PDF lab module attached HERE
Contents
1. LLU ESSE21 Land Use Land Cover Classification Module Honduras Land Use Land Cover Case Study ESSE 21 Land Use Land Cover Classification Module Table of Contents Overview of Land Use Land Cover Classification 3 Description of Study Area 4 Land Use Land Cover Classes 5 Overview of ERDAS Imagine Tools and Menus 6 Classification Process Summary 8 Image Preparation 8 Classification Algorithms 9 Supervised Classification 9 Training Sites Creation 10 Training site Evaluation 11 Band Selection 20 Apply Decision Rule 21 Evaluation 21 Recode and Smooth 22 Accuracy Assessment 23 Unsupervised Classification 25 Clustering Algorithms 25 Cluster Analysis 26 Resolving Problem Classes 27 Recode and Smooth 28 Accuracy Assessment 28 Questions 29 Example Exercises 29 References 30 ESSE 21 Land Use Land Cover Classification Module Overview of Land Use Land Cover Classification Digital image data are frequently the basis to derive land use land cover information over large areas Classification of image data is the process where individual pixels that represent the radiance detected at the sensor are assigned to thematic classes As a result the image is transformed from continuous values generally measured as digital numbers DN or brightness values BV to discrete values that represent the classes of interest Traditionally the algorithms employed for this process differentiate and assign pixels based2. ESSE 21 Land Use Land Cover Classification Module gt edit parallelpiped limits gt set either max min or 2 standard deviations Check box to indicate overlap and leave white as color to highlight signature overlap ld Viewer 3 Alarm Mask OB File Utility View AOI Raster Help SeDHESPVHU Kue xQQu PA 494278 14 1749430 02 UTM WGS 84 The colors displayed over the image data bands 4 3 2 are those assigned to the training data If no colors are displayed over the original data then the pixels do not fall within boundaries determined by the training data and additional training sites may need to be defined The image alarm shows that some of small holdings and pasture areas within the lowland plain have not been well characterized by the training sites and additional sites are needed The image alarm also indicates that the offshore water needs additional signatures Also remember that the software will assign each pixel in the image to a class whether or not you have designated an appropriate class For example because clouds are visible in the image the classifier will place them in the spectral class to which there is the best match To avoid confusing clouds with other classes such as urban concrete or agriculture bare soil or dirt roads 19 ESSE 21 Land Use Land Cover Classification Module it is best to create training sites for the clouds Thus you may have training sites that are not
3. gt Operators Choose subtraction from pull down menu assign new file name for output OR ii Create a model using Modeler tool The inputs will be your mask original clustered image and your re clustered image Use a conditional function statement to select the original clustered image if mask equals 0 or the re clustered file if mask equals 1 4 Recode and Smoothing a As in the supervised classified you must now group all of the spectral clusters into the information classes Again you should use a unique numeric value code for each information class and these values should be the same values that you used for the supervised classification In main menu click on Interpreter gt GIS Analysis gt Recode b To remove speckle due to isolated pixel and create more contiguous classes run a majority filter over the recoded classified image In main menu click Interpreter gt GIS Analysis gt Neighborhood Functions 5 Accuracy Assessment see description under supervised classification 28 ESSE 21 Land Use Land Cover Classification Module Questions 1 How separable were your signatures based on each of the measures that you used for evaluation 2 Which signatures training sites gave you the most difficulty Why 3 Did you need to add signatures to capture all of the variability within the scene 4 Which bands did you select for the classification 5 How did results compare between supervised a
4. 84 Manual digitizing of African oil palm training sites in yellow 2 Evaluation you should review your training sites to ensure that 1 each is homogeneous 2 that all classes in the image data have been captured and the 3 training sites are spectrally separable Various tools in the software will allow you to make those assessments After your evaluation you may 11 ESSE 21 Land Use Land Cover Classification Module wish to delete perhaps merge very small training sites or add more signatures In evaluating separability between training sites remember the distinction between information classes and spectral classes That is a single information class will be represented by multiple spectral classes that may not probably will not have good spectral separability However you should strive for good spectral separability among training sites signatures that represent different information classes The tools in Imagine will also allow you to select those bands that are most suited to your classification Typically not all bands are used because of the high degree of redundancy in image bands e g visible bands tend to be highly correlated Use of feature space plots and transformed divergence analysis will provide insight into how many and which bands are best a Modality and Variation Evaluation using Histogram tool for a Single Band In signature editor tools click on histogram icon gt in dialog click buttons f
