image analysis is not complete before an accuracy assessment is performed Accuracy of thematic maps resulting from classification can be defined as the degree of correct representation of land cover The most widely accepted assessment tool in image classification is the confusion matrix Canty 2010 Stehman 1992 Tso and Mather 2009 The confusion matrix is an array of i X i dimension where i is the number of classes The labels in the first row represent the training data classes whereas the labels in the first column represent the testing data classes The
36. 00 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000 7 0 41000 0 02710 0 00860 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000 8 0 41200 0 03310 0 00860 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000 9 0 41400 0 03980 0 00950 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000 _ n A1endlo nazala nann a anaona nnnnn n nananalo nannn n nonnai nnana n annnn o nanasl_n annnnl_ n nanani A nnnnn _ annonl_n ananol_n qnnanl_a anannl_n annnal_n an ar m Chapter 10 Appendices 72 b Save the document both in ods format and as ASCII file e g txt Start ENVI Go to Spectral gt Spectral Libraries gt Spectral Library Builder Choose ASCII file as Input Spectral Wavelength and click OK Browse to the filter function file saved in txt An Input ASCH File window opens automatically Make sure that the wavelength column number is correct e g 1 if you placed it in the first column You can leave the FWHM column blank Set Micrometers as Wavelength Unit Leave 1 as Y Scale Factor Click OK In the Spectral Library Builder window click Import gt From ASCII file Browse to the same txt filter function file and click Open In t
37. 17 19 21 23 25 master O LI 24 6 8 10 12 14 16 18 20 22 24 ASTER l 2l 3 4 slol s L L L L L 1 1 1 L 1 1 1 1 1 1 1 L 1 1 L L L L L L 1 0 0 0 5 1 0 1 5 2 0 2 5 Wavelength km 22 25 OO 8 i 20212324 7 28 29 30 1 2 33 3435 36 B 27293133353739 41 43 45 47 49 I man Mi MASTER 26 28 30 32 34363840 42 44 46 48 50 1011 12 13 14 ASTER L L 1 L 1 L L 1 L L 1 L L L 1 L 4 4 1 L L 4 L L L 4 3 4 6 8 10 12 14 15 Wavelength pm Fig 2 Channels and wavelength regions of the MODIS ASTER and MASTER radiometers a reflective channels b thermal channels Boxes represent the bandwidth for each sensor channel The numbers refer to the band numbers for each sensor Image from Li and Moon 2004 Chapter 3 Material 9 A full description of MASTER Sna er aia i e i n z Number of channels 50 optical and electronic devices is given Number of pixels 716 Instantaneous field 2 5 mrad by Hook e al 2001 They also of view Total field of view 85 92 provide specific information about in Platforms DOE King Air Beacheraft B200 NASA ER 2 and NASA DC 8 flight calibration depending on the Fixed size DC 8 oe Pixel size ER2 50 m er Pixel size B200 5 25 m utilized aircraft platform Spectral 25 ea ein P i 2 without refueling and radiometric calibration of 1200 range d seio mi without refueling channels is performed on a regular DC 8 range 5403 statute miles without refueling basis pre and postflight see Arnol
38. 2 3 1025 1036 Chang C and Lin C 2011 LIBSVM a library for support vector machines ACM Transactions on Intelligent Systems and Technology TIST 2 3 27 Chapter 11 References 83 Chen X Vierling L Deering D 2005 A simple and effective radiometric correction method to improve landscape change detection across sensors and across time Remote sensing of environment 98 1 63 79 Clevers J Bartholomeus H M cher S De Wit A 2004 Land cover classification with the medium resolution imaging spectrometer MERIS EARSeL eProceedings 3 3 2004 Coates A Jackson J Collins L Cronin T Dowsett H Bybell L Jung P Obando J 1992 Closure of the Isthmus of Panama the near shore marine record of Costa Rica and western Panama Geological Society of America Bulletin 104 7 814 828 Colby J 1991 Topographic normalization in rugged terrain Photogrammetric Engineering and Remote Sensing 57 531 537 Cole R 2010 Social and environmental impacts of payments for environmental services for agroforestry on small scale farms in southern Costa Rica International Journal of Sustainable Development amp World Ecology 17 3 208 216 Congalton R 1991 A review of assessing the accuracy of classifications of remotely sensed data Remote sensing of environment 37 1 35 46 Congalton R 1988 A comparison of sampling schemes used in generating error matrices for assessing the accuracy of maps generated from remotely sensed
39. 7 2 Data collection The major issues faced in the field by applying the response protocol were in order i plot access and ii location of the sampling plot in the ground For the former it is very important to consider legal and security issues as access to private land is not always possible Distinctive signs such as an official car or a disclaimer document should help in most cases To locate more easily the plot an idea is to follow the path to reach it on a topographic map or a digital support such as Google Earth It was in fact experienced during the Chapter 7 Discussion 50 field work that access to the Google Earth platform is possible even without internet connection given that the program was preliminary opened roughly on the area of interest In this way also the zooming in out tool remains active Recording the camera height as described in the field manual Appendix I can be time demanding when the plots are numerous and superfluous for the objectives so it can be omitted in similar studies No error assessment Van Genderen et al 1978 was conceived for evaluating the accuracy of field data the main reason being the simplicity of the legend applied However it is recommended to employ at least two operators for the field data collection in order to reduce even more chances of error in plot location and or class attribution Time gap between the sensed image and the field data collection is an issue to be
40. 757 6 0 042 0 711 0 71 31 0 1524 3 9134 3912 7 0 0418 0 7499 0 75 32 0 1548 4 0677 4 067 8 0 042 0 8 0 8 33 0 153 4 2286 4224 9 0 0417 0 8658 0 866 34 0 153 4 3786 4374 10 0 0407 0 9057 0 906 35 0 1446 4 5202 4 522 ll 0 0403 0 9452 0 946 36 0 1608 4 6684 4 667 12 0 0542 1 6092 1 608 37 0 1521 4 8233 4 822 13 0 0526 1 6645 1 666 38 0 1487 4 9672 4 962 14 0 0514 1 7196 1 718 39 0 1495 5 116 5 117 15 0 0521 1 7748 1 774 40 0 1578 5 2629 5272 16 0 0506 1 8281 1 826 al 0 3645 7 7599 7815 17 0 0457 1 8751 1 874 42 0 4333 8 1677 8 185 18 0 0575 1 9244 1 924 43 0 3543 8 6324 8 665 19 0 0504 1 9807 1 98 4 0 4253 9 0944 9 104 20 0 0481 2 0806 2 08 45 0 4083 9 7004 9 706 21 0 0511 2 1599 216 46 0 3963 10 116 10 115 2 0 0508 22106 2 212 47 0 5903 10 6331 10 554 23 0 0513 22581 2 258 48 0 6518 11 3293 11 365 24 0 0683 23284 232 49 0 4929 12 117 12 097 25 0 0641 23939 2 388 50 04618 12 8779 12 876 Table 2 Spectral characteristics of MASTER channels On the left channels 1 25 VNIR and short wavelength infrared SWIR band definition On the right channel 26 50 mid thermal infrared mid TIR Table from Hook e a 2001 12 13 16 19 s CHL ld J i 893114 11415 21718 5 26 6 7 A 13 57 911 13 15
41. G gt 450 bands were selected This means that the bands used for further processing from MDA index were all except 13 to 15 18 to 20 and 23 to 25 and for the MDG index all except 16 18 to 20 and 22 to 25 Last as mentioned in 5 3 3 EnMAP Toolbox does not provide a statistical comparison among bands However EnMAP Toolbox like ENVI on the other hand offers the possibility of identify noisy bands by visual interpretation A screen shot with one significant band and the three noisy ones is presented in Fig 12 A 1 3 sel_master lele l x Zoom 056 4 tt A FE A 2 16 sel_master ies i351 y 351 geox 205907 42 geoy 1095198 1 Grey 1 00854 Zoe O56 eg ala rat a k 351y 351 ge0x 205907 42 geoy 1095198 1 Grey 0 431210 Fig 12 Screen shot of 4 MASTER spectral bands displayed in EnMAP Toolbox Clockwise from top left corner band 3 16 17 and 18 The image is MASTER 1B March 11 2005 over the study site in Turrialba Costa Rica Only these 3 bands 16 17 and 18 were classified as noisy in EnMAP Toolbox Chapter 6 Results 39 6 4 Image analysis 6 4 1 Results of unsupervised classification The Gaussian Mixture algorithm could be applied only to images with reduced number of features Therefore the algorithm was only tested on PCA at 4 bands MNF at 6 bands first 5 bands of the radiance at sensor image and first 5 bands of the
42. MODIS ASTER airborne simulator MASTER data and NDVI a case study of the Kochang area Korea Canadian Journal of Remote Sensing 30 2 123 136 Lichtemberg PdSF 2010 Occurrence incidence and grower perception of Fusarium oxysporum f sp cubense in banana intercropped with coffee and trees in Costa Rica and Nicaragua smallholders Masters thesis Universit t Bonn Germany Loveland T Reed B Brown J Ohlen D Zhu Z Yang L Merchant J 2000 Development of a global land cover characteristics database and IGBP DISCover from 1 km AVHRR data International Journal of Remote Sensing 21 6 7 1303 1330 Lu D Mausel P Brondizio E Moran E 2002 Assessment of atmospheric correction methods for Landsat TM data applicable to Amazon basin LBA research International Journal of Remote Sensing 23 13 2651 2671 Mahiny A and Turner B 2007 A comparison of four common atmospheric correction methods Photogrammetric engineering and remote sensing 73 4 361 Chapter 11 References 88 Malinverni E Tassetti A Mancini A Zingaretti P Frontoni E Bernardini A 2011 Hybrid object based approach for land use land cover mapping using high spatial resolution imagery International Journal of Geographical Information Science 25 6 1025 1043 Mather P 2004 Computer processing of remotely sensed images Wiley Online Library Matthew M Adler Golden S Berk A Felde G Anderson G Gorodetzky D Paswaters S Shippert M 2003 Atmospheric
43. Siegmund A 2003 Automatic land cover analysis for Tenerife by supervised classification using remotely sensed data Remote Sensing of Environment 86 4 530 541 Chapter 11 References 87 Kleinn C 2007 Lecture notes for the teaching module Forest Inventory 1 revised ed Kottek M Grieser J Beck C Rudolf B Rubel F 2006 World Map of the Koppen Geiger climate classification updated Mezeorologische Zeitschrift 15 3 259 263 Kruse F 2004 Comparison of ATREM ACORN and FLAASH atmospheric corrections using low altitude AVIRIS data of Boulder CO Paper presented at the meeting of the 13th JPL Airborne Geoscience Workshop Kumar BM 2006 Carbon sequestration potential of tropical homegardens In Kumar BM and Nair PKR Tropical Homegardens A Time Tested Example of Sustainable Agroforestry 185 204 Dordrecht the Netherlands Springer Lagemann J and Heuveldop J 1983 Characterization and evaluation of agroforestry systems the case of Acosta Puriscal Costa Rica Agroforestry systems 1 2 101 115 Landgrebe D 1999 Information extraction principles and methods for multispectral and hyperspectral image data Information processing for remote sensing 82 3 38 Law K and Nichol J 2004 Topographic correction for differential illumination effects on IKONOS satellite imagery International Archives of Photogrammetry Remote Sensing and Spatial Information 641 646 Li P and Moon W 2004 Land cover classification using
44. Truth Percent Class sh coffee 3 sh coffee 2 sun coffee sugar cane grassland Unclassified 295 14 0 00 24 0 00 tree crown co es 2450 0 00 2 10 0 00 barren land 0833 019 0 00 0 29 0 00 paved road 0 65 1 82 0 00 0 04 0 00 building 0 00 0 00 0 00 0 07 0 00 water 0 00 0 00 0 00 0 00 0 00 sh coffee 3 83 96 Cale 0 00 1 96 0 00 sh coffee 2 6 55 77 67 0 78 ela Al 0 00 sun coffee 0 00 1 63 DEl 2A 0 25 0 00 sugar cane 2 78 ae 0 00 81 53 0 00 grassland 0 00 0 00 0 00 0 00 100 00 mixed tree gr 1 64 6 25 0 00 9 84 0 00 Total 100 00 100 00 100 00 100 00 100 00 Ground Truth Percent Class mixed tree gr Total Unclassified 2 68 1235 tree crown co 7 94 17291 barren land 0 10 2 88 paved road 2 34 2 58 building 5 83 10 21 water 0 00 0 67 sh coffee 3 ae 8 55 sh coffee 2 HASI 20 02 sun coffee 0 10 1 42 sugar cane 7 69 18 88 grassland 0 00 0 45 mixed tree gr 60 20 15 09 Total 100 00 100 00 Fig 16 Confusion matrix of sieved amp clumped MASTER radiance at sensor image 25 spectral bands classified with ENVI ML algorithm using 11 LUC classes The shown part of the confusion matrix indicates the matches between the testing column and the training row areas expressed in percentage Overall accuracy 74 8 Kappa Coefficient 0 7002 Chapter 6 Results 45 Fig 17 Examples of rule images resulting from ML classification of MASTER reflectance scene On the left rule image of class 2 shade coffee at 23 strata On
45. a set which appears in the list of Select Input File In Spectral Subset highlight the non thermal bands 1 25 Click OK In the Resize Data Parameters window maintain the default values and set Output to Memory if the output exceeds the available memory save to File CQ Resize Data Parameters Output File Dimensions Samples 716 xfac 1 000000 Lines 20000 yfac 1 000000 Output Size 1 432 000 000 bytes Resampling Nearest Neighbor File Output Result to Memory X OK Queue Cancel b On the main menu bar click on File gt Open External File gt Generic Formats gt HDF Select the same MASTER 1B product and click OK In the HDF Dataset Selection highlight the PixelLatitude and PixelLongitude datasets Chapter 10 Appendices 68 Select HDF Datasets 20000x1 Aircraft Altitude 65 20000 BlsckBody1Counts 5020000 Black Body2Counts 50x20000 Head 1Counts 71 amp 20000 PixelElevation 71 amp 20000 SensorZenithAngle Number of items selected 2 Select All tems Clear All tems Click OK c On the main menu bar click on Map gt Geometry from Input Geometry gt Super Georeference from IGM As input file select Memory 1 the calibrated non thermal bands Click OK Select as Input X Geometry Band PixelLongitude and click OK Select as Input Y Geometry Ba
46. according to their spectra and is not subject to the user s interaction However since in our case different classes have little spectra variations Fig 14 training the classifier with reference areas in supervised classification outputs the best results In terms of processing time ML was above all the fastest making the Gaussian Mixture and SVM more demanding to implement Among supervised algorithms the one that performed best was Maximum Likelihood Table 6 It has to be considered on the other hand that the accuracy indexes in a confusion matrix only refer to that specific thematic map and cannot be generalized Most scientific efforts focus nowadays on machine learning algorithms like SVM or artificial neural network for their promising potentials Erbek e al 2004 Huang et al 2002 In addition some authors Malinverni e al 2011 Walter 2004 Yu e al 2006 embraced object based classifications rather than based on single pixels with positive results Canty s modified accuracy assessment lead to slightly lower results than the ENVI standard accuracy tests The modified Canty s algorithms were repeated a number of times and the accuracy indexes kept on being lower than when all reference pixels were used as training set Probably the 1 3 reference data excluded in Canty s variant from the training set did contribute to better train the classifiers One possible explaination could be the high spectral variance within ea
47. aly Copenhagen G ttingen and Costa Rica for being just so gr eat I am very thankful to you all Table of content List f Feuer i SE Noble ae ee ii List of Abbrevidtio nsee nane Asa Eee OE ER AE iii Physical constant eek v Hleformas unse ie vi Abstractas ireen einen esse Vii ZUSIN MER ASUS neuster viii 1 Introdueti n unse ie 1 2 Study Site nennen A E E R RA EN 5 2 1 Stone E AE E ai 5 2 2 Climate an ag 5 23 Geomorphology u ee 5 2A Land Use Reese 6 3 Matetal een a aa 7 SLMSSLER masse es E aeia denaii 7 3 1 1 Instrument and data desenpton ausge 7 3 12 CARTA missione ee 9 Dad Ge bye l image an 11 Did PERO LECCE aD es essen E 12 SA Digital clevation mode an ee nea neaereee 12 3 5 Softwar Programs essen 12 4 Land use and cover LUC cases a 13 5 Methodology cananan iae ae ee 19 Sl Sampling desien nun 19 5 141 Population user einen 19 5 12 Sampling ameniona aa AA T A AE E 20 3 13 Sampling plot nd un ae u R stan 23 5 2 Data collet ara ERR EEE 24 Did mage process 25 5 3 1 Georeferencing and subsetting the image 25 5 32 AlmospHerie COME CON sie aukekee res R 26 5 3 3 Fe t re SOLE CHAO een e a a eg 27 Orthogonal transformations with PCA and MNE usesnenesessesesnenenn 28 Feature seleetion in Rue en 28 Feature selection with EnMAP Toolbar 29 54 LG OS analyse ee ei RER 29 5 4 1 Unsupervised classification ne en 29 Gaussian M AE e a a sextape a R S 30 54 2 Supefvised classificatio
