Home

Общее описание программы NeRIS

1. Create polyline vie Set drawing style Fit to window Set selection style Utilities blank vector map slots importing 00000000 Principal components Open principal components calculation tools emma SOM utilities Classification BL ANN Create working window EI Create a new window for raster display Synchronize windows Synchronize the location of scrollers in all windows according to the current window coordinates of all windows centers are equal to the coordinates of the current window center Vertical Arrange raster display windows vertically Close all windows a Close all windows MANUAL ABOUT THE PROGRAM m About NeRIS application Navigation 62
2. True Color Select Haste f Primary Secondary f 256 Values Blue Show SOM 50 Vectorize Classification to MIF Select File MIF Style Field ict Topology Value E Run Area 0000 Perm 20000 Preview Cancel ok Besides the abovementioned controls the dialog has the group of classification results vectorizing controls and Show SOM button meant for palette control It should be noted that all controls designed for slot values transformation are in the unit transformation status and inhibited which is important for correct classes numbers interpretation Button Show SOM enables to open palette creation dialog using the trained neural network List box Select File enables to select vector layer where the vectorized classification results will be put classes se lected in the palette creation dialog using the trained neural network are subject for vectorizing List box Field enables to select the vector layer attributive table field where the vectorized objects code will be save to field must have integer numbers Checkbox Topology if checked partial topology is built prior to saving vector objects in one object is inside the other bigger one an island is cut out in the outer object This operation takes some time however it is convenient for further vector objects processing area calculation etc
3. Current Layer Sample Field Hame Statist Object Type Region Square 1 139333 53 Count of Arcs 1 Penmeter 153526 46354 Count of Modes 11 Table Field name Value contains the list of the vector layer attributive table fields There are two columns available for each of them e Column with the field name e Editable column with the field value Statistic data is available at the bottom about the selected object object type area perimeter number of arcs and nodes 39 New vector layers creation control Create New MIF File dialog To create new vector layers MIF MID format files in the current software version use the following dialog Create New MIF File n ajx MIF File Name E SampleSROI mif a Field Mame Field Integer Up Down 10 Add Field 1 Delete Field Field Properties M Type Field Length 10 File Name selection button to record new vector layer 15 used to select file type and name only MIF MID format files creation is supported by the current software version Table contains the list of vector layer attributive table fields There are two columns available for each of them e Column Field Name e Column Field Type Selection of the current field in the table enables to edit its properties using the following controls Field Name specifies and mo
4. e 3 channels value True Color means that the display is formed of three slots which values are interpreted accord ingly as red Red green Green and blue Blue color tones of the created images whereas the derived colors are defined by the linear transformations and gamma correction of these values e 1 channel value 256 values means that the display is formed of one slot which values are interpreted either as brightness or as the class number depending on the Type control Type group of radio buttons specifies the slot value interpretation method when making a display e SOM selection means that the values are interpreted as numbers of classes received as a result of neural net classifi cation In this case the colors for each class display are set via the corresponding dialog control e Grey selection means that the values are interpreted as brightness and the grey levels for each value are defined by linear transformation and gamma correction of these values e Legend selection means that the values are interpreted as numbers of classes received as a result of neural net classi fication or image filtration In this case the colors for each class display are set via the corresponding dialog control 37 Making RGB view Dialog used to make RGB display is shown below View Parameters 2 Hed 12 Lz21 7021 02120000905 BU Green 10 21171021 021 20000305 B4 Blue 8 Lzt171 021_02120000905_ 20 255 Gamma
5. Button Save saves neural net to file Button Comments enable to create and edit comments stored for neural net In particular parameters used during the neural net training and calibration are stored here including the slots composition and normalization method Stan dard multi line edit window appears on the screen after pressing this button as shown below The user can add his own comments which are stored into file when neural network 1s recorded Button Teach initiates neural net training Button Calibration initiates calibration of the previously trained neural net 55 Contextual Kohonen neural network classification Classify using Pattern SOM Neural Net dialog This dialog can be opened by selecting Classify using Pattern SOM item in Classification menu and is designed for classification control using previously trained neural net and primary raster slots Input tab contains the group of input parameters for classification using pattern SOM Classify using Pattern 50 Neural Net 3 xl Input ROI Pattem SOM Input Parameters SOM used to create classes Comments Palette was Used While Teaching SOM Classes Plane 7 2111021 02120000905 10 chl warp oF Window Size Field SOM used to create classes specifies the neural net for classification Button Comments enables to view loaded neural net parameters Checkbox Palette Was Used While Teaching must be checked if the palette was used in tr