5. On Imagine Main top menu click Classifier gt Signature Editor Be sure to name the file and save it occasionally Outline training sites groups of homogeneous pixels that represent a given class by a clicking on the Viewer menu bar AOI gt Tools gt then in tool palette window on Squashed Polygon tool b Select your area within the image for the signature Left click to add vertices double click to finish Use the viewer magnification and reduction tools to enlarge or reduce image as needed and ensure that the pixels within the training site are homogenous c Add the outlined area to your signature editor using i L A plus arrow icon or under signature editor menu Edit gt Add As you add your signatures you should enter a meaningful signature name Itis also helpful to assign a color that corresponds to the information class of this signature right click on color box next to signature name and select color You will need to be able to identify the training sites as you evaluate them prior to classification and recoding following classification Continue adding signatures until you have examples of all possible variations for each information class For example forests in mountainous regions will vary considerably in their spectral response because slope and aspect affect illumination and will change spectral properties of the forest In addition you may need to add training sites for anomalies such as
6. The feature space plots and transformed divergence tools are useful to choose which and how many bands should be used for the classification Redundancy and correlation between two bands are easily assessed visually using feature space plots The transformed divergence analysis is more effective as a quantitative measure of the separability using different number of bands and band combinations a Select bands Open your training signature file in the signature editor if not already open Select the bands for the classification by clicking on signature editor menu Edit gt Layer Selection Highlight the bands that you wish to use use shift key to highlight multiple bands b SAVE your signature file 20 ESSE 21 Land Use Land Cover Classification Module 4 Apply decision rule a Select decision rule From the signature editor menu click Classify gt Supervised Enter aname for your output file and check box and enter a name for an output distance file The software provides choices for the rules that will be used to assign image pixels to one of the spectral classes represented by the training sites Generally non parametric rules are faster but less robust than parametric rules These partition the feature space into regions and assign pixels if they fall within one of the defined region A non parametric rule can be used where little ambiguity exists for assigning a pixel to a class while the more rigorous but slower p
7. an individual pixel match the class to which the pixel was assigned If the match is close or good then the distance is small and will appear dark in the image If the match is poor or far then the distance between that pixel and its class is longer These pixels will be much brighter The brighter the output pixel the higher the likelihood that it has been assigned to an incorrect class 21 ESSE 21 Land Use Land Cover Classification Module A technique called thresholding can be used to interactively isolate the pixels in the image that have a long distance i e have a higher likelihood of misclassification The process of thresholding uses a chi square distribution of the distance measurements and isolates pixels in the tail of the histogram The user can interactively determine where the distance threshold should be set Once a threshold based on distance is set all pixels that exceed that threshold are set to zero The user can then reassign them to a correct category based on some other criteria ERDAS Imagine threshold tools are under the classifier menu gt threshold 6 Recode and Smoothing a In this step you will group together all of the spectral signature classes that you created to characterize each of the information classes This is the process of assigning a new class the information class to the spectral classes generated from the classification Each of the information classes must have a
8. clouds or cloud shadows to prevent these areas being classified as one of the legitimate land cover categories Region Growing Automatically generates a training site from a representative seed pixel a From viewer menu click on AOI gt seed properties In the region growing properties dialog select the 9 pixel neighborhood 1000 pixel area 3 pixel search distance Euclidean distance to mean center of 8 10 ESSE 21 Land Use Land Cover Classification Module ls Region Growing Properties Neighborhood Geographic Constraints Es Sie 100000 fatpinets y Constrain by region area H Constrain Constrain by region area region area Spectral Euclidean Distance 8 00 Grow at Inquire Set Constraint AOI Options Redo Invert Close Help b Select region growing tool from AOI palette eyedropper S and click on a representative pixel for your signature After the Add the signature generated to signature editor After you add a signature using the region growing tool look at the pixel count Eliminate any signatures within pixel counts of less than number of input bands plus 1 Using both seed pixel and manual methods generate enough signatures to adequately represent all classes and their variability Viewer 3 2003_03_06_sub img Layer_4 Layer_3 L EBR File Utility View AOI Raster Help 2 EDak 1 xqaQ 492907 16 1733686 87 UTM WGS
9. explore the data by a Examining band histograms and statistics b Determine correlation of bands using scatterplots c Calculate correlation for pairs of wavelength bands II A Supervised Classification 1 Select training sites image signatures the quality of a supervised classification depends on the quality of the training sites Particular care should be taken as outlined below to create evaluate and edit training sites a Characteristics training sites are spectral signatures and the terms are used interchangeably An individual training site must be i Contiguous group of pixels that is representative of a class ii Homogeneous as possible When examining the histogram for the training site it should be unimodal and have a relatively narrow range of values low standard deviation and iii Be comprised of n 1 number of pixels where n is the number of bands in the image As a group the training sites must iv Capture all of the variability within an information class While an individual training site should be homogeneous with little variance multiple training sites will be required so that all ESSE 21 Land Use Land Cover Classification Module of the possible variation present in a given information class is captured A single information land use class will be represented by multiple training sites signatures land cover b Creating training sites i 11 Manual Delineation Manual Digitizing