48. antations FRA 2000a Other thresholds are taken from the FAO or from the IGBP LCCSs FRA 2010 According to the FAO FRA 2010 agroforestry systems are classified as Other land more precisely as sub category Other land with tree cover given that trees are gt 10 Note that coffee agroforestry systems might or might not be associated with trees and still be classified as such as the coffee bushes themselves are occasionally used as woody fuel Thacher et al 1996 Although more speculations on tree cover thresholds in agroforestry can be discussed Zomer ef al 2009 the agroforestry systems observed in the region were only shade coffee plantations thus implying a degree of tree cover tree plants are estimated to cover approx 50 of land according to a previous work by Chapter 4 Land use and cover LUC classes 17 Lichtemberg 2010 on agroforestry systems at 23 strata in the same region No ID Color Class name Definition 1 10 11 Tree crown cover Shade coffee at 23 strata Shade coffee at 2 strata Sun coffee Sugar cane Mixed tree grassland Grassland Bare soil Paved road Building Water bodies Non agricultural area gt 60 covered by the nadir tree crown projection on the ground Coffee shrubs grown under trees with gt 3 strata vegetation structure Coffee shrubs grown under trees with 2 strata vegetation structure Sun grown coffee planta
49. be applied only to floating point images and not to the 2 byte integer scaled reflectance Chapter 6 Results 40 The required processing time for clustering was above 60 minutes No pixels were left unclassified though unsupervised algorithms contemplate this possibility 6 4 2 Supervised classification As for the unsupervised case also for supervised classification the algorithms were tested on all images This was meant to evaluate both the need for atmospheric correction feature selection and evaluate different classification algorithms The confusion matrices of the highest scoring supervised classification algorithms are summarized in Table 6 The three most accurate thematic maps are presented in Fig 13 The graph showing the average spectral emission of different land use classes follows in Fig 14 and Fig 15 a Fig 13 Clockwise starting from the top left corner subset of the original MASTER image over Turrialba standard Maximum Likelihood ML in ENVI using the radiance at sensor MASTER image 25 spectral bands Canty s modified ML classification of the same image Canty s modified ML classification of MNF at 25 bands 21 Ina 32 bit processor 2 GB RAM 2 16 GHz Chapter 6 Results 41 Algorithm Bands Notes Overall accuracy Kappa coefficient MaxLikelihood 22 no 16 to 18 Reflectance img 70 0 0 6432 MaxLikelihood 22 no 16 to 18 71 7 0 6638 SVM 22 no 16 to 18 R
50. ces and easy to implement by one two operators Stratified sampling is suggested as alternative approach Among the tested classification algorithms supervised methods performed better than unsupervised ones in terms of accuracy and in some cases also processing time Among the former the standard ENVI ML lead to the highest accuracy followed in order by Canty s modified accuracy assessment for ML SVM in EnMAP Toolbox standard SVM and Canty s modified accuracy assessment for SVM In order to reduce data size the most relevant bands for spectral land use determination were researched In particular two orthogonal transformations PCA and MNF two out to out feature selections with R and EnMAP Toolbox and FLAASH atmospheric correction model were applied Surprisingly none of these attempts performed better than the full MASTER image at 25 bands calling for further investigations Accuracy was mainly improved by post processing operations particularly sieving amp clumping 1 3 and class combining 1 9 Ideas for possible study developments are referred to the Outlooks chapter Chapter 9 Outlooks 58 9 Outlooks e Analysis of MASTER TIR bands for land use classification French et al 2000 Keuchel et al 2003 Nemani and Running 1997 e comparison of classification algorithms and outputs with images taken from other sensors see for example Xie e al 2008 e comparison of MASTER classification performance
51. ch LUC class This needs further investigations Chapter 7 Discussion 55 targeting both the minimum sample size the reference area selection and the legend Very interesting would also be to analyze the spectral variability and overlapping between LUC classes in order to evaluate their separability In the case of vegetation classification special relevance might take in this context a NDVI band composite One possible way to investigate and minimize class variance could be through the bootstrap method Efron and Tibshirani 1986 McRoberts and Meneguzzo 2007 PCA and MNF did not perform better than the raw image in none of the classifications in terms of accuracy However the orthogonal transformations at 4 and 6 bands did require less processing time as expected Considering the aim of the study the size of the study area and the limited number of visible near and short wave infrared VNIR SWIR bands we would not recommend to manipulate the feature space There is not yet clear consensus in the literature on the thresholds above which to consider a classification map accurate Thomlinson e al 1999 proposes 85 as Overall Accuracy and no class less than 70 The considered MASTER image with 10 m spatial resolution showed to have high potential in this regard for shade coffee agroforestry systems detection in the region classified respectively with 86 and 71 producer s and user s accuracy Ihe possible reasons why 86 o
52. ch in Agroforestry IDL Interface Description Language IGM Input GeoMetry LCCS Land Cover Classification Scheme LP DAAC Land Processes Distributed Active Archive Center LUC Land use cover MASTER MODIS ASTER airborne simulator MDA Mean Decrease Accuracy MDG Mean Decrease Gini ii ML Maximum Likelihood MNF Minimum Noise Fraction MODIS Moderate Resolution Imaging Spectroradiometer MODTRAN MODerate resolution atmospheric TRANsmission NASA National Aeronautics and Space Administration OSM OpenSteetMap PCA Principal Component Analysis RGB Red Green Blue ROI Region Of Interest SGL Super Geometry Lookup SVM Support Vector Machine SWIR Short Wave InfraRed TIR Thermal InfraRed ToA Top Of Atmosphere VNIR Visible Near InfraRed Physical constants km kilometer m meter msl meter above sea level WW micro Watt sr steradian W Watt File formats CSV Comma Separated Values HDF Hierarchical Data Format HDR High Dynamic Range ODS OpenDocument Spreadsheet SLI Spectral LIbrary TIF Tagged Image Format TXT TeXT Abstract Agroforestry is an old practice Despite the increasing attention paid in the last decades to its ecological and economical benefits little is known so far on the actual spatial extend One major problem for spectral identification of agroforestry systems through remote sensing is the wide range of species and structure compositi
53. ching the image histograms using for example ERDAS Imagine The feathering option in ENVI ENVI Tutorial Mosaicking seems in fact too simplistic and does not solve the problem throughout but just in the overlapping area The described atmospheric correction models assume the Earth surface being flat and having a Lambertian reflectance behavior Tso and Mather 2009 In this way the radiance pixel values are homogeneously corrected throughout the image without taking 25 Interatively Re weighted Multivariate Alteration Detection Chapter 7 Discussion 52 into account that different elevations and roughness of the landscape cause different solar incident and radiance angles Though not applied in this study two top methods for topographic correction are suggested band ratioing and the Minnaert method Colby 1991 Law and Nichol 2004 Tso and Mather 2009 Both take into account the satellite position as well Despite the promising results by Matthew e al 2003 the reflectance image atmospherically corrected using FLAASH did not perform better than the radiance at sensor image Table 6 and was therefore abandoned during classification The need for atmospheric correction of a satellite scene of one acquisition date can be called in question if no accuracy benefits arise from it However a deeper analysis of the methodology applied could reveal involuntary misuse of the model For suggestions how to improve the app
54. coarse resolution sensors Working Paper 29 FAO Rome FRA 2000b On definitions of forest and forest change FAO Rome FRA 2010 Global forest resources assessment terms and definitions Working paper 144 E FAO Rome Chapter 11 References 85 Franklin S Wulder M 2002 Remote sensing methods in medium spatial resolution satellite data land cover classification of large areas Progress in Physical Geography 26 2 173 French A Schmugge T Kustas W 2000 Discrimination of senescent vegetation using thermal emissivity contrast Remote sensing of environment 74 2 249 254 Fuchs HJ Kleinn C Zamora S Aberle H 2010 BP TechReport1 AWF G ttingen University Unpublished technical report Gelli G and Poggi G 1999 Compression of multispectral images by spectral classification and transform coding Image Processing LEEE Transactions on 8 4 476 489 Gomez C Mangeas M Petit M Corbane C Hamon P Hamon S De Kochko A Le Pierres D Poncet V Despinoy M 2010 Use of high resolution satellite imagery in an integrated model to predict the distribution of shade coffee tree hybrid zones Remote Sensing of Environment 114 11 2731 2744 Green A Berman M Switzer P Craig M 1988 A transformation for ordering multispectral data in terms of image quality with implications for noise removal Geoscience and Remote Sensing LEEE Transactions on 26 1 65 74 Gretton A Herbrich R Chapelle O Rayner PJW 2001 Estimatin
55. ction Chapter 5 Methodology 19 5 Methodology In this chapter the methodological theory applied to process and classify the MASTER image is outlined When needed more technical references on the procedures are available in the appendices 5 1 Sampling design When performing any classification with remote sensed resources it is indispensable to ensure references for the output accuracy The references can be already validated thematic maps interpreted space or airborne images ground truth collection or a combination of both Plourde and Congalton 2003 Stehman and Czaplewski 1998 In this study we opted for a visual interpretation of the MASTER and GeoEye 1 image accompanied by field data collection The choice of the population size and sampling design was conceived to be as efficient as possible in terms of logistic time and financial resources At the same time the aim was to maintain a scientifically sound and statistically valid approach The sampling design is the protocol adopted to select the observation units Stehman and Czaplewski 1998 As the majority of land use and forest inventories the applied sampling technique in this study is systematic sampling This means that the location of the observation plots follows a systematic pattern Kleinn 2007 Drafting a sampling protocol requires delineations of the population sampling frame sampling plot and sampling unit hereinafter defined 5 1 1 Population A
56. d Scan speeds 6 25 12 5 25 rps Products Radiance at sensor Level 1B et al 1996 for procedure details ee ey ae ee Calibration MIR TIR 2 on board blackbodies Data Format Hierarchical Data Format HDF Examples for MASTER data can be Digitization 16 bit ordered free of charge from the Table 3 Summary of MASTER sensor characteristics NASA Jet Propulsion Laboratory B200 DC 8 and ER 2 refer to different aircraft models website used by NASA Table from Hook e al 2001 lt http masterweb jpl nasa gov gt at a maximum of 5 scenes per order They are compiled as Level 1B product which means that data are already available in radiance at sensor values and contain ancillary information about geo location navigation and calibration Eurimage 2011 MASTER data are stored in the Hierarchical Data Format HDF4 3 1 2 CARTA mission Costa Rican Airborne Research and Technology Applications CARTA missions are a cooperation between the Centro Nacional de Alta Tecnologia CENAT and NASA The main objectives are the acquisition of airborne images of Costa Rica for national use and scientific research Until now there have been two CARTA missions one in 2003 and one in 2005 In both these occasions four sensors have been used to collect digital scanner data and and analogue airphotos MASTER airborne simulator Leica RC 30 metric camera Cirrus Digital Camera System DCS and HyMap hyperspectral scanner ISDA 2011 These informati
57. dance MSc Henning Aberle for kindly translating the abstract and sincere aid MSc Paul Magdon for the inspiring discussion on sampling and classification the whole AWF department in particular Basanti Bharad Becksch fer Philip Buschmann Axel Dockter Ulrike Fischer Christoph Heydecke Hendrik Kywe Tin Zar Malla Rajesh Schlote Reinhard Vega Araya Mauricio for the amazing availability kindness and good advice Sabine Schreiner for having being a wonderful mentor and example Prof Dr Niels Stange and the whole SUFONAMA team for allowing this great experience the truly inspiring teachers I had in the last year especially Dr Stergios Adamopoulos Dr Gehrard B ttner Prof Dr Hanns H fle Dr Ronald K hne Prof Dr Bo Larsen Prof Dr Ralph Mitl hner Dr Carsten Schr der Among the Costa Rican partners Dr Miguel Dita por el apoyo excepcional y disponibilidad Dr Pablo Siles and Dr Oscar Bustamante para las observaciones siempre constructivas y la ayuda incondicional en el trabajo de campo Dr Ana Tapia por las ideas MSc Nancy Chaves por hacerme sentir como en casa y por la experiencia inolvidable Dr Christian Brennez and MSc David Brown por la gentil ayuda y cooperativas Ligia Quezada and Karol Araya para la inmensa disponibilidad y asistencia And last but not least mamma e Serena per avermi insegnato la cura e l amore ih Ly agile ya Ava ua ala Lb Saum Ld Som all the friends met in It
58. e 4 in Latin America and the Carribean Paper presented at the meeting of the ISHS Acta Horticulturae 897 International ISHS ProMusa Symposium on Global Perspectives on Asian Challenges Pocasangre L and P rez Vincente L 2009 Impacto potencial de la entrada de raza tropical 4 del mal de Panama Fusarium oxysporum f sp cubense en la industria bananera y platanera de Am rica Latina y el Caribe In Taller de entrenamiento sobre diagn stico y caracterizacion de la marchitez por Fusarium o Mal de Panama Bioversity International Raunkiaer C and Gilbert Carter H 1937 Plant life forms The Clarendon Press Reisner Y De Filippi R Herzog F Palma J 2007 Target regions for silvoarable agroforestry in Europe Ecological engineering 29 4 401 418 Richards JA and Jia X 1999 Remote Sensing Digital Image Analysis An Introduction 3rd Edited by Ricken DE and Gessner W Secaucus NJ USA Springer Verlag New York Inc Russo R and Budowski G 1986 Effect of pollarding frequency on biomass of Erythrina poeppigiana as a coffee shade tree Agroforestry systems 4 2 145 162 Schwarz M and Zimmermann N 2005 A new GLM based method for mapping tree cover continuous fields using regional MODIS reflectance data Remote Sensing of Environment 95 4 428 443 Chapter 11 References 90 Smith GM and Milton EJ 1999 The use of the empirical line method to calibrate remotely sensed data to reflectance International Journal of Remote Sensing 20