6. Dimesions and Comer coordinates Lower Lett Come Upper Aight Carne x o o x 100000 0 E Y 100000 0 Define Geo Reference Detine Cancel Count of Pixel bu x 1000 bw T 1000 Fisel Siz by 100 by Y 100 Lower Left Corner fields contain the lower left corner coordinates of the raster being created in the target coordinate system Upper Right Corner fields contain the upper right corner coordinates of the raster being created in the target coordi nate system Pixel Size fields by X contains the horizontal size of the of the created raster pixels by Y contains the vertical size of the of the created raster pixels Count of Pixels fields by X contains the horizontal size of the of the created raster in target coordinate system units by Y contains the vertical size of the of the created raster in target coordinate system units Projection EN mu One Datum arth z 5 1984 z Shift Dos Shift Shift 02 i i Standard parallels 0 0000 0 0000 0 0000 Rotation Rotation Rotation Ez Morth 0 0000 Scale 1 000000 0 000000 0 000000 0 000000 South B5 0 0000 0 00000 scale factor e Prime Meridian 0 000000 0 000000 Origin Spheroid Longitude 0 0000 False Easting 0 000 wes Latitude 0 0000 False Marthing 0 000 63781 37 000 f 298 257
7. NONE The application provides for two selection options using vector map in MIF format as mask and using the slot pre viously classified by another neural net by specifying hierarchy nodes containing the required classes in this case mask control 15 done via the hierarchy tab Checkbox Use Vector Layer as Border indicates the requirement to limit the training area by the regions contained in vector layer file File selection push button enables to select the vector map file containing training area borders Slots names and hierarchy nodes used for mask creation are indicated below If both vector map and hierarchy mask are applied only pixels within both masks are used for training SOM tab enables to display neural net parameters and to load the net from file if the neural net stored in the hierar chy node is not used SUM Classification BEES 7 xj Planes Lll Hierarchy SOM Horizontal Size Vertical Size Comments Use weights Distance Neighbour Alpha Made Save SOM Legend Hierachy Class NONE Button Load SOM loads the neural net from file Button Comments enables to view neural net comments Information about the reference of the current neural net to a certain hierarchy node 1 available in the bottom of the tab 5 Hierarchy tab enables to control the neural net hierarchy corresponding to the previously created hierarchy of classes Neural net 1 on the node o
8. after finding the neural network class to detect which thematic class the pixel 1s re lated to Image structure detection and storing To display the detected structure of the multi layer image after training NeRIS application is using the Sammon s mapping when the classes in the initial p dimensional space are mapped in the planar grid so that the distances between the classes characters are the same as between the classes themselves After normalizing received Sammon s coordinates we obtain the invariant structure of the initial image 1n terms of distance functions used It s easy to see that such struc ture 1s invariant with respect to the linear transformations of the values of individual initial image planes provided this transformation is not putting the values beyond the permissible value range Therefore if we are using Sammon s coordinates for the creation of two new image planes they keep all the data about the initial image structure In this way the trained and thematically calibrated neural network which coordinate classes have been transformed into Sammon s space can be used for thematic interpretation of new images obtained under different atmospheric conditions provided the image 15 also transformed into the Sammon s space Besides Sammon s coordinates can be also used as the compression of the initial image when it 1s transferred to dif ferent classification procedures NeRIS application provides for the assessment
9. 0 Remote sensing data thematic interpretation software ScanEx NeRIS version 4 0 Neural network raster interpretation system 5 NeRI5S Gr kd uo Pen 0279 0 08 all som uer Or bae SOM 0 003 0 25 SOM By SOM E 1 14 0464 Chi 50 Che 10 33 ChE Cha 26 Ch4 15 O O 1 00 07 1 el 247 ______ 330882 34 6200028 49 1 113 327 User s manual Moscow 2005 NERIS SOFTWARE KEY FEATURES cc 4 DATA MODEL T E E IIa III td mM 5 VECTOR DATA EE idi E iis 5 RO 5 R 6 6 Rarer ana ane a ier ees 6 KOHONEN NEURAL NETWORKS SELF ORGANIZED MAPS 7 7 Hed w dioe N r DATA MANIPULATIONS aerea EA 8 DISTANCE MEA 8 Added Pari Me PRSETER MN DM
10. 16 slots in each slots of the same raster have similar size and coordinate system coordinate systems and pixel sizes of the primary and additional raster are independent and may vary Both pixel coordinate system axis from left to right and from bottom to top and target coordinate system Cartesian system parallel to pixel rows and columns with random dimensionality normally expressed in meters are used in the application From the viewpoint of image visualization view or presentation both raster are equal whereas processing opera tions can be done only using primary raster secondary raster can be used only for mosaicking from several images Kohonen neural networks self organized maps SOM Neural network topology Self organized maps proposed by Kohonen are single layer neural networks with a competitive layer where the in put vector is fed at the same time into all neurons inputs Only one winning neuron matching better with the input vector is identified based on one of the following methods to calculate distance between input vector and each neuron s weights vector On the other hand we interpret neuron s weight vectors as classes centers defined as distance segmentation using the same distance function In this way one to one correspondence between neural network neurons and their classes 15 defined in the initial feature space Neural network training implies updating of winning neuron s weight vector
11. Checkbox Value controls the code saving the code is saved only if this control 15 checked Field Value enables to display vectorized objects code Field Min Area enables to specify minimal area of vectorized objects Field Min Perim enables to specify minimal perimeter of vectorized objects Button MIF Style enables to specify drawing style of the vectorized