10. of interest in your classification scheme but are required to deal with the actual phenomena present in the imagery The image alarm also indicates where potential confusion among classes may occur The sprinkling of white is an indication of overlapping signatures for example between upland and lowland forests f Edit signatures based on your analysis delete merge or add training sites as needed In addition determine which and how many bands should be used for your classification Each of the various tools used to examine and evaluate the training data provides different insights Visually you are able to assess the homogeneity and distribution of the training sites using histograms statistics and the ellipses You can also use the histograms and ellipses to examine overlap vs separability of signatures in individual bands and band pairs with these tools Quantitative information on the separability of pairs of signatures is provided by the transformed divergence analysis The confusion matrix also provides quantitative data by showing which signatures are likely to be confused Finally the image alarm generates a quick pre classification This is also a visual tool that gives an overview of where the classes will be assigned in the image and whether additional classes are required SAVE your edited signatures to a file with a sig extension from signature editor menu File gt Save or Save As 3 Select Bands
11. on the values recorded at that pixel for each of the wavelength regions bands in which the sensor records data A universal land use land cover classification system does not exist Instead a number have been developed to reflect the needs of different user s Typically the systems are hierarchically arranged with the ability to consolidate lower level classes into the next highest level and with a consistent detail for all classes at a given level in the hierarchy In selecting a classification system for use with remotely sensed data the classes must have a surface expression in the electromagnetic spectrum For example a crop such as pineapples reflects electromagnetic radiation but an automatic teller machine ATM on the side of a building cannot easily be detected especially from most down looking sensors In addition the resolution characteristics of the imagery selected must be compatible with the classification system That is the imagery must have the spatial detail spectral discrimination and sensitivity and temporal characteristics required for the classes of interest For purposes of applying the system for image classification we must differentiate the information classes represented by the land use land cover classification system from the spectral classes that we can obtain from the imagery Often an information class will have a range of spectral responses that represent the inherent variability within a class that is inte
12. age with corresponding ground reference data One dimension of the matrix is the ground reference data the other dimension are the corresponding classified pixels 23 ESSE 21 Land Use Land Cover Classification Module i 11 111 1v Overall accuracy percent of number of correctly classified pixels to total pixels in sample Errors of omission commission Accuracy by class Within classes the accuracy may vary substantially from the average or overall accuracy Two analyses using the error matrix can provide additional insight Errors of omission determine the total number of reference pixels within a class that were classified corrected The reference pixels that were not assigned to the correct class are omitted Errors of commission determine the number of pixels that were correctly assigned in the classified image output class Pixels that were included in a class that should not have been are errors of commission Kappa analysis The K hat statistic generated by Kappa analysis is used to determine whether the classification results are different better than results that could be achieved by chance Perform an accuracy assessment a Generate Sample from Image On main menu click Classifier gt Accuracy Assessment gt File gt Open the file for your final recoded classification In Accuracy Assessment dialog menu click View gt Select Viewer and click in the viewer where you have displaye
13. arametric rule is better used to resolve similar spectral responses that employs statistics The parametric rules use statistics of the training sites as the basis for assigning pixels to a class The three primary decision rules in ERDAS Imagine are 1 minimum distance 2 Mahalanobis and 3 maximum likelihood As indicated the first two assign based on training sites that are the shortest distance in spectral feature space from the pixel Maximum likelihood assigns pixels based on probability that the pixel belongs to a class Maximum likelihood is the most robust of the three but does require that the training signatures are normally distributed and is also slower than other algorithms For the parametric rule use parallelpiped and maximum likelihood decision rule for parametric rule Accept the defaults of parametric rule for overlap and unclassified rule 5 Evaluate results a display your output file using the pseudo color option under raster options from viewer menu bar click File gt Open gt Raster Layer gt in Select Layer To Add dialog enter file name gt click Raster Options tab gt from pull down menu for display select pseudo color gt OK The colors in the output classified image correspond to those assigned when you created the training sites In a separate viewer display the output distance file The output distance file is a grey scale image that measures how closely the spectral characteristics of