59. e decision on the class variance with n t s A where n is the sample size is the t distribution s is the standard error and A is the confidence interval width Foody e al 2006 Kleinn 2007 This is in line with the consideration that the model of data distribution is crucial for example a classifier such as Maximum Likelihood 5 4 2 requires values for the mean vector and the variance covariance matrix per each class to operate effectively Tso and Mather 2009 e other statistic formulas accounting for the data distribution are discussed by Fitzpatrick Lins 1981 and Congalton 1991 e generalizing this concept Foody ef al 2006 state that the for supervised classification the training set size is not fixed but can vary considerably depending on the properties of the chosen classification algorithm e the estimations mentioned above might be severely influenced by erroneous or unrepresentative samples included in the training set Though the effect of 14 n BIL 1 II b where n is the sample size B the upper a x 100 percentile of the chi square distribution a the required confidence lever amp the number of classes I the percentage of land covered by class 4 and 4 the required precision in percentage Chapter 5 Methodology 22 outliers should be limited for large samples Mather 2004 presents a supportive method to weight the observations Applied sample size In spite of al
60. eb jpl nasa gov NASA 2005 CARTA 2005 Final Flight Summary Report Airborne sensor facility NASA Ames Research Center Chapter 11 References 89 Nemani R Running S 1997 Land cover characterization using multitemporal red near IR and thermal IR data from NOAA AVHRR Ecological Applications 7 1 79 90 Olschewski R Tscharntke T Benitez P Schwarze S Klein A 2006 Economic evaluation of pollination services comparing coffee landscapes in Ecuador and Indonesia Ecology and Society 11 1 7 Palma J Graves A Bunce R Burgess P De Filippi R Keesman K Van Keulen H Liagre F Mayus M Moreno G others 2007 Modeling environmental benefits of silvoarable agroforestry in Europe Agriculture ecosystems amp environment 119 3 4 320 334 Ploetz R and Churchill A 2009 Fusarium wilt the banana disease that refuses to go away Paper presented at the meeting of the Proceedings of International ISHS ProMusa Banana Symposium Ploetz RC Pehh KG Jones DR Stover RH Lomerio EO Tessera M Quimio AJ 1999 Fungal diseases of the root corm and pseudostem In Disease of banana abaca and enset CABI Publishing Plourde L and Congalton R 2003 Sampling method and sample placement How do they affect the accuracy of remotely sensed maps Photogrammetric engineering and remote sensing 69 3 289 297 Pocasangre L Ploetz R Molina A Vicente LP 2011 Raising awareness of the threat of Fusarium wilt tropical rac
61. echnical record of the steps follows in Appendix G The two indexes used by randomForest to rank feature importance are the Mean Decrease Accuracy MDA and the Mean Decrease Gini MDG MDA ranks variables according to their difference in prediction accuracy MDG is the normalized sum of all Gini decreasing impurities along a tree over the total number of trees Calle and Urrea 2011 It is implicitly assumed that noisier bands will score low in the list Chapter 5 Methodology 29 Feature selection with EnMAP Toolbox EnMAP Toolbox 1 1 is a software developed by Humbold University of Berlin and DLR Deutsches Zentrum f r Luft und Raumfahrt for multi and hyperspectral data analysis It is currently developed specifically for hyperspectral data from EnMAP a German scientific satellite mission with envisaged launch in 2013 However with some modifications EnMAP Toolbox can process images from other remote sensing sources too Through EnMAP Toolbox it is possible to invoke imageSVM an IDL tool for support vector machine regression and classification analysis based on LIBSVM by Chih Chung Chang and Chih Jen Lin 2011 imageSVM is freely available for scientific purposes at http www2 hu berlin de hurs projects imageSVM php Alternatively imageSVM can run from the IDL Virtual Machine without necessarily owning a license for IDL or ENVI Van der Linden et al 2010a Feature selection is an indispensable part of data preparation
62. ed on the sieved image Sieving parameters 2 min threshold 8 number of neighbors Clumping parameters 3 operator size row 3 cols The highlight row shows the best combination in terms of accuracy 22 Ina 32 bit processor 2 GB RAM 2 16 GHz 23 Ina 64 bit processor 16 GB RAM 4 double processors Chapter 6 Results 44 Clumping was performed on the sieved image In these case if clumping was performed prior sieving directly on the MASTER classified image the final accuracy would have been lower results not shown Sieving leaves some pixels unclassified their number decreases after clumping but does not disappear At last we performed Combine Classes for the best sieved and clumped output sieving and clumping of MASTER image classified with the standard ML algorithm We started by analyzing carefully the confusion matrix Fig 16 and rule images Fig 17 Ground Truth Percent Class tree crown co barren land paved road building water Unclassified 1 60 1 20 90 09 0 00 tree crown co 65 14 0 00 0 45 0 00 0 00 barren land 0 33 88 43 0 00 0 07 0 00 paved road 0 87 0 00 88 34 1 89 0 00 building 2 65 0 00 4 04 85 96 0 00 water 0 00 0 00 0 00 1 24 98 68 sh coffee 3 bal ale 2 41 0 00 0 00 0 00 sh coffee 2 8 34 Sachs 4 04 19109 0 00 sun coffee 0 48 0 00 0 00 0 00 0 00 sugar cane 2 20 a 0 00 0 29 0 00 grassland 0 03 0 00 0 00 0 58 0 00 mixed tree gr 219 2 65 2 24 7 78 1 32 Total 100 00 100 00 100 00 100 00 100 00 Ground
63. ees with a percent canopy cover gt 60 and height vegetate d h Broadleaf Forests exceeding 2 meters Consists of seasonal broadleaf tree communities with g i N an annual cycle of leaf on and leaf off periods Cultivated 5 Mixed Forests Lands dominated by trees with a percent canopy cover gt 60 and height aquatic areas exceeding 2 meters Consists of tree communities with interspersed E SEHE mixtures or mosaics of the other four forest cover types None of the forest 4 Aquatic or types exceeds 60 of landscape regularly flooded 6 Closed Shrublands Lands with woody vegetation less than 2 meters tall and with shrub il Semi natural canopy cover is gt 60 The shrub foliage can be either evergreen or aquatic deciduous vegetation 7 Open Shrublands Lands with woody vegetation less than 2 meters tall and with shrub canopy cover is between 10 60 The shrub foliage can be either L a nd su rfa ce evergreen or deciduous 8 Woody Savannas Lands with herbaceous and other understorey systems and with forest Artificial canopy cover between 30 60 The forest cover height exceeds 2 meters sues Savannas Lands with herbaceous and other understorey systems and with forest canopy cover between 10 30 The forest cover height exceeds 2 meters 10 Grasslands Lands with herbaceous types of cover Tree and shrub cover is less than Terrestrial 10 f A i Permanent Lands with a permanent mixture of water and herbaceous or woody y Direne Wetlands vegetation that
64. eflectance img 68 0 0 6085 SVM 22 no 16 to 18 68 3 0 6109 SVM 25 68 5 0 6144 Canty ML 22 no 16 to 18 Reflectance img 71 1 0 6558 Canty ML 22 no 16 to 18 71 3 0 6588 Canty SVM 22 no 16 to 18 67 5 0 6016 Canty SVM 25 67 6 0 6024 Canty ML 16 MDA 67 7 0 6182 Canty ML 17 MDG 69 3 0 6364 Canty SVM 16 MDA 66 6 0 5906 Canty SVM 17 MDG 66 3 0 5872 Canty ML PCA 4 48 8 0 4199 Canty ML MNF 6 59 0 0 5223 En MAP SVM 22 no 16 to 18 See a 71 1 0 6790 En MAP SVM 22 no 16 to 18 See b 64 6 0 6069 Table 6 Summarizing table of overall accuracies and Kappa coefficients obtained from confusion matrices of several supervised algorithms ML Maximum Likelihood SVM Support Vector Machine MDA Mean Decrease Accuracy MDG Mean decrease Gini a Min g 0 01 Max g 10 g and C Multipliers 2 Min C 1 Max C 10000 b Min g 0 01 Max g 1 g and C Multipliers 2 Min C 100 Max C 10000 The highlight rows indicates the highest accuracy scores Results from different algorithm band combination are not shown because of lower accuracies Chapter 6 Results 42 Spectral Library Plots Fig 14 Average land use class spectra of reference ROIs associated with MASTER radiance at sensor image 25 bands Line colors correspond to the legend on page 17 Unit of the x axis wavelength um y axis radiance W m sr Spectral Library Plots Fig 15 Average
65. en the first csv file with OpenOffice Calc In the Text Import window under the Separator option section tick Comma Semicolon Space and Merge delimiters Have a quick look at the preview at the bottom and click OK In the Calc document delete the extra header rows leaving just one row which should display the titles ID X Y LAT LON B1 B2 B3 B22 Delete also the extra columns and make sure that all values are aligned in the right place Go to File gt Save as and tick Filter options You can overwrite the modified document on the original csv file if you like Repeat the last 5 steps for all ASCII files b Prepare the R environment for feature selection Start R Load the Remdr package by clicking on the main menu bar Packages gt Install packages Choose Germany Goettingen as CRAN Mirror Then scroll down the list until Remdr and click OK The Remdr window is useful to save the input commands for records and future developments Alternatively to the Remdr interface you can use TINN R downloadable for free Load the Random Forest package by clicking on the menu bar Packages gt Load packge Scroll down the list until randomForest select it and click OK To save this step type in the the R Commander window gt install packages randomForest dependencies TRUE _ Chapter 10 Appendices 76 Then type gt library randomForest Assign an internal
66. f the shade coffee pixels were correctly detected whereas only 71 of the pixels classified as shade coffee were actually belonging to that class is probably linked with the mistake mentioned at the beginning of this sub chapter alias the existence of class mixed tree grassland Interestingly was the fact that not always sieving and clumping different classification results improved the accuracy Sometimes both were useful others clumping would increase the accuracy only if applied on the sieved image pg 43 others it would do it only when applied directly on the original classification and other times both decreased the accuracy No particular pattern could be identified which linked the first classified NIR VIS 27 Normalized Difference Vegetation Index It is given by the reflectance value ratio were NIR VIS NIR is the near infrared and VIS is the red visible region of electromagnetic spectrum In MASTER they would be channel 7 to 11 and 5 respectively see also Li and Moon 2004 Chapter 7 Discussion 56 image and the post processing results We recommend therefore to test each time the accuracy through the confusion matrix when undertaking sieving and clumping Combine classes on the other hand proved to have more tangibly effects in all circumstances 1 9 and 0 0125 overall accuracy and Kappa coefficient respectively This follows the knowledge that decreasing the number of classes increases the overall
67. g the Leave One Out Error for Classification Learning with SVMs Citeseer Haklay M and Weber P 2008 OpenStreetMap User generated street maps Pervasive Computing IEEE 7 4 12 18 Hook S Myers J Thome K Fitzgerald M Kahle A 2001 The MODIS ASTER airborne simulator MASTER a new instrument for earth science studies Remote Sensing of Environment 76 1 93 102 Hsu C and Lin C 2002 A comparison of methods for multiclass support vector machines Neural Networks IEEE Transactions on 13 2 415 425 Huang C Davis L Townshend J 2002 An assessment of support vector machines for land cover classification International Journal of Remote Sensing 23 4 725 749 Hubert Moy L Cotonnec A Le Du L Chardin A Perez P 2001 A comparison of parametric classification procedures of remotely sensed data applied on different landscape units Remote Sensing of Environment 75 2 174 187 Hughes G 1968 On the mean accuracy of statistical pattern recognizers Information Theory IEEE Transactions on 14 1 55 63 Chapter 11 References 86 IAAST 2008 Agriculture at a crossroads Global report International Assessment of Agricultural Knowledge Science and Technology for Development Washington DC Island Press ICRAF 1993 Annual report 1993 International Centre for Research in Agroforestry Nairobi Isaac M 2009 Ecological and social interactions in sustainable agroforestry management Cocoa in Ghana PhD diss U
68. g the QuantumGIS eVis plugin Appendix C 6 3 Image processing 6 3 1 Georeferencing The maximum observed spatial offsets between the georectified MASTER against the OpenStreetMap OSM layer and GeoEyel mosaic were 3 and 4 pixels 30 and 40 m respectively Fig 9 Surprisingly the spatial offset between the GeoEye 1 manually georeferenced mosaic and the OSM layer was 10 15 m lower than the dispatched original 5 band GeoEye 1 image 20 m offset against OSM On the other hand orthorectification of the GeoEye 1 standard product with the available DEM using ENVI 4 8 improved the spatial accuracy to 5 7 m Chapter 6 Results 36 Fig 9 MASTER image left and GeoEye 1 mosaic right overlayed with OpenStreetMap OSM layer 6 3 2 Atmospheric correction From FLAASH one reflectance image was created and scaled for a factor of 10 000 Appendix F A sample spectra profile of the reflectance image compared to the original radiance at sensor is shown in Fig 10 The 3 bands automatically excluded by FLAASH because classified as noisy were band 16 17 and 18 1 78 to 1 98 um 4 R ROT Resize Band 5 sub1 moasic 020104419 6 EI amp IT Cy 2 R sand Math FLAASH Band 5 sel master 0201 SIE X File Overlay Enhance Tools Window File Overlay Enhance Tools EL a x NE PB is ve E un Ls Ga Q 4 Spectral Profile sel_ mas
69. gons Chapter 10 Appendices 80 u E i E g u a I E H m E E E E m E Fig 19 Grid overlaid to the selected area of 7 x 7 km each quadrant is 200 x 200 m real scale Once the grid was created 36 quadrants were systematically selected in red The first quadrant was chosen randomly by the software the remaining 35 accordingly this is to ensure an unbiased selection 5 At the end of each on screen digitization a class key number is assigned to the polygons This number corresponds to one of the land cover classes listed FIELD MEASUREMENTS Following there is a complete description of the steps required for the field sampling 1 2 4 Before going to the field upload some relevant and easy to find GPS points on a GPS device These will help identifying the target quadrant in the field The corresponding waypoint number can be written on the quadrant hard map that you will bring along you can take a print from the available imagery Choose one system to name the quadrants and therefore facilitate the field work One idea is to name them with decimal numbers the whole number before the comma represents the row number while the decimal number after the comma represents the column number For example
70. he Input ASCH File window make sure that all 25 Bands are selected in the Select Y Axis Columns Set the Wavelength Units to Micrometers Click OK In the Spectral Library Builder window now all 25 bands should appear Click on Select all and then Plot to view a graph of the transmittance responses of the different bands Save the plot as Image Save the filer function as Spectral Library file sli Run FLAASH Model Open ENVI if you haven t done yet so Go to Spectral gt FLAASH In the FLAASH Atmospheric correction Model Input Parameters window Select as Input Radiance Image the MASTER georeferenced and subset image In the Radiance Scale Factor windows that opens type in 10 0 as single scale factor for all bands this is because of the scale conversion from W lum x sr x m to UW nm x sr x cm Choose a name for the Output Reflectance File Choose an Output Directory for FLAASH Files Choose a Rootname for FLAASH Files e g flaash_ In the same window convert the Scene Center Location to DD and type Lat 9 89692700 Lon 83 68173500 Sensor Type UNKNOWN MSI Sensor Altitude 8 457 Km Ground Elevation 0 770 Km Pixel Size 10 m Flight Date March 7 2005 Chapter 10 Appendices 73 Flight Time 15 16 52 As Atmospheric Model choose Tropical Water retrieval No Water Column Multiplier 1 00 Aerosol Model Maritime see ENVI Atmospheric Correction Module Aerosol Retrieval None Initial vis
71. he sample size For Support Vector Machine SVM classifier 4 very interesting strategies are proposed by Foody et al 2006 for improving accuracy and reducing considerably sample size 16 In Quantum GIS Vector gt Research Tools gt Random selection 17 Distance between any point in a quadrant and its equivalent in the adjacent quadrant 49000000 1361111 1111 1200m 36 Chapter 5 Methodology 23 5 1 3 Sampling plot and unit Theoretically a systematic grid should be seen as a large cluster plot made of several sub plots Kleinn 2007 Stehman and Czaplewski 1998 define sampling plots as all the sites that are actually sampled However for simplicity in this thesis we will use the terms quadrants and sampling polygons to refer respectively to the grid sub plots and to the actual sampling plots Sampling polygons were digitized on screen within each quadrant by partitioning the image into regions of homogeneous neighboring pixels A LUC label was assigned to each polygon following the legend on page 17 by doing so all the polygon belonging pixels were also automatically labeled For a full procedure description see Appendix C This process called visual interpretation was performed by one person only using the georectified MASTER image and GeoEye 1 mosaic Fig 7 eeror Fig 7 Left one of the 36 selected sampling quadrants red frame whose land use classes have been digitized and visually inter
72. hms for land use activities International Journal of Remote Sensing 25 9 1733 1748 Eurimage 2011 Product level description online available at lt www eurimage it products products html gt accessed on 15 August 2011 Eva H Belward A De Miranda A Di Bella C Gond V Huber O Jones S Sgrenzaroli M Fritz S 2004 A land cover map of South America Global Change Biology 10 5 731 744 FAO 2003 WTO agreement on agriculture the implementation experience Developing countries case studies FAO 2006 World Soil Resources World reference base for soil resources USS Working Group WRB World Reference Base 103 ISBN 92 5 105511 4 Fitzpatrick Lins K 1981 Comparison of sampling procedures and data analysis for a land use and land cover map Photogrammetric Engineering and Remote Sensing 47 3 343 351 Foody G 2000 Accuracy of thematic maps derived by remote sensing Paper presented at the meeting of the Proceedings of the accuracy 200 conference Amsterdam Foody G and Mathur A 2004 Toward intelligent training of supervised image classifications directing training data acquisition for SVM classification Remote Sensing of Environment 93 1 2 107 117 Foody G Mathur A Sanchez Hernandez C Boyd D 2006 Training set size requirements for the classification of a specific class Remote Sensing of Environment 104 1 1 14 FRA 2000a Forest cover mapping amp monitoring using NOAA AVHRR amp other
73. ibility 40 0 Km orrection Model Input Parameters Input Radiance Image C Users Marina Desktop tesi Maps images sel_master_020104 tif Output Reflectance File C Users Marina Desktop tesi master_refl Output Directory for FLAASH Files C Users Marina Desktop tesi flaash Rootname for FLAASH Files flaash_ Scene Center Location DD lt gt DMS Sensor Type UNKNOWN MSI Flight Date neta M 7 2005 Lat 9 89692700 Sensor Altitude km 8 457 Mec Zur Right Time GMT HH MM SS 15 16 52 a Lon 83 68173500 Ground Elevation km 0 770 E Pixel Size m 10 000 fea i pees Mods Water Retrieval No 4t Aerosol Retrieval 2 Band K T Water Column Multiplier 1 00 Initial Visibility km 40 00 Apply Cancel Haie Mutiopectei Settings J Advanced Settings Then click on the Multispectral button in the Multispectral Settings window leave GUI as Channel Definer As Filter Function File browse to the sli file Click OK Click Apply c Scale the values to 2 bytes integer data space In b you have created an image in reflectance values You might notice that the reflectance unit in the Z Spectrum is above 1 the ratio value you would normally expect is between 0 and 1 This is because it has to be scaled from floating point to 2 bytes integer data space Therefore Go to the main ENVI menu bar Basic Tools gt Band Math Ty
74. ical point of view the grid cell size 200 x 200 m appeared to be suitable for the objective of land use observation around Turrialba as each quadrant could be easily surveyed However the spectral variance among polygons of the same LUC class but belonging to different quadrants could be statistically studied using the intracluster correlation coefficient This information should help in finding unconsidered spatial autocorrelation and be potentially incorporated into a stratified sampling scheme A way to increase statistic accuracy might be to switch the square systematic grid to a triangular one this leads to a systematic unalignment of plots Though the precision increases as the observation points are scattered even more homogeneously over the area the square grid is just slightly less precise Kleinn 2007 Therefore for simplicity a square grid is often preferred As referred in the methodology several are the formulas proposed in the literature to estimate the sample size for remote sensed data within a certain confidence level For Mather 2004 rule of thumb 30 x 25 bands 750 pixels should be sampled per class this gives a total of 8250 pixels 750 x 11 classes below our sample size of 13 744 Before data collection other formulas were not possible to apply as no variance Chapter 7 Discussion 49 estimations were available After classification it is possible to obtain statistical information on the processed data in
75. ichards and Jia 1999 pixel g is in class i provided d g 2 d g forally TD At given that the discriminant function is d g log X l e X W The moments J and u are directly estimated from the training data Due to the limited number of parameters estimations the computational time is also limited Canty 2010 Being a parametric model ML strongly relies on the prior assumption of data structure If the normal distribution assumption is correct parametric models show stronger results as compared to non parametric ones Whitley and Ball 2002 ML classification in ENVI outputs also a gray scale rule image consisting of the discriminant functions for each class and training observation Canty 2010 In the case of ML the rule image values are the probability of pixel belonging to class This rule image can be used to evaluate the classification and possibly to modify the discriminant thresholds ITT ENVI Tutorial Classification methods ITT ENVI 1999 Support Vector Machine Support Vector Machine SVM classifiers are state of the art machine learning algorithms Iso and Mather 2010 Like in ML methods the structural misclassification error is minimized by accounting for data variance But whereas in ML the variance is empirically estimated SVM partitions the feature space by taking a subset of the training data the support vectors without invoking their distribution Canty 2010 Being a non parametric model SVM has
76. icking with the eVis Event ID tool on each GPS point that has an associated picture should display the photo Save the project and all layers used Especially when sharing the document remember that all files pictures vector layers associated files like SHP SHX QPJ PRJ DBF etc must be located in the same folder Chapter 10 Appendices 66 10 4 Appendix D Field form QUADRANT Date PICTURES Polygon ID GPS point Direction Notes Sequence number Pic No Camera height cm Polygon ID GPS point Direction Notes Pic No Camera height cm Polygon ID GPS point Direction Notes Pic No Camera height cm add an appropriate number of similar sections to survey all polygons Chapter 10 Appendices 67 10 5 Appendix E Georeferencing and subsetting MASTER 1B products The following is an adjusted version of the ENVI tutorial Geo referencing images using Input Geometry ENVI Tutorial Georeferencing to suit the specific case of MASTER 1B images Required files MASTER 1B product HDF file vector file with study area to be subset Required softwares ITT ENVI a Start ENVI Click on File gt Open External File gt Thermal gt MASTER Browse to the MASTER 1B hdf file On the main menu bar click on Basic Tools gt Resize Data Spectral Spatial Click on the loaded MASTER dat
77. in imageSVM Since the band selection is mainly based on visual assessment this method has some limitations for large hyperspectral datasets However it can be considered a potentially interesting complement to feature selection with ENVI and R For familiarizing with imageSVM in this context see Appendix H 5 4 Image analysis 5 4 1 Unsupervised classification Unsupervised classification algorithms use statistical techniques to cluster dimensional data according to their spectral values IT T ENVI 2001 They require the user to input the number of desired clusters Several such algorithms are available in ENVI and were applied in this study For compactness and relevance reasons only the Gaussian Mixture algorithm is presented here Chapter 5 Methodology 30 Gaussian Mixture The Gaussian Mixture is an iterative algorithm somehow superior to other unsupervised classification methods because it uses multivariate normal probability densities thus allowing higher clustering flexibility Canty 2010 This classification tool is only available in ENVI with the IDL extension The algorithm used in this study is actually Canty s modified version of Expectation Maximization EM applied to a Gaussian Mixture model Unfortunately this extension is not suited for large or high dimensional data sets because it does not exploit the ENVI tiling facility it allows though a more flexible algorithm parametrization 5 4 2 Supervised classificat
78. ing a column with wavelength values from 0 4 to 2 55 um and 25 adjacent columns one for each band containing the associated transmittance response Add a 0 in all blank cells The results looks similar to 14 4 D0 Sheett allt r Sheet1 1 Default STD Sum 0 04710 o _e 100 File Edit View Insert Format Tools Data Window Help 2 Bus Anar SE xhe D eu bo a oma gir kle g E aria lp EE ew RN EEE 0 2 A Cm m URL He a B10 Az un A CS PEN ss lA Ee hts re ina mPa ae se a ea ans Sel eT 1 Wavelength mu Band 1 Band2 Band3 Band4 Band5 Band6 Band Band8 Band9 Band 10 Band 11 Band 12 Band 13 Band 14 Band 15 Band 16 Band 17 Band 18 Band 18 2 0 40000 0 01280 0 00440 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000 3 0 40200 0 01670 0 00440 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000 4 0 40400 0 01860 0 00640 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000 5 0 40600 0 02020 0 00640 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000 6 0 40800 0 02500 0 00860 0 00000 0 00000 0 00000 0 00000 0 00000 0 00000 0 000
79. ion Contrary supervised classification requires the user to select and input training areas which are used to train the classifier Though the procedure might be tedious the results are ofter more satisfactory than those from unsupervised classification Tso and Mather 2009 The disadvantage is that for defining the classes some prior knowledge on the study site is necessary Like for unsupervised classification ENVI offers several options as algorithms The two that gave the most promising results in this study are the Maximum Likelihood ML and the Support Vector Machine SVM which were implemented into 5 variants standard ML and SVM in ENVI ML and SVM with accuracy assessment modified by Canty Canty 2010 and SVM in EnMAP Toolbox Maximum Likelihood The Maximum Likelihood classifier ranks the pixel belonging probability to each class and assigns the pixel to the class with the highest probability Tso and Mather 2009 This assignation is based on the reflectance value or digital number per each band ML classifier assumes that all classes have a normal probability distribution prior band analysis However developments of this algorithm not applied in this study proved that there is the potential to use expected class distribution or pixel context to improve the final ML classification accuracy Strahler 1980 Hubert Moy et al 2001 Chapter 5 Methodology 31 The decision rule used by ENVIML classifier is Canty 2010 R
80. ion behind Raukizr s life form distinction is the position of the buds in the plant during the resting season highlighted in the drawing Banana trees and plantains can be considered a Herbaceous subgroup of Phanerophyte For this and other subdivisions see Ellenberg ez al 1967 Chapter 10 Appendices 63 10 3 Appendix C Digitizing sampling polygons update shape LUC values and link digital photos with eVis plugin in QuantumGIS a Create a vector file Start QuantumGIS Load your remote sensed image s e g the MASTER subset and the GeoEye 1 mosaic by clicking on the Open Raster Layer icon If you wish to change the image color composite right click on the name in the left list and choose Properties gt Style and modify the RGB bands e g for MASTER you can choose the true color composite RGB 5 3 1 Tick Use Standard Deviation and select 2 as factor Under Contrast enhancement gt Current select Stretch To MinMax Click OK In case of systematic sampling load the grid vector file or create it by clicking on the main menu bar Vector gt Research tools gt Vector grid Select randomly one quadrant by clicking Vector gt Research tools gt Random selection and select the other grid cells accordingly Save the so created selected grid cell vector file From the main menu bar Layer gt New gt New Shapefile Layer In the New Vector Layer window tick Polygon specify the Coordinate Reference Sys
81. l these considerations usually costs access to sampling sites and resource availability constrain data collection In addition in our case no estimations on land use could be exploited prior the actual data collection Thus being a pilot study it is acceptable to base the decision of the sample size simply on the objectives and context constrains postponing the evaluation of sample size adequacy to the Discussion and Outlooks chapters The sampling plots were distributed into 36 quadrants each of 200 x 200 m 4 ha and all the surface within these quadrants was scheduled to be surveyed The above seemed in fact reasonable figures to cover the highest number possibly all land use types and limit time and travel costs The systematic sampling frame was drawn as follows overlay of a square grid of quadrants to the 7 x 7 km study site using a geographical information system GIS software Quadrant grid cell size 200 x 200 m Grid orientation North South Random selection of one quadrant by the GIS software Systematic selecting other 35 quadrants at 1200 m distance No slope correction was applied to these quadrants since the object of interest was the number of pixels which are our sampling units see below and not the surface area 15 For example in the case of Maximum Likelihood ML supervised algorithm the mean vector and the variance covariance matrix per each class necessary to perform the classification influence t