14. art with an arbitrary assignment of initial clusters in the image data Individual pixels are then assigned to the cluster to whose mean is closest in spectral space The cluster means are then recalculated based on the new assignment of image pixels The process is then repeated each pixel is compared to the recalculated cluster mean and assigned to the cluster which is closest spectrally then cluster statistics are recalculated This iteration continues until one of two criteria set by the user is reached a A convergence criterion specifies some maximum number of pixels that are allowed to change cluster assignment If that number of pixels or fewer change between iterations the clustering is said to have converged In some cases this convergence will never occur and the process needs to be stopped based on performing the user specified number of iterations In addition the user must specify the number of classes that should be created in the clustering As you saw in the supervised classification you must specify enough spectral classes to cover all of the variability within an information class Therefore when specifying the number of classes you should overestimate rather than underestimate Cluster image click on main menu Classifier gt Signature Editor or Unsupervised Classification You can initiate the unsupervised classification from the signature editor menu if you wish to use some subset of the image bands but not
15. create a separate image However you must also have a signature file sig associated with the image If you have created a separate image with bands of interest then you may initiate the clustering from the classifier menu Classifier gt Unsupervised Classification 25 ESSE 21 Land Use Land Cover Classification Module Imagine uses the ISODATA Iterative Self Organizing Data Analysis Technique for unsupervised clustering The algorithm is robust and has the advantage that the clusters generated are not biased to any particular location in the image In the Unsupervised Classification dialog enter names for your output cluster layer and output signature file sig Under clustering options the initialize from statistics box should be ON The minimum number of classes should be based on your experience with training data sets but probably at least 50 60 Set maximum iterations to 20 and leave convergence threshold at 0 95 Click OK once all relevant information is filled in Note that unsupervised clustering can be used to create a classified image AND generate spectral signatures This can be an effective means to create and or training sites for supervised classification The software is effective at discriminating subtle spectral differences especially for classes that may be difficult for the user to define 2 Cluster Analysis After the image has been clustered the user must identify and label the clusters If
16. d your final recoded classification Select edit create add random points and continue until you have the number of points needed This will create a random distribution of points in your classified image b Collect Enter Ground Reference Points In Accuracy Assessment dialog click on Edit gt Show Class Values Reference values may be obtained from some other image e g high spatial resolution aerial photography or GIS data layer e g a prior land use classification Once reference values have been entered the color of the points should change from white to yellow c Generate an accuracy report Options include an error matrix accuracy totals and kappa statistics Be aware that the matrix can be quite lengthy To create report from dialog click Report gt Options gt Accuracy Report 24 ESSE 21 Land Use Land Cover Classification Module II B Unsupervised Classification As an alternative to a user designating the training sites for classification you can use software to find group the image data into spectrally homogeneous clusters In addition to the image preparation outlined above you should also evaluate the spectral response and correlation among the input image bands This will allow you to select the optimal image bands for the classification 1 Clustering Algorithms While hundreds of clustering algorithms have been developed certain characteristics are common Generally all of the algorithms st
17. discrimination but also may be source of confusions between training sites 12 ESSE 21 Land Use Land Cover Classification Module iii Discrimination for the three cover types bare soil growing crop and water is better in bands 4 6 than bands 1 3 where there is overlap among all of the categories iv The four water signatures exhibit overlap but since this allows you to capture the variability within the information class the overlap will not a problem However the overlap of near shore water signature with vegetation and bare soil is a problem This coupled with its variability and lack of mode makes it an unsuitable training site ks Histogram Plot Control Panel E ts Signature Editor 2003sig sig Plot Options File Edit View he e Cia Signatures E Bands Single Signature Single Band i Green Blue Value Order Count Pro All Selected Signatures All Bands 23 r i ES ts sig pineapple bare soil sig 1 Band Number histogram histogram histogram LIL 177 007 Sel x ws sigl pineapple bare soil sig 1 D T Band Number 6 histogram 12 a a a 92 0128 124 009 4 84 0128 Print Save Help a Print Save Help al Print Save Help hl b Separability and Band Correlation based on Feature space scatterplots for Pairs of Bands In signature editor click on feature gt create feature space layers Use the subset as the input file check output to