82. land use class spectra of reference ROIs associated with MASTER FLAASH reflectance image 22 bands Line colors correspond to the legend defined on page 17 Unit of the x axis wavelength um y axis reflectance ratio Chapter 6 Results 43 If we consider the computational effort PCA at 4 bands MNF at 6 and 25 bands transformed images took in general slightly less time for classification processing as compared to the MASTER image at 22 and 25 bands in all classification algorithms Time difference was more pronounced among algorithm types with SVM classification being the slowest approximately 30 minutes in ENVI vs about 1 2 minutes for ML However SVM processing time in EnMAP Toolbox highly varied depending on the input parameters especially sampling type g and C For the tested combinations it was in the range of 3 to 200 minutes 6 5 Post classification processing The 3 most accurate classification images were first sieved and then clumped The respective accuracy assessment results are presented in Table 7 Algorithm Image Tool Overall acc Kappa coef ML MASTER 25 bands sieve 72 7 0 6765 D as eee ee Canty ML sieve 72 3 0 6712 clump 73 8 0 6889 MNF 25 bands sieve 72 2 0 6706 clump 73 8 0 6887 Table 7 Overall accuracies and Kappa coefficients of 3 post classified images ML stands for Maximum Likelihood MNF 25 for Minimum Noise Fraction at 25 bands Clump was perform
83. lassification and the fact that actually no comparison among different time scenes was required In feature selection both FLAASH ENVI and EnMAP Toolox pointed as bands 16 17 and 18 as noisy 1 803 to 1 953 um This is conform to the knowledge that water is absorbed by the atmosphere of electromagnetic wavelength of channels that span around 1 45 1 95 and 2 50 um Jensen 2000 The R package randomForest proved to be somehow less effective in identifying noisy bands in this study The Mean Decrease Accuracy and Mean Gini Accuracy indexes showed no accordance Fig 11 and no apparent accuracy benefit arose from choosing the best appointed bands for classification Table 6 7 4 Image analysis As mentioned in chapter 4 the legend applied it is clearly suited for the considered study site and should require modifications if extended to a different area The classification scheme suggested on page 16 should provide the backbone for modifications It has been shown that by reducing the number of classes the classification accuracy increases therefore we suggest to keep the legend as condensed as the objectives allow or alternatively to merge some classes at the end of the classification The most prominent classification misattributions arising from the confusion matrices belong to the classes tree crown cover and mixed tree grassland This might be imputed to the legend and scale of observation polygons were in fact defined as pixel
84. lication of FLAASH see Outlooks In addition other atmospheric modeling techniques could be tested like Atmospheric CORrection Now ACORN and ATmospheric REMoval ATREM ACORN uses MODTRANA4 radiative transfer code like FLAASH for estimating atmospheric parameters but contrary to the latter it retrieves water vapor concentration using bands 0 94 and or 1 15 um Kruse 2004 ATREM on the other hand is based on a 6 S atmospheric correction code improved by Vermote e al 1997 and briefly discussed in Tso and Mather 2009 A comparison of these 3 atmospheric modeling techniques is given by Kruse 2004 There are some proves that empirical methods such as DOS Dark Object Subtraction and ELC Empirical Line Calibration could perform better than theoretically superior but generalized models in calibrating remote sensed images for DOS see Lu e al 2002 for ELC see Karpouzli and Malthus 2002 though the issue is still debated Smith and Milton 1999 Mahiny and Turner 2007 Empirical techniques require some knowledge of ground conditions which have to be retrieved either from the image itself in DOS or by ground measurements also subsequent in ELC Tso and Mather 2009 As last remark for atmospheric correction we can conclude that FLAASH model was 26 ATREM is not available in ENVI but see ITT ENVI 2001 Chapter 7 Discussion 53 superfluous in this study the main two reasons being the poorer performance in c
85. mote Sensing 58 3 4 22 238 Wang G Gertner G Anderson A 2005 Sampling design and uncertainty based on spatial variability of spectral variables for mapping vegetation cover International Journal of Remote Sensing 26 15 3255 3274 Whitley E and Ball J 2002 Statistics review 6 Nonparametric methods Critical care London 6 6 509 513 Woodcock C Strahler A Jupp D 1988 The use of variograms in remote sensing Il Real digital images Remote Sensing of Environment 25 3 349 379 Xie Y Sha Z Yu M 2008 Remote sensing imagery in vegetation mapping a review Journal of Plant Ecology 1 1 9 Yu Q Gong P Clinton N Biging G Kelly M Schirokauer D 2006 Object based detailed vegetation classification with airborne high spatial resolution remote sensing imagery Photogrammetric Engineering and Remote Sensing 72 7 799 Zomer RJ Tabucco A Coe R Place F 2009 Trees on farm analysis of global extend and geographical patterns of agroforestry ICRAF working paper no 89 World Agroforestry Centre Nairobi Kenya Declaration of authorship Hiermit versichere ich gem 9 Abs 5 der Master Pr fungsordung vom 22 07 2005 dass ich die vorliegende Arbeit selbst ndig verfasst und keine anderen als die angegebenen Quellen und Hilfsmittel benutzt habe 22 September 2011 Marina Martignoni
86. n anal 30 Maximurn Likelihood na 30 Support Vector Machine enan enaa a ea tens 31 SVM with EnMSP Foolbaresteunntiisnese east 32 MEand SVM by Cine anne ae T AE 32 54 3 Accuracy assessment nee 33 5 5 Post classification Processing ne aapetvebsvaducsseihSevaneeatele 34 6 Results isra ar RE 35 Gol Sampling desien nenne ETE EE E EATE EEEN 35 6 2 Data collection en en ee ea EETA E andes 35 6 3 Imi ge processing ee 35 6 3 1 GeoreferenciNg ena este 35 6 32 Atmospheric Corfeclion nn rear 36 03 3 Beature selection eenn a een ee 37 6 4 Tage analysis anne A E AAS 39 6 4 1 Results of unsupervised classifications kann ea 39 6 4 2 Supervised classificati Neee E RE EER 40 6 5 Post classification Processing ersehen 43 7 Discussions inna tice sl tes Sesas ace eae es hecho ig ea tieeses 47 71 Sampling COSI GTi saczesticseacatasusesmeasensonncagdsstabensou whereendesesssaueoeass arennantenysescasseeseeiae 47 YD Dataeolleet1 on ee ee 49 1 3 WAV BS processing ee ee 50 74 Image analysis innii aa E E E Ni 53 S Conclusio eene e e oaa E E RA EEA E dite 57 2Oitlooer essen 58 10 Appende en ee SR a 60 10 1 Appendix A FRA Ben ee ee een 60 10 2 Appendix B Raunkiaer s life form classification ueesseesesesesenenenenennenenennenn 62 10 3 Appendix C Sampling plot digitization a hese Bee sede 63 102 Append amp D Field formi am 66 10 5 Appendix E Georeferencing MASTER Hles u 67 10 6 Appendix F FLAASH atmospheric correction
87. n den letzten Jahrzehnten deren kologische und konomische Vorteile zuteilwurde ist bisher wenig ber die tats chliche r umliche Verteilung bekannt Ein gro es Problem f r die spektrale Identifizierung von Agroforstsystemen durch Fernerkundung ist die breite Palette der Arten und der Strukturzusammensetzung Aufgrund des gestellten Rahmens dieser Arbeit beschr nkt sich der Fokus auf Kaffee Bananen Mischkulturen Diese ist ein besonders h ufiges obwohl bedrohtes Agroforstsystem in Zentralamerika Die Hauptgr nde hierf r sind intensive Produktion und die Verbreitung spezifischer Krankheiten Das Ziel ist eine effiziente Methode zur Identifizierung von Kaffee Bananen Agroforstwirtschaft in dem Kontext der Klassifikation von Landnutzung und Landbedeckung zu finden Die Pilotstudie liegt in der N he von Turrialba Costa Rica Dennoch ist ein weiteres Ziel die Methodik auch f r gr ere Fl chen zu diskutieren Hyperspektrale MASTER und r umlich hoch aufgel ste GeoEye 1 Satellitendaten werden als Grundlage f r die Bildverarbeitung und analyse verwendet Eine Reihe von Feature Auswahl Techniken einschlie lich orthogonaler Transformationen FLAASH atmosph rischer Korrektur und Klassifikationsalgorithmen werden getestet Die Ergebnisse zeigen dass der Maximum Likelihood ML Algorithmus zu der besten Genauigkeit der Klassifizierung f hrt 77 im Vergleich zu Gaussian Mixture und Support Vector Machine SVM f r das ausgew hlte Untersuchu
88. n to the field crew and to the reader the observation units are spread homogeneously over the area and thus less autocorrelated all parts of the population are covered assuming that there is no cyclic land use distribution Kleinn 2007 There is however the risk of under or over represent some classes for example too few reference pixels belonging to rare classes are collected to train properly the classifier or too many energies are invested in digitizing the umpteenth polygon which does not contribute in increasing information on that land type Plourde and Congalton 2003 Thus considering the purpose of LUC classification and the drawbacks of systematic sampling it could be possible to apply a stratified sampling scheme in further studies as it would guarantee that all land use classes are represented equally In stratified sampling homogeneous objects on the ground are isolated and visually classified Object isolation can be performed through segmentation followed by semi automatic classification best suitable for high spatial resolution images alternatively unsupervised classification can also work to identify land cover with similar spectra in lower resolution products The selection of the objects to be sampled can be random or follow a n un aligned scheme Fitzpatrick Lins 1981 One disadvantage of stratified sampling is that transport logistic plot location and permission to access the sites might balanced out the time
89. named Sieve Clump and Combine Classes Sieve smooths the image by removing isolated pixels after a minimum cluster size has been input Clump merges similarity classified areas Combine classes allows the user to join classes with very similar spectra into a more generalized one ITT ENVI Tutorial Classification methods When appropriate all three tools were tested over the classification results Chapter 6 Results 35 6 Results 6 1 Sampling design A total of 207 sampling plots were digitized belonging to 12 land use and cover classes The number of selected reference pixels was 14 842 over a population of 490 000 6 2 Data collection 198 polygons were actually field validated as the remaining 9 sites could not be accessed leaving a total of 13 744 pixels to be used as reference areas During field data collection the LUC labeling of 38 polygons and the shape of 17 ones was modified due to visual misinterpretation of the remote sensed images The most commonly misattributed LUC classes from visual interpretation compared with field survey were mixed tree grassland 15 shade coffee at 23 strata 13 shade coffee at 2 strata 7 The LUC type changed in only in two sites between 2005 and 2011 it was anyway possible to trace back to the original land use being mainly agricultural fields under crop or fallow rotation All GPS point and digital pictures acquired during data collection were incorporated into vector files usin
90. natural breaks The class threshold values are inspired either by the FAO Forest Resource Assessment report FRA 2010 or by the IGBP definitions depending on the context and objectives Agricultural land is separated from non agricultural land on the basis of produced good crop vs wood respectively FRA 2000b 12 Decision Trees are a branch of classification algorithms based on the mapping of decision rules and relative outputs in a tree like format normally dichotomous Chapter 4 Land use and cover LUC classes 16 Tree Crown cover S Tree plantation Mixed tree orassland 10 lt Trees lt 60 u Shrubs Forest succession Non agri Pasture Urban greenery land Trees s 10 Grassland Shade coffee 3 strata A Other agrol systems Agricultural A Shade coffee 2 strata land Tueesat2 stuata Vegetated land Sun coffee Trees lt 10 Sugarcane Other crops Barren soil ail Mine Dirt road 4 Paved road Non vegland feg Building aa Water River Lake Fig 6 LUC classification scheme applied in this study The underlined LUC class names on the right are those selected for the final legend the others written in smaller font are just possible LUC classes not found or not applied in this study Vegetated land is defined as land with 210 vegetation cover FRA 2000b Agricultural land is defined on the basis of produced crop type and excludes tree pl
91. nd PixelLatitude and click OK Select as Input Projection Geographic Lat Lon as Datum WGS 84 as Units Degrees as output Projection UTM as Datum WGS 84 as Units Meters as Zone 17 N Input Projection of Geometry Bands OSS rep a E UTM State Plane NAD 27 State Plane NAD 83 Argentina Zone 1 Argentina Zone 2 Argentina Zone 3 won Output Projection for Georeferencing New Geographic Lat Lon State Plane NAD 27 4 State Plane NAD 83 Argentina Zone 1 Argentina Zone 2 Argentina Zone 3 Argentina Zone 4 Datum WGS 84 a a Zone 17 ON 28 IGM Input GeoMetry Chapter 10 Appendices 69 Click OK d A Build Geometry Lookup File Parameters window opens automatically As Output Pixel Size round the default value e g 9 6 gt 10 m As Output Rotation choose 0 Accept the other default values Enter an Output SGL file name Save to File the Georef Output and name it SGL Parameters Output Pixel Size 10 Output Rotation 0 Kemel Size Min 3 Ma5 Minimum Pixels to Resample 1 gt Enter Output SGL Filename Choose C Users Marina Desktop SGL_output Georeference Background Value 0 0000 Output Result to File Memory Output Georef Filename Choose El Compress C Users Marina Desktop Georef _output OK Queue Cancel Click OK e To sub set the study area over
92. ndaries of the sampling polygons checking the polygon shape and assigned class value registration of GPS coordinates and acquisition of digital photos from the observation point useful in case of doubt or as a record for future comparison Due to the broad class definition in the legend it is possible to assess the class belonging by simple visual survey of the area If any land use change occurred between the MASTER 2005 or the GeoEye 1 2010 image acquisition and the field survey 2011 we attempted to trace back to the land use in 2005 and annotate the inconsistency To help data collection the used field form template is shown in Appendix D Once the field data collection is concluded it is necessary to upload the land use information and GPS points into vector layers The digitized and field validated polygons become in this way the regions of interest ROIs used as reference areas for accuracy assessment and to train the supervised classifiers chapter 5 4 2 Chapter 5 Methodology 25 Data collection designed in this way can be preformed by one operator alone though two is the recommended number Other equipment details are present in the field manual The average sampling speed with the given location resources and conditions was 4 quadrants per day Data collection for this study was carried out in May This field procedure is rather economic in terms of time and money with transportation being the main facto
93. ngsgebiet Schatten Kaffee Agroforstsysteme wurden mit einem hnlichen Grad an Vertrauen klassifiziert obwohl es noch nicht mit Sicherheit m glich war die Anwesenheit von Bananen und Kochbananen zu ermitteln Stichworte Landnutzung bedeckung und Klassifizierung Agroforstwirtschaft MASTER Maximum Likelihood SVM Vili Chapter 1 Introduction 1 1 Introduction Agroforestry is in short the integration of forestry and agriculture A more precise definition by ICRAF 1993 regards it as systems and practices where woody perennials are deliberately integrated with crops or animals in the same land management unit either at the same time or in sequence with each other Following this agroforestry systems can occur in different forms from shifting cultivation to complex vegetation structures Nair 1985 In his definition Nair 1993 also mentions about the usual presence of two or more plant species the production of at least two outputs and the significant interactions among them Whereas most studies focused on the ecological Jose 2009 Palma e al 2007 Tornquist ei al 1999 economic Alavalapati e al 2004 Olschewski et al 2006 van Asten e al 2011 and social Cole 2010 Isaac 2009 Dahlquist et al 2007 benefits of agroforestry systems little is known on their actual spatial extend Unruh and Lefebvre 1995 Zomer et al 2009 Some of the figures present in the literature are summarized in Table
94. niversity of Toronto ISDA 2011 Project Large geospacial datasets Costa Rica 2050 Image Spatial Analysis Group http isda ncsa uiuc edu Costarica ITT ENVI Atmospheric Correction Module QUAC and FLAASH User s guide http www ittvis com portals 0 pdfs envi Flaash_Module pdf ITT ENVI ENVI Tutorial Classification methods http www ittvis com portals 0 tutorials envi Classification_Methods pdf ITT ENVI ENVI Tutorial Georeferncing Images using Input Geometry http www ittvis com portals 0 tutorials envi Georef_Input_Geometry pdf ITT ENVI ENVI Tutorial Mosaicking in ENVI http www ittvis com portals 0 tutorials envi Mosaicking pdf ITT ENVI 2001 ENVI Tutorials Research systems ed 620 pg Jensen JR 2000 Remote sensing of the environment An Earth resource perspective Prentice Hall Jose S 2009 Agroforestry for ecosystem services and environmental benefits an overview Agroforestry Systems 76 1 1 10 Karpouzli E and Malthus T 2003 The empirical line method for the atmospheric correction of IKONOS imagery International Journal of Remote Sensing 24 5 1143 1150 Kass D Jimenez M Kauffman J Herrera Reyes C 1995 Reference soils of the Turrialba valley and slopes of the Irazu volcano In Soi Brief Costa Rica ISRIC Kearns M and Ron D 1999 Algorithmic stability and sanity check bounds for leave one out cross validation Neural Computation 11 6 1427 1453 Keuchel J Naumann S Heiler M
95. nted in Appendix A Classification abstract representation of the reality as resulting from diagnostic criteria It describes both the discerning rules used by the classifier and the associated outputs A classification scheme can be bierarchical or non hierarchical Di Gregorio and Jansen 1998 As the term explains in the former the classes are assigned with progressively level of detail whereas in the latter all classes are defined at once The FAO LCCS is an example of hierarchical classification Fig 4 while the IGBP legend is non hierarchically defined Fig 5 CLASS CLASS NAME DESCRIPTION Cultivated amp Evergreen Lands dominated by trees with a percent canopy cover gt 60 and height i Needleleaf Forests exceeding 2 meters Almost all trees remain green all year Canopy is Managed arcas never without green foliage pes 7 2 Evergreen Lands dominated by trees with a percent canopy cover gt 60 and height Terrestrial Broadleaf Forests exceeding 2 meters Almost all trees remain green all year Canopy is never without green foliage N Semi natural 3 Deciduous Lands dominated by trees with a percent canopy cover gt 60 and height vegetation Needleleaf Forests exceeding 2 meters Consists of seasonal needleleaf tree communities with e an annual cycle of leaf on and leaf off periods Mainly 4 Deciduous Lands dominated by tr
96. o perform similar processing with other images or as reference material the interested reader is recommended to consult also the Atmospheric Correction Module QUAC and FLAAS User s gui ITT ENVI Atmospheric Correction Module a Before running the FLAASH Module you need to create a Filter Function file Go to the MODIS Airborne Simulator webpage were the MASTER channels are defined http mas arc nasa gov data srf_html TC4_master_srf html Towards the end of the page click on Download the SRF s for all bands zip and save the folder Unzip the files The resulting folder contains a list of c files Right click on the first c file on the list named msr0807a c01 and choose Open with Browse to OpenOffice Calc program and click OK A Text window pops up automatically make sure that the field Space is ticked Click OK Copy and paste the column with the wavelengths values into a new OpenOffice Calc document Make sure that in both the original and new document the cell format is set to contain at least 5 significant digits to change it go to Format gt Cells gt Numbers tab Copy and paste the column with the radiance values from the original file to the new Calc document Make sure that transmittance response values are matched with the corresponding wavelength Correct the number of significant digits to 6 Repeat the previous 2 steps for each band The result is a Calc document contain
97. oduct covering the considered study site was ordered for evaluation and further research developments within the funding project To facilitate LUC classification a raster file was created based on the GeoEye image 432 screen shots were taken from 1 15 km zenith distance from the ground and mosaiced using Photoshop CS4 This mosaic was then manually georeferenced in Quantum GIS 7 Rounded to 0 5 m due to US Government restrictions on civilian imaging Chapter 3 Material 12 3 3 OpenStreetMap OpenStreetMap OSM is an open source geospatial project where geographic data are collected and shared The aim is to create freely editable and accessible maps of the whole world licensed under the Creative Common Attribution ShareAlike 2 0 This means that it is possible for example to download spatial references such as street maps which have been traced with in situ GPS records without the burden of legal or technical restrictions A very handy plug in allows to invoke and download target OSM layers directly from the QuantumGIS working project The OSM layers are considered to have generally a good degree of spatial accuracy Haklay and Weber 2008 Mooney et al 2010 3 4 Digital elevation model A digital elevation model DEM of Costa Rica with 30 m spatial resolution was available and used to retrieve basic information such as elevation and slope when not directly possible from the MASTER image This DEM was produced from the ASTER
98. on The focus of this thesis is limited to coffee banana mixed cropping due to the work frame constrain This is a particularly common though threatened agroforestry system in Central America the main reasons being intensive production and specific disease spread The objective is to find an efficient methodology for identification of coffee banana agroforestry in the context of land use and cover LUC classification The pilot study site is located near Turrialba Costa Rica nevertheless a further goal is to discuss the methodology for larger areas as well Hyperspectral visible near and short wave infrared MASTER and spaceborne high spatial resolution GeoEye 1 data will be used as basis for image processing and analysis A number of feature selection techniques including orthogonal transformations FLAASH atmospheric correction and classification algorithms are tested The results show that Maximum Likelihood ML algorithm lead to the best LUC classification overall accuracy 77 as compared to Gaussian Mixture and Support Vector Machine SVM for the selected study site Shade coffee agroforestry systems were classified with similar degree of confidence though it was still not possible to detect with certainty the presence of bananas and plantains Key words land use cover classification agroforestry MASTER Maximum Likelihood SVM vi Zusammenfassung Agroforstwirtschaft ist eine alte Praxis Trotz der zunehmenden Aufmerksamkeit die i
99. on about the CARTA mission are provided here as Chapter 3 Material 10 context background and as sparkle for further sensor comparison studies The selected MASTER image for this research was taken on March 11 during the CARTA 2005 mission tile code MASTERL1B_0500311_04_20050307_1516_1543_ V01 Fig 3 As the study site was fully covered by this scene and without cloud cover it was not necessary to mosaic several tiles The High Altitude WB 57 aircraft used flew at 8438 8459 m height allowing a 10 m spatial resolution on the ground 0 100000 200000 300000 1200000 1200000 Caribbean Sea 1100000 1100000 Pacific Ocean 1000000 1000000 0 50 100 150 sk 100000 200000 300000 Fig 3 Study site indicated by the red frame 7x 7 km around the city of Turrialba Costa Rica The MASTER tile MASTERL1B_0500311_04_20050307_1516_1543_V01 from March 11 2005 was georectified using ENVI Tile swath width 14 2 km scanline length 200 km Grid projection system UTM Zone 17N Datum WGS 84 Map produced in ArcMap 10 0 Seasonal phenological variation can be a problem for remote sensing LUC classification especially when focusing on natural environments The selected MASTER scene was taken in March the driest month of the year which lowers the chances of cloud and haze presence but could also mean lower vegetation cover Thus although for LUC 6 Data measured with a GPS device on
100. oncise protocol for field data collection of reference areas e profiling an applicable classification scheme and legend e evaluate the accuracy and handiness of several unsupervised and supervised classification algorithms namely Gaussian Mixture Maximum Likelihood and Support Vector Machine in the context of this study e evaluate the advantage of atmospheric correction and feature selection for the purpose of classification Chapter 2 Study site 5 2 Study site 2 1 Location The study area is located in the central highlands of Costa Rica approximately 50 km East of San Jose Fig 3 on page 10 The frame of 7x 7 km includes the city of Turrialba and the research institute of CATIE 2 2 Climate The annual precipitation in Turrialba is about 2600 mm prevalently concentrated between June and December March is the driest month with lt 100 mm rainfall Precipitation usually exceeds evapotranspiration and relative humidity is above 85 throughout the year The mean annual temperature is 21 5 C with limited excursion Kass e al 1995 FAOCLIM data from Fuchs e al 2010 Because of these characteristics the area is classified in the K ppen Geiger eco climate scale as equatorial rainforest fully humid updated version for the second half of 20 century by Kottek e al 2006 Like in other parts of Costa Rica movements of air masses are influenced by land topography which is briefly outlined below 2 3 Geomorphology
101. particular the class standard deviation covariance and correlation per each band using ENVI Class Statistics However due to handling complexity of considering 25 spectral bands spectral autocorrelation of neighboring pixels feature space correlation class distribution definition and separability and lack of literature on the topic it was not possible to evaluate the chosen sample size or infer on the appropriate one within this study It would be extremely fascinating to read in the future of studies on the topic Once the sampling scheme and sample size have been selected for classification purposes it is also important to consider the sample placement and cluster size Plourde and Congalton 2003 suggest to create training samples in homogeneous areas rather than selecting larger clusters which include many mixed pixels Testing this effect on accuracy would be interesting by comparing our classification results with those from new digitized sampling polygons where all the edge pixels have been discarded When evaluating these results it is however important to consider possible sample size differences as the quadrants might no longer be wholly classified Congalton 1988 moreover showed that the size of each training areas should not be larger than 10 pixels as no information is added by extra contiguous pixels This is due to spectral correlation and reinforces the suggestion to digitize smaller plots within the original ones
102. pe in the expression b1 10000 00 Save the new image 31 These information can be retrieved from the Header file of the original MASTER file except for the Scene Center that should be calculated in QGIS Vector gt Geometry Tools gt Polygon centroid Chapter 10 Appendices 74 10 7 Appendix G Feature selection in R Random Forest package Required files one vector file e g shp for each land use class with all sampling polygons a Export land use ROI to ASCII format readable by R Start ENVI On the main menu bar File gt Open Image file and browse to the MASTER image Click OK In the Available Bands List window right click on the image name and choose Load True Color In the Image Display window click Overlay gt Vectors In the Vector Parameters Cursor Query window click on File gt Open vector file and browse to the vector files with the land use polygons Remember to set the file extension to the appropriate format in the search menu e g shp An Import Vector File window opens This is because to read shp format files ENVI needs to convert them first into evf Check the default parameters and fill in the blank fields Click OK In the Vector Parameters Cursor Query window highlight the first click layer name on the list and then on File gt Export Active Layer to ROI In the Export EVF Layers to ROI window choose Convert all records of an EVF layer to one ROI
103. preted over MASTER 1B georectified product RGB 5 3 1 Right the same quadrant over the GeoEye 1 mosaic The colored polygons represent different land use classes legend on page 17 Spatial offset is discussed in 6 3 1 Care was taken to avoid polygon overlapping At the end of the digitization a vector file was created for each LUC class The digitized polygons represent our sampling plots Contrary sampling units are defined as the reference units upon which the accuracy assessment is based Stehman and Czaplewski 1998 Thus the ultimate sampling units Chapter 5 Methodology 24 of this study are the image pixels and should not be confused with the sampling plots If preferred sampling plots can be thought as spatial pixel clusters In order to be considered ground truths the polygons need field validation Congalton 1991 Due to unavoidable human error they will be hereinafter referred to as reference areas and not exactly as ground truths 5 2 Data collection The response design defines the methodology to collect information on the LUC reference sites Stehman and Czaplewski 1998 Though in this sub chapter only a brief overview on the methodology is provided a more detailed description is listed in the field manual Appendix I To validate the polygon land use on the field the followed procedure involved location of the quadrant s position using GPS device and hard soft maps identification of the virtual bou
104. r to be considered 5 3 Image processing 5 3 1 Georeferencing and subsetting the image Once received the MASTER 1B tile in HDF format it is necessary to assign precise coordinates to the image georeferencing and subset the study area The ENVI tutorial Geo referencing images using Input Geometry provided a suitable backbone scheme for geo rectification though some modifications were required Appendix E Note that to follow these steps platform geometry information are necessary In case of MASTER products they are incorporated into the HDF file specifically under the datasets PixelLatitude and PixelLongitude After georeferencing the MASTER image will be overlapped with both the GeoEye 1 mosaic manually georectified and layers from OpenStreetMap OSM to test the accuracy of geometric correction Subsetting the image was performed by overlapping the study area vector file to the georeferenced MASTER tile Alternatively the exact geographic extension of the study area can be used The MASTER image was also spectrally subset selecting only the first 25 visible near infrared and short wave infrared VNIR SWIR bands and excluding the remaining 25 thermal infrared TIR ones This choice was driven by the preference of reducing data size and by the lower spatial resolution of TIR channels Gelli and Poggi 1999 Tso and Mather 2009 In addition TIR bands requires different calibration methodologies and handling as they record emis