18. ed red and green wavelengths bands 4 3 2 PU The data set is a subset of 1633 x 1280 pixels from ETM acquired on March 6 2003 The subset contains bands 1 5 and 7 and has been registered to UTM zone 16 WGS 84 ESSE 21 Land Use Land Cover Classification Module Land Use Land Cover Classes Evaluacion Nacional Forestal de Honduras Level I Level II Level HI Forest Lowland Broadleaf Coniferous Mixed Forest Mangrove Nonforest Other natural areas w woody cover Shrubs Pasture w trees Savanna w trees Other lands w out trees except agri forest Natural pasture Savanna Wetlands Bare soil Agri forest Annual crop Permanent crop Animal husbandry Human settlement At level IV and lower forest classes are grouped further by age class then cover From Carla Ramirez Zea y Julio Salgado eds Manual para levantamiento de campo para la Evaluaci n Nacional Forestal Honduras 2005 ESSE 21 Land Use Land Cover Classification Module ERDAS Imagine 9 0 Throughout this module the classification processes and routines described are from LeicaGeosystems ERDAS Imagine 9 0 image processing software This software operates through a series of menus that open dialog boxes tools bars and editors Each editor and viewer window will have its own menus and tools bars For image classification the Menu bar clicks and command sequences are identified in the text in bold italics Parameters for dialog boxes are listed in ord
19. er but vary when that input occurs in the process In a supervised classification the user identifies spectral signatures that are representative of the classes of interest A decision rule is then implemented that will assign each pixel in the image to a class based on how closely it matches the spectral characteristics of the input spectral signatures also called training sites In contrast an unsupervised classification allows the computer to cluster partition the image into spectrally homogeneous clusters The user then assigns a class name to those clusters The following outlines the steps to perform first a supervised classification then an unsupervised using Imagine image processing software Successful classification requires knowledge of the area of interest and the spectral reflectance characteristics of the land use land cover classes Before undertaking the classification you should familiarize yourself with the area and the imagery Information and images of the north coastal region of Honduras with a virtual tour can be found starting at http resweb llu edu rford ESSE21 LUCCModule see Introduction Compare this information with the image data and the land use classes used for the Honduras data set so that you recognize the land use land cover categories of interest in the imagery and understand how that response varies with wavelength In addition to visual examination of the imagery and ancillary non image data you can further
20. er of entry Below are the primary tools that you will use Main Menu Ms ERDAS IMAGINE 9 0 Session Main Tools Utilities Help gt Viewer Import DataPrep Composer Interpreter Catalog io r Modeler Vector Si AutoSync For greater efficiency 1t is useful to set location for default input and output directories by clicking on Session menu gt Preferences Imagine offers two different viewers for displaying imagery Classic Viewer or Geospatial Light Table Classic Image Viewer with Menu and Tool Bar ld Viewer 2 Jm fx File Utility View AOI Help c w DESA gaam HAKE KQAQ PF 260 00 23 00 Viewing an image click on File gt Open gt Raster Layer This opens a dialog box in which you enter or browse for the image file name If you click on the Raster Options tab that control display characteristics including image bands size of image etc You may open multiple classic viewers and display different images or band combinations of the same image ESSE 21 Land Use Land Cover Classification Module Geospatial Light Table ls Untitled GLT Viewer File Utility View AOI Interpreter Help 2 EDDSZ BRAA gN enun A P l 65 Pl AAR Display Enhance Zoom Rotate Roam General y OR O09 O66 tke tO ae KOE _ 0pm 027 Of 300 Anm a s Zatut p Opm 0 2 MEA ED ASSAR This performs the same functions as the image viewer b
21. ign a new output name Use Majority from the pull down menu for the Function Definition This will apply a moving window filter that will assign the value that occurs most often within the window to the center pixel within the window The default window size is 3x3 but can be adjusted depending on your image and data 7 Accuracy Assessment Land use land cover data are used for a variety of purposes and consequently some understanding is needed as to how accurately it represents reality A comparison of a random sample of the classified data with ground reference data is the generally the basis of accuracy assessments a Sample the sample must be of adequate size for amount of variation present in the imagery and desired level of confidence Frequently this requires very large samples As a rule of thumb Jensen 2005 cites studies that suggest 50 samples per class are adequate The sampling design is also important Random sampling is often recommended to achieve a representative sample However this may under sample classes with few members In this case some form of stratified sampling may be required The sample should not include pixels used for training sites Instead a different set of pixels should be generated for accuracy assessment b Error matrix An error matrix is similar to the contingency matrix described in evaluation of training sites The matrix provides a cross comparison of pixels taken from the classified im
22. ing input classes will be confused Again if the two training classes represent variability within the same information class confusion is not a problem In this example most of the training site pixels were correctly assigned Some confusion was found between lowland and upland forests 15 ESSE 21 Land Use Land Cover Classification Module ls Editor Dir File Edit Yiew Find Help o2Oo Ss tA ERROR MATRIX Reference Data Classified sigl pin water of water ne sigl pin sigl pin water of water ne siga du sig 5 lo 0o0o0o0o0ocooo low growth Column Total Classified siga af pasture 0 00 0 00 0 00 0 00 0 00 0 00 100 00 0 00 0 00 0 00 sigl pin sigl pin water of water ne sig a du sig 5 lo siga af pasture upland for low growth m o 0o0o0o0o0o0oooOo D 0 D 0 D 0 D 0 D 0 D 7 D 0 0 0 0 2 0 0 Column Total 25 Reference Data Classified Data upland for low growth Row Total sigl pin sigl pin water of water ne sigda du sig 5 lo siga af pasture upland for low growth OJOONOOOOO 0o0o0o0oooooo Hm o Column Total 16 ESSE 21 Land Use Land Cover Classification Module d Quantitative Separability and Band Selection from Transformed Divergence calculation using All Bands In signature editor menu click evaluate gt separability Use options for transfo