105. ributes are correlated and must therefore be handled differently than standard statistical rules e sample size is related to the observation scale as the scale determines the sampling units and thus the variability within them For being adequate the sample size should account for variance between classes Tso and Mather 2009 e researchers dealing with multi and hyperspectral data should bear in mind that the higher the number of features the larger sample size should be in order to maintain same accuracy degree Hughes phenomenon Hughes 1968 e Mather 2004 advises to equal the minimum sample size per each class to the Chapter 5 Methodology 21 number of features multiplied by 30 He bases this figure on the notion that in univariate statistics a sample size gt 30 is considered large e Congalton 1991 experienced that 50 is a reasonable sample size for each vegetation class increased to 75 100 if the area is gt 1 mln acres or considering gt 12 classes Van Genderen ef al 1978 propose some tables in their article to estimate the required sample size depending on different confidence probabilities for example 20 and 30 should be respectively the class sample size for 85 and 90 accuracy e Congalton and Green 1999 propose a formula to estimate the minimum sample size for specific confidence intervals that reflects the class multinomial distributions e another possible approach is to base th
106. s Montagnini 2006 Pocasangre e al 2011 The association of trees to small and medium scale coffee banana farms is common as it provides additional income to the producers relieves the exposure of crops to pest diseases and contributes to maintain soil fertility Lagemann and Heuveldop 1983 However along with these facts small and medium scale coffee banana producers face several challenges namely fluctuating market prices competition and diseases which particularly affect indigenous cultivars One dramatic example is the so called Panama disease Fusarium oxysporum cubense FOC a fungi which spread in the first half of 20 century across Central America and swept away entire banana plantations of the Gros Michel cultivar Ploetz ei al 1999 amp 2009 Blomme e al 2011 Export oriented farms rapidly switched to banana cultivars not affected by FOC like Cavendish But among locals Gros Michel is still highly preferred for paid price taste and storage quality Pocasangre e al 2009 As the fungi affects mainly production rather than plant survival many small medium scale producers maintain Gros Michel in their farms mixed with other crops in the hope of a near future eradication of the problem To promote conservation profitability and development of coffee banana agroforestry systems in Latin America a project called Mejorando la producci n y mercadeo de bananos en cafetales con rboles de pequenos productores
107. s Orthogonally transformed images usually appear more colorful than color composites from spectral images because features are less correlated Richards and Jia 1999 Feature selection on the other hand does not transform the feature space but simply ranks the bands according to their importance in determining the land use class Orthogonal transformations with PCA and MNF The Principal Component Analysis PCA is a common technique to orthogonalize data sets It ranks and transforms features according to the highest variance Richards and Jia 1999 However this does not imply that noisy bands are excluded The Minimum Noise Fraction MNF transformation overcomes this problem by generating feature spaces that maximize the signal to noise ratio Canty 2010 It is actually a double Principal Component transformation where first noise is isolated and rescaled and in the second step the remaining bands are regularly processed as in a PCA ITT ENVI 2001 Performing PCA and MNF transformations is a rather automatized process in ENVI and easy to perform In both cases it is possible to select each time a different number of features bands to describe the final image For more theoretical background on PCA and MNF see respectively Landgrebe 1998 and Green e al 1988 Feature selection in R The package used in this study from the statistical program R is randomForest developed by Leo Breiman and Adele Cutler Breiman 2001 A t
108. s mentioned in 2 1 a study area of 7 x 7 km has been selected Having a spatial resolution of 10 m it is equivalent to about 490 000 pixels This can be considered our population better defined as the the interest domain of the study Kleinn 2007 Chapter 5 Methodology 20 5 1 2 Sampling frame The sampling frame delimits and identifies the sample population Stehman and Czaplewski 1998 In the case of a map a sampling frame can be areal boundaries which include all the sampling plots The question on how large should the sample population be in order to fulfill a certain accuracy confidence interval does not have an easy answer This is why a summarized literature review on the topic is introduced before presenting the chosen sample size for this study Short literature review on sample size for remote sensed data The following selections are mainly based on Foody e al 2006 Tso and Mather 2009 and Van Genderen ei al 1978 It emerges that e the sampled data must adequately represent the temporal state of the observed phenomena especially if the target classes are of a temporally changing nature Iso and Mather 2009 e several authors Atkinson 1996 Curran 1988 Van der Meer et al 1999 Wang et al 2005 Woodcock et al 1988 recommend the use of geostatistical methods to choose the most appropriate sampling scheme in remote sensing because since sampling sites have fixed locations in space their att
109. savings from manual polygon digitization of the whole quadrants In addition in the case of unsupervised clusters labeling it is possible to perpetrate systematic observer bias if the operator is not familiar with the study scene thus leading again to an unbalanced number of samples per class However a successful implementation of stratified random sampling scheme was done for example by Foody e al 2006 who also suggest 4 interesting tricks to reduce sample Chapter 7 Discussion 48 size when the focus is on one class only The population size of this study was not random as 49 km is the minimum required area for ordered GeoEye 1 or IKONOS satellite imagery This is an important factor to consider if the research is extended using different sensors or sites The sampling grid orientation was North South and this could be arguably objected as user bias choice However if there is not a cyclic pattern in the landscape e g systematic arrangements of sugar cane fields at 1200 m distance in the North South direction the error variance of the land use classes should not be much higher than from a sampling grid randomly oriented Kleinn 2007 Moreover it is easier to visualize and explain a systematic grid with cells oriented North South rather than one oriented at different azimuths Therefore from the trade off between increased error variance and higher practicality we decided to apply a North South orientation From a pract
110. scene archive during the ASTER Global DEM GDEM mission ASTER GDEM validation team 2009 The GDEMs are available for download from the Earth Remote Sensing Data Analysis Center ERSDAC of Japan or the NASA s Land Processes Distributed Active Archive Center LP DAAC websites 3 5 Software programs To perform the study the following software programs were used Licensed Open source e Adobe Photoshop CS4 e DNRGarmin 5 4 1 e ESRI ArcGIS Desktop 10 0 e EnMAP Box 1 1 64 bit operating systems e ITT ENVI 4 8 e JabRef 2 6 e OpenOffice 3 1 QuantumGIS 1 7 e R2 13 Chapter 4 Land use and cover LUC classes 13 4 Land use and cover LUC classes The LUC classes used in this study are a modified version of the FAO land cover classification systems LCCSs and IGBP legend The applied modifications mainly aim at reducing the number of classes while highlighting the prominent land features of the study area Before proceeding with describing the classes used it is important to clarify the meaning of the following terms Land cover the bio physical cover on the Earth surface Di Gregorio and Jansen 1998 E g tree mixed shrub tree stand water body bare soil Land use the arrangements activities and inputs people undertake in a certain land cover type to produce change or maintain it Di Gregorio and Jansen 1998 E g tree plantation agroforestry system swimming pool pit Legend the
111. sivity and not reflectance 18 In this thesis the terms georeferencing and georectification will be used interchangeably Chapter 5 Methodology 26 5 3 2 Atmospheric correction As indicated in Table 3 MASTER 1B products are already dispatched as radiance at sensor values this means that digital numbers DNs have already been translated into radiance values using the offset and gain metadata of instrument calibration Radiance is the radiant energy emitted from a particular area per unit time at a given solid angle in a specified direction Jensen 2000 It is measured in W m xsr On the way to the space this radiant flux interacts with various atmospheric gases and particles Thus radiance at sensor indicates the radiant energy recorded by the sensor which is affected by atmospheric scattering absorption reflection and refraction The task left to the user is to convert the radiance at sensor data first into true radiance and then to reflectance True radiance is free from atmospheric distortions which affect radiant fluxes from the surface to the sensor Reflectance is the dimensionless ratio between the radiant flux emitted by a surface and the radiant flux incident to it the sun light Jensen 2000 We want to evaluate if reflectance is more effective in land use spectra characterization than radiance at sensor as it does not carry atmospheric distortions and accounts for the intensity of the incoming radiating energy
112. spatial clusters irrespectively of their size provided that the LUC was homogeneous This means that if there were trees in a grassland field e g pasture the area was classified in one polygon as mixed tree grassland In the case of MASTER medium spatial resolution images with many mixed pixels this can be accepted as long as within the digitized polygon no pixel cluster larger than a certain size e g 3x3 could fall into a different class If this is the case it would be appropriate to re shape the polygon or revise the Chapter 7 Discussion 54 legend Thus in future studies we recommend first of all to digitize only pure pixels or pixels whose LUC is unambiguous Second to clearly define the size of the legend unit as this is crucial both for the classification accuracy and for the class definition Alternatively if the spatial resolution allows in between classes such as mixed tree grassland or shade coffee can be omitted and eventually only at the end created by merging neighboring pixel clusters e g tree crown cover pixels merged with grassland or coffee clusters within a specified spatial distance into a new class Generally all supervised classifications led to higher accuracies than the unsupervised Gaussian Mixture This result agrees with similar findings in literature Richards and Jia 1999 Tso and Mather 2009 Conceptually unsupervised classification is a rigorous method because classifies pixels purely
113. t include land that is predominantly under agricultural or urban land use Explanatory notes 1 The definition above has two options e The canopy cover of trees is between 5 and 10 percent trees should be higher than 5 meters or able to reach 5 meters in situ or e The canopy cover of trees is less than 5 percent but the combined cover of shrubs bushes and trees is more than 10 percent Includes areas of shrubs and bushes where no trees are present 2 Includes areas with trees that will not reach a height of 5 meters in situ and with a canopy cover of 10 percent or more e g some alpine tree vegetation types arid zone mangroves etc 3 Includes areas with bamboo and palms provided that land use height and canopy cover criteria are met Chapter 10 Appendices 61 TERM definition and explanatory notes OTHER LAND All land that is not classified as Forest or Other wooded land Explanatory notes 1 Includes agricultural land meadows and pastures built up areas barren land land under permanent ice etc 2 Includes all areas classified under the sub category Other land with tree cover OTHER LAND WITH TREE COVER sub category of OTHER LAND Land classified as Other land spanning more than 0 5 hectares with a canopy cover of more than 10 percent of trees able to reach a height of 5 meters at maturity Explanatory notes 1 The difference between Forest and Other land with tree cover is the land
114. tem CRS by browsing to the appropriate set and add a new attribute named for example LUC Class type Whole number width e g 10 You can also remove the default ID attribute present in the list Click OK Save the vector layer 5 files where preferred b Digitize polygons Click on Settings gt Snapping options Tick the newly created the cell grid vector layers the field Enable topological editing and all the vector layers whose polygons you do not want to overlap Avoid Overlapping Click OK Highlight the vector layer name in the left list if not already highlighted Click on the Toggle editing icon on the main menu bar Chapter 10 Appendices 64 Zoom in to the area you want to digitize You may help LUC interpretation by overlapping a higher resolution image e g the GeoEye 1 mosaic Once you are sure which pixels to include in the sampling polygon click on the Capture polygon icon Start drawing the verteces of your target polygon by left clicking on the image Terminate digitizing and close the polygon shape by right clicking Type in the LUC class value correspondent to your legend You can continue in this way to digitize the other polygons To facilitate drawing play with the Snapping options Mode and Tolerance values Every once in a while click again on the Toggle editing icon to save your work Once you have digitized all target polygons you can separate them in different vec
115. ter 020104 Ele 2 Spectral Profie sel master refl BM10000 I 0201041 2 cae Filem Editon Options Plot Functionm Help File Edit Options Plot Functi ctral Profile S 10 5 10 15 Index Nu Index Number Fig 10 Radiance left and FLAASH reflectance right images of MASTER scene over Turrialba The associated spectral profile refer to the same pixel Index band number Chapter 6 Results 37 6 3 3 Feature selection From PCA and MNF analysis three images have been created one PCA at 4 bands and two MNF at 6 and 25 bands respectively The number of bands is subjectively chosen by the operator However the aim was to condensate spectral information while preserving essential traits for classification The R randomForest MDA and MDG ranking results are shown in Fig 11 Bit Bi B9 B3 B5 B17 B14 0 34 ME Ei 0 40 mo m m m Fig 11 Feature selection using the Mean Decrease Accuracy MDA and the Mean Decrease Gini MDG indexes available in R randomForest package The bands B are ranked top to bottom according to the importance in determining the LUC class Only 22 bands appear in each list as the other 3 23 to 25 in both cases were considered less influential thus excluded This forest was obtained using the first 25 bands of MASTER data set seed 4543 trees 1000 Chapter 6 Results 38 From these rankings the first 16 MDA gt 0 38 and 17 MD
116. the advantage to fit well most data sets Whitley and Ball 2002 Though SVMs were originally conceived for binary classification a number of multiclass applications have been investigated Hsu and Lin 2002 Huang e al 2002 Foody and Mathur 2004 The multiclass approach used in ENVI to determine the membership class is regarded as a one to one voting scheme Canty 2010 Chapter 5 Methodology 32 SVM with EnMAP Toolbox With EnMAP Toolbox imageSVM described on page 29 is possible to classify an image using the SVM algorithm Though the results should be very similar to SVMs applied in ENVI in reality different adjustable model parameters can originate different outputs In particular the user can modify the width of the Gaussian kernel function parameter g in K x x exp g x x used to separate class distribution and control the training data individual influence C on classification van der Linden e al 2010a ML and SVM by Canty In addition to standard ML and SVM classification methods also a modified version of both was applied Canty 2010 in fact developed some ENVI IDL extensions for standard classifications with a smart consideration for accuracy assessment Normally the confusion matrix chapter 5 4 3 is used to evaluate the classified map accuracy ENVI built in procedure e uses pixels from all training areas to calculate average band values per each class e these averages are used to classify the