23. nd on the application Typically this may include 1 Image import ingest imagery from its transfer format e g TIFF to format used by the software e g Imagine img file 2 Image registration rectification usually performed to geometrically register to a coordinate system and to remove geometric distortions if present 3 Image subset limit the processing to the area of interest 4 Radiometric correction may be required if sensor anomalies are present or if atmospheric contamination is severe Removal of atmospheric effects can be difficult and time consuming and generally will not improve classification results unless the effects are non uniform across the scene and or extreme e g visible haze 5 Data transform discrimination among classes may be enhanced by creating new spectral bands based on some combination of the image data A variety of transforms can be used for this purpose including bands derived from a principle components analysis PCA indices used for vegetation moisture or geologic properties and linear transforms such as the tasseled cap transform 6 Layer stack combining bands of the original image with transformed data to create a new image ESSE 21 Land Use Land Cover Classification Module II Classification Algorithms The traditional spectral based classification is approached by using one of two methods supervised or unsupervised Both require input from a knowledgeable us
24. nd unsupervised approaches Which classes were more successfully classified under each approach Other Examples of Classification Exercises http www cas sc edu geog rslab 751 e10 html John Jensen Image Classification Exercise http www cas sc edu geog rslab Rscc fmod7 html Remote Sensing Core Curriculum http www nr usu edu Geography Department rsgis Remsen1 ex5 ex5 html Utah State University Exercises 5 8 29 ESSE 21 Land Use Land Cover Classification Module References Jensen John R 2005 Introductory Digital Image Processing New York Prentice Hall DiGregorio Antonio and Louisa J M Jansen 1998 Land Cover Classification System LCCS Classification Concepts and User Manual Rome Food and Agriculture Association of the United Nations Carla Ramirez Zea y Julio Salgado eds Manual para levantamiento de campo para la Evaluaci n Nacional Forestal Honduras 2005 Financiado por la Organizaci n de Naciones Unidas para la Agricultura y la alimentaci n a trav s del Proyecto de Apoyo a la evaluaci n e inventario de bosques y rboles TCP HON 3001 A 30
25. nded to capture like activities This may be due to composition of covers that are necessary to express the class e g residential class would include materials for roads lawns gardens rooftops and other building materials Or multiple land covers may individually satisfy the criteria for a given land use class e g crops of barley corn lettuce sugar beets etc are all in the class field crop but would have different spectral responses The diurnal and seasonal aspect also contributes to spectral variability due to variation in planting dates vegetation phenology and illumination Thus multiple spectral classes called signatures or training sites that capture the variability are required to represent a single information class ESSE 21 Land Use Land Cover Classification Module Description of Study Area In this module we work through techniques to classify a portion of the North Coast of Honduras using Enhanced Thematic Mapper imagery for March 2003 The classification scheme is one used by Forestry Department in Honduras and is based on the FAO Land Cover classification system For purposes of this exercise we are working with 30m spatial resolution data and that is consistent with Level II of the system and in some cases Level III The image processing routines described are based on Leica Geosystems ERDAS Imagine software AE eget tt ale Northern Honduras from Enhanced Thematic Mapper on March 6 2003 in near infrar
26. omething that will stand out from the background right click on the color column for that cluster This will display the color chosen for the cluster over the image Enter a label Change opacity back to 0 iv Continue toggling opacity between 0 1 assigning colors and adding labels until all clusters have been labeled File Edit Help Row Histogram 12624 12327 forest mtns 13798 forest mtns 14119 forest mtns forest mtns forest mtns forest mtns forest water mix forest water mix forest water mix forest water mix forest water mix forest water mix forest mix forest water Press Left Button to Edit 51 1542 80 1748963 42 UTM WGS 84 Assignment of colors and labels to unsupervised classification using opacity toggle c The blend swipe flicker utilities Viewer menu gt Utilities are another means to overlay clusters on the image 3 Resolve problem classes Typically one or more of the clusters generated will still contain more than one information class In the above example a number of the classes are labeled forest water indicating that confusion has occurred i e two information classes are spectrally similar and have been placed in the same cluster This is particularly likely if the user specifies too few output classes Nonetheless problem clusters are usual and can be resolved through a re iteration of the clustering
27. or all selected signatures and all bands click on plot button to view histograms To view statistics using the signature menu bar click on view gt statistics Use the histogram and statistics to determine if your signatures are homogeneous and normally distributed Discard any multimodal training sites or data that does not approximate a normal distribution Multiple modes indicate that more than one spectral class was captured The training signatures should be approximately normally distributed to satisfy assumptions of parametric decision rules used to assign pixels from the image to an output class By highlighting multiple signatures in the signature editor and clicking on the histogram icon you can display multiple signatures This provides an indication for a given band of whether there is overlap among signatures As noted above you will probably have overlap if the spectral signatures are in the same information class but should NOT if they represent different information classes Determine whether some of your signatures should be discarded The plot on the following pages illustrates several issues i Most of the signatures are unimodal except for the one of the water signatures near shore water that is variable and lacks a mode 11 In bands 5 and 6 the growing pineapple category shows a range of spectral response rather than a clustering around the mean A wider response more variability in one or more bands may aid in
28. process sometimes referred to as cluster busting a Recode the clustered image to two classes 1 Recode all good clusters as O These are the clusters that represent a single information class have no confusion and further analysis is not required 27 ESSE 21 Land Use Land Cover Classification Module ii Label all bad clusters as 1 These are the clusters that represent more than one information class and thus are confused These are the clusters that you will further analyze b Using the recoded image mask the original image data that you used for the unsupervised clustering Click on main menu Interpreter gt Utilities gt Mask The input file is the original image data the mask file is the recoded image and assign a name to the output file c Perform unsupervised clustering on the masked image The number of clusters is dependent on how much confusion there was after the first clustering d Analyze and label clusters Repeat process if necessary e Combine the masked new clusters with the good classes from the first clustering You can combine the two layers by i Creating a mask that is the inverse of your original mask and applying it to the first clustering output Your good classes will have their assigned value while bad classes are all set to zero Add this masked clustered image to the results you obtained in step d Main menu gt Interpreter gt Utilities
29. rmed divergence ASCII output and summary report Start with 3 bands per combination but you may want to consider more 4 5 or 6 band combinations The tools will evaluate separability between each pair of signatures for a given number and image bands The best average or minimum transformed divergence calculated can be used to determine how many and which bands to use for classification The report is in 3 sections The header material lists the image file used distance measure image bands and number of bands considered In the next section the training sites signatures are listed In the third section the divergence between each pair of signatures for a given set of bands is listed The signature pairs are listed first then below is the divergence value Values greater than 1900 indicate good separability while values less than 1700 have poor separability In the summary report rather than list all possible band combinations the information is given only for those bands that produce the best average separability and the best minimum separability In the following example the separability of all of the signatures except signatures 10 lowland rainforest and 17 upland rainforest was high This is consistent with the results for the contingency matrix in which these two classes showed some confusion Both the minimum and average separability results indicate that bands 1 and 6 would provide good discrimination The results differed
30. s the large blue ellipses for one of the water training sites This has a large variance overlaps with multiple non water signatures and should be eliminated Feature space is also useful to determine if two bands are highly correlated that is information in the bands are redundant If so use of both bands will probably add little to discrimination of classes of interest 14 ESSE 21 Land Use Land Cover Classification Module do Viewer 18 EANT_cO0856 D E S g ike Bands 1 and 2 Bands 5 and 6 In the above plots the two band pairs shown are highly correlated particularly bands 1 and 2 and will be more limited for feature discrimination Compare this to the spread in scatterplots for band pairs 4 and 6 and pairs 2 and 5 C Quantitative Separability Confusion determined by Contingency matrix computed for All Bands In signature editor cell array select highlight all signatures Then in signature editor menu click evaluate gt contingency Use parallepiped as the non parametric rule parametric rule for overlap and unclassified and minimum distance as parametric rule check on pixel percentages and click ok Print report and save to text file The purpose of the contingency matrix is to classify the training site pixels and assess how many are assigned to the correct class Ideally the output class for the pixels will be the same as the input training class The matrix is useful to determine if two or more train
31. the clusters will be used as training data for supervised classification then the analysis and evaluation of training data described above should also be performed The following tools and techniques can be used to label clusters a Display clustered image in one view and original unclassified imagery in a separate viewer The clustered image will be a grey scale You can automatically assign colors that can aid in distinguishing the classes i Click on view menu Raster gt Atttributes In Attribute dialog click on Edit gt Colors The default options are effective as a first cut IHS Slice Method Slice by Value with Maximal Hue Variation b Alternatively you may display the clustered image OVER the original image data i In Viewer display three bands of the image data Then open the clustered image Use the Raster Options tab and UNCLICK box to clear display The clustered image now appears over the image data ii In View Menu click on Raster gt Attributes In the dialog select all rows right click on column labeled Opacity gt Formula In the formula dialog enter 0 zero in Formula box at bottom of dialog click Apply After setting opacity to zero your original 3 band image data will again appear 26 ESSE 21 Land Use Land Cover Classification Module iii Analyze and label clusters one at a time For a single cluster change the opacity value back to 1 and in color column change the color to s