117. there are usually 2 species coffee and one tree species whereas in a system at 23 strata they are usually more coffee and several tree species and or Musa spp Cloud and shadow classes are absent in the legend This was clear from simple visual inspection of the MASTER image though the decision became definitive after evaluating the results of preliminary unsupervised classifications Burned land usually sugar cane fields could also not be detected from visual inspection of the MASTER and GeoEye 1 images even if likely present on the ground Other potential land use classes e g mixed shrubs grassland forest clear cuts were excluded as no matching data were observed A draft of LUC classification was drawn before going to the field though the most suitable classes were delineated only after data collection The FAO Di Gregorio et al 1998 regards this classification scheme as a posteriori approach like the Braun Blanquet system The disadvantages of this approach are the limited application scale of the legend on wider landscapes site specificity and the risk of incurring in ambiguous field notes However for the identification of agroforestry systems in the selected area and of specific land use types not recognizable from the MASTER image a posteriori classification proved to be the most suitable For example distinction between cropping land with trees at 2 and at 23 strata could be defined only during data colle
118. time frame is used 3 Includes forest roads firebreaks and other small open areas forest in national parks nature reserves and other protected areas such as those of specific environmental scientific historical cultural or spiritual interest 4 Includes windbreaks shelterbelts and corridors of trees with an area of more than 0 5 hectares and width of more than 20 meters 5 Includes abandoned shifting cultivation land with a regeneration of trees that have or is expected to reach a canopy cover of 10 percent and tree height of 5 meters Includes areas with mangroves in tidal zones regardless whether this area is classified as land area or not Includes rubber wood cork oak and Christmas tree plantations Includes areas with bamboo and palms provided that land use height and canopy cover criteria are met ne Be ee Excludes tree stands in agricultural production systems such as fruit tree plantations oil palm plantations and agroforestry systems when crops are grown under tree cover Note Some agroforestry systems such as the Taungya system where crops are grown only during the first years of the forest rotation should be classified as forest OTHER WOODED LAND Land not classified as Forest spanning more than 0 5 hectares with trees higher than 5 meters anda canopy cover of 5 10 percent or trees able to reach these thresholds in situ or with a combined cover of shrubs bushes and trees above 10 percent It does no
119. tion Sugar cane plantation Non agricultural land covered with herbaceous species and 10 stree covers60 Non agricultural land covered with herbaceous species and lt 10 tree cover Land with lt 10 vegetation cover and exposed soil sand or rocks Roads covered with asphalt Buildings or other similar man made structures Visible surface water Table 4 Legend of land use and cover LUC classes applied to classify the MASTER image over the study site of 7 x 7 km near Turrialba Costa Rica at a spatial resolution of 10 m Agricultural land is defined on the basis of produced crop type and excludes tree plantations FAO 2000b Other thresholds are taken from the FAO or from the IGBP LCCSs FRA 2010 13 Silvopastoral systems are included in the class mixed tree grassland Although they are agroforestry systems they are not considered in a separate class in this thesis due to the spectral similarity with for example urban greenery Chapter 4 Land use and cover LUC classes 18 Musa spp are often integrated in these shade coffee cropping systems However due to the reduced sample size and resolution of the MASTER scene it was not possible to infer on the presence or absence of banana and plantains in shade coffee plantations only from remote sensing Thus we will limit the agroforestry class definition to shade coffee at 2 or at more vegetation strata implying that in a 2 strata agroforestry systems
120. tion contained in the confusion matrix and corrects the classification rate for the probability of classifying pixels correctly by chance Canty 2010 Alternatively to the equation given in Fig 8 k can be defined as Pr correct classification Pr chance classification a 1 Pr chance classification where Pr stands for probability Tso and Mather 2009 K is considered somehow more accurate than other indexes because provides two statistical tests of significance into one It indicates in fact the degree of significant difference between the result of the classification and a random classification Second it can be used as reference for comparing two matrices as in the case of classification algorithm evaluation Although Stehman 1992 points that k could slightly lose its assessing power when used in context other than simple random sampling other authors Congalton and Green 1999 Foody 2000 Li and Moon 2004 used it for accuracy assessment also in systematic sampling The confusion matrix accuracy assessment can be applied to both supervised and unsupervised classifications Instead of the confusion matrix it is possible to obtain an accuracy assessment also by reiteratively repeat the leave one out test Gretton et al 2001 Kearns and Ron 1999 not applied in this study though 5 5 Post classification processing Some post classification tools in ENVI can be used to improve the accuracy These functions are
121. tor files according to the LUC class by opening the Attribute Table and select the polygons with the same attribute field Note when a pixel cluster lays inside another land use cluster e g a building in the middle of a forest draw the larger polygon around the smaller one It is important in fact to maintain active the option Avoid overlapping the vector layers in order to avoid pixel miscounting c Update polygon shape LUC value Start a new Quantum GIS project Open the SHP file with the quadrants and the digitized land cover polygons Go to File gt Save as and type in a new file name In this way you create a copy of the original SHP file without altering the latter Modify the shape of the wrong polygon s using the Move feature s and Node tool from the main menu bar If interested annotate of the number of changes Modify the LUC by adding a column to the attribute table and insert the correct value the Toggle editing mode should be active Remember to save your work regularly d Upload GPS points into vector file Connect the GPS receiver to the computer Open DNR Garmin Click File Load from gt File Browse to the GPX file with the recorded waypoints make sure that the file extension is set to GPX Chapter 10 Appendices 65 Select the file and click Open As Feature Type chose Waypoint The waypoints and their associated metadata are uploaded in the Data Table
122. tree gr Total Class Unclassified tree crown co barren land settlement water sh coffee sun coffee sugar cane grassland mixed tree gr Total Ground Truth Percent Class tree crown co barren land 1 60 1 20 65 14 0 00 0 33 88 43 2 82 0 00 0 00 0 00 19 51 5 54 0 48 0 00 2 20 2 0 03 0 00 tf Ab 2 65 100 00 100 00 Ground Truth Percent sun coffee sugar cane 0 00 oz 0 00 2 10 0 00 0929 0 00 omi 0 00 0 00 0 78 5 67 EJE a CA 0 25 0 00 81 53 0 00 0 00 0 00 9 84 100 00 100 00 settlement co OWNJOOOoOHrHMmMoooHr Hr 07 06 06 43 07 aul O0 325 50 O4 00 water 0 00 0 00 0 00 0 00 98 68 0 00 0 00 0 00 0 00 I g E 100 00 grasslandmixed tree gr OOOO0000000 e eS amp oO 00 00 00 00 00 00 00 00 00 00 00 2 68 7 94 0 10 8 17 0 00 1323 0 10 ae 0 00 60 20 100 00 sh coffee 1 48 oo ono o Ol iO nN 425 222 60 D0 88 433 85 D0 40 D0 Fig 18 Confusion matrix of sieved and clumped MASTER image classified with ML modified algorithm after 2 LUC classes have been merged The shown part of the confusion matrix indicates the matches between the testing column and the training row areas expressed in percentage Chapter 7 Discussion 47 7 Discussion 7 1 Sampling design Implementing a systematic sampling scheme has several advantages it requires no prior knowledge on the area it is simple to explai
123. un supervised mixed classifications Gomez ef al 2010 e combine soil data for land use potentials and correlation finding In addition soil agroforestry and FOC mapping should go arm in arm as the fungi propagates through the soil e extend the study to areas with pure coffee and pure banana plantations and larger agroforestry range for comparison e test of Canty s CT RUN accuracy assessment against ENVI confusion matrix Chapter 10 Appendices 10 Appendices 10 1 Appendix A FAO Forest Resource Assessment FRA legend FRA 2010 TERM definition and explanatory notes FOREST Land spanning more than 0 5 hectares with trees higher than 5 meters and a canopy cover of more than 10 percent or trees able to reach these thresholds in situ It does not include land that is predominantly under agricultural or urban land use Explanatory notes 1 Forest is determined both by the presence of trees and the absence of other predominant land uses The trees should be able to reach a minimum height of 5 meters in situ 2 Includes areas with young trees that have not yet reached but which are expected to reach a canopy cover of 10 percent and tree height of 5 meters It also includes areas that are temporarily unstocked due to clear cutting as part of a forest management practice or natural disasters and which are expected to be regenerated within 5 years Local conditions may in exceptional cases justify that a longer
124. use criteria 2 Includes groups of trees and scattered trees in agricultural landscapes parks gardens and around buildings provided that area height and canopy cover criteria are met 3 Includes tree stands in agricultural production systems for example in fruit tree plantations and agroforestry systems when crops are grown under tree cover Also includes tree plantations established mainly for other purposes than wood such as oil palm plantations 4 Excludes scattered trees with a canopy cover less than 10 percent small groups of trees covering less than 0 5 hectares and tree lines less than 20 meters wide INLAND WATER BODIES Inland water bodies generally include major rivers lakes and water reservoirs Chapter 10 Appendices 62 10 2 Appendix B Simplified version of Raunkizer s life form classification The following appendix is meant to be consulted in case of doubt when classifying life forms or for the interested reader In the present work we distinguished vegetation types in the following three main categories trees shrubs and herbs grasses With simplifications these categories can fall respectively in Raunkizer s classes Phanerophytae Chamaephyte and Hemi Cryptophytes Raunki r et al 1937 A simplified graphic representation of Raunkizer s life form classification is where 1 Phanerophyte 2 3 Chamaephyte 4 Hemicryptophyte 5 9 Cryptophyte The underling assumpt
125. using other CARTA mission sensors e georectification of MASTER image using the available DEM and topographic correction e for atmospheric correction with FLAASH change of atmospheric parameters possibly retrieved from other sources e other atmospheric correction options ATREM ACOR Dark Object Subtraction use of EFFORT Empirical Line Calibration Flat Field correction IARR Internal Average Relative reflectance calibration Multivariate Alteration Detection MAD QUAAC Temporally Invariant Cluster TIC Chen e al 2005 modules is ERDAS IMAGINE buy a Lever 3A product e investigate on sample size of multispectral geospatial data improve polygon digitization by excluding ambiguous pixels No pixel spatial sub cluster within each polygon should be able to fall into another LUC class e object based classification segmentation object extraction module in ENVI EX called feature extraction module e apply a stratified sampling scheme e feature selection n D Visualizer create new bands e g NDVI or other ratios Plourde and Congalton 2003 perform band correlation analysis e extract endmembers Pixel Purity Index PPI SMAAC Chapter 9 Outlooks 5 e investigate potential spatial correlation between classes e g through the bootstrap method McRoberts and Meneguzzo 2007 e for classification try Spectral Angle Mapper SAM Artificial Neural Networks fuzzy methods Decision Trees
126. utomatically excluded from the reflectance image 5 3 3 Feature selection Alternatively or complementary to atmospheric correction it is possible to select the spectral bands that are most relevant for land use class determination Normally noisy bands will be discarded in this process which is useful for reducing the size and processing time of large spectral data sets Two are the selection approaches taken in this study orthogonal transformation PCA and MNF and out to out band selection with R and EnMAP Toolbox Before proceeding in describing them few definitions are clarified Spectral bands can be regarded as features as features describe an object Each pixel the objects can thus be represented as a point in a k dimensional plot called feature space with amp being the number of features Tso and Mather 2009 Orthogonal transformations transform the feature space into a new set of axes with lower correlation These axes are orthogonal to the originals and start at the data mean The new number of axes can be 19 MODTRAN is a highly accurate program developed by Spectral Sciences Inc which models the electromagnetic propagation of atmospheric radiation Berk e al 1998 Currently the latest available version is MODTRANS Chapter 5 Methodology 28 lower or equal to the original data set however since the first axes display the highest decorrelation it is often preferred to reduce the number of feature
127. y save them in evf format if you haven t done so yet In the Available Vector List window choose Select All Layers gt File gt Export Active Layers to ROIs As image file to associate choose the MASTER orthorectified and spatially subset image In the Export EVF Layers to ROI choose Convert all records of an EVF Layer to one ROI In the Available Bands List right click on the MASTER image and choose Load True Color In the main menu bar Basic Tools gt Region of Interest gt ROI Tools You can now change 33 Currently En MAP is able to read files with an ENVI type header For more info see the Classification Manual van der Linden e al 2010a Chapter 10 Appendices 78 the colors of the classes by right clicking on the color names 6 From the main menu bar Classification gt Create Class Image form ROI Select the MASTER image Select all ROIs Finally choose an appropriate path and file name This corresponds to the so called file of the Alpine_foreland En MAP tutorials Feature selection With the just created image file you can now proceed to feature selection Please follow chapter 2 of the EnMAP Box Regression Tutorial Suess e al 2010 You will create in this way a spectrally subset and scaled image to use for SVM classification Band selection is based on mainly visual assessment therefore not very practical for hyperspectral data without prior knowledge on likely atmospheric and spectral windows

Land use and cover classification using airborne MASTER and

Contents

Download Pdf Manuals

Related Search

Related Contents