32. unique numeric value code assigned to it In the main menu click on Interpreter gt GIS Analysis gt Recode Enter the name of your classified image and assign a new name for the output Click on Setup Recode Button This opens up a Thematic Recode window Click on a row in the value column enter the value of the information class in the box at the bottom labeled New Value then click button to Change Selected Rows s Thematic Recode DER Value NewValue Histoqram Red Green Blue A Oof 0 000 0 000 0 000 371526 0 0 827 0 827 0 827 0 827 0 827 10501 0 0 000 0 392 0 000 0 827 0 827 0 000 0 824 0 706 0549 es Change Selected Rows Cancel Help 22 ESSE 21 Land Use Land Cover Classification Module Continue the process of highlighting each of the spectral classes listed entering the new value in the box and clicking Change Selected Rows until all of the spectral classes have been recoded Do NOT attempt to enter the new value in the column labeled New Value as 1t will not be saved When all values have been reassigned click OK gt OK b Smoothing the process of classification may result in isolated pixels and an overall speckled appearance To reassign these stray pixels a smoothing operation can be employed Click on Interpreter gt GIS Analysis gt Neighborhood Functions Enter file names for the classified and recoded image as input and ass
33. ut differs in the interface Many of the tools that can be accessed from menus in the classic viewer are incorporated in the geospatial light table tool bar In addition the geospatial light table will display multiple images in a single screen Either viewer may be used but the examples are based on the classic viewer O aH 0 Squashed Polygon Eyedropper aia ier oF Q ETOSOS O Mr NI ls Signature Editor sig1 sig File Edit View Evaluate Feature Classify Help 2 O e Wilt VA Color Red Green Blue Value Order 0 000 1 0001 0 000 1 000 0 000 0 000 cua pog mog Me dtmMmmME 5 amp M y jiii 4 3 A Signature Editor AOI Tool Palette ESSE 21 Land Use Land Cover Classification Module Classification Process Classification is a multi step undertaking and is summarized in following outline flowchart I Image preparation IL Algorithm Selection A Supervised Develop Training Sites Image Signatures Evaluate training data Band Selection Apply Decision Rule Evaluate classification Recode Assessment B Unsupervised 1 Clustering algorithm and parameters 2 Cluster analysis 3 Resolve problem classes and clean up 4 Recode 5 Assessment NAMAMNMNN I Image preparation Image data must be in a form that can be used for classification A number of processes may be involved and are lumped under the term preprocessing The exact tasks will depe
34. viewers and use default output root name Click OK For each band pair a feature space scatterplot window will be opened that displays the spectral values for the two bands also called a 2 D histogram The frequency of occurrence in the image of a pair of values is indicated by the color That is pairs of values that occur least frequently are in magenta while those that occur most frequently are in yellow to red 13 ESSE 21 Land Use Land Cover Classification Module To plot signature ellipses in the feature space viewers highlight one or more signatures in cell array of the signature editor click in column under Class In the signature editor menu click Feature gt Objects In the dialog box enter the number of a feature space viewer check plot ellipses 2 standard deviations labels The center of the ellipse is the mean value of the signature for the two bands displayed and the size of the ellipse outer boundary represents the variation of the signature You are looking for signature ellipses that 1 have relatively narrow boundaries 2 do not overlap between information classes and 3 cover different regions of the feature space Perform this analysis for each pair of input bands ve Viewer 17 EANT_O00856 2 w DESA 2 3k a 175 23 208 00 Bands 4 and 6 Bands 2 and 5 In the above two feature space plots most of the ellipses do not overlap and exhibit a small degree of variance One exception i
35. with respect to the third band either band 4 in near infrared or band 5 in mid infrared These bands provide somewhat different information with respect to vegetation and moisture and selection should be based on goals of the classification 17 ESSE 21 Land Use Land Cover Classification Module M Editor Dir File Edit View Find Help File c documents and settings sally_westmoreland my documents imagery honduras_subsets 20 Distance measure Transformed Divergence Using bands 123456 Taken 3 at a time Class sigl pineapple bare soil sigl pineapple growing water offshore water nearshore sig a dunes sig 5 lowland rainf sigba african oil palm pasture low veg upland forest low growth crop pasture Best Minimum Separability MIN lass Pairs C 1 E 3 4 Best Average Separability MIN Class Pairs 5 3 Bcd i EL ae AR pd 4 4 17 6 19 10 15 1519 2000 2000 2000 2000 2000 2000 e Completeness of signature set using Image alarm with All Bands To highlight all pixels in the image that are estimated to belong to a class use the image alarm The image alarm performs a quick pre classification of the image data Before using the alarm select distinguishable colors for your signatures see color column if you have not already done so Highlight some or all of your signatures for the evaluation From signature editor menu click view gt image alarm 18
Download Pdf Manuals
Related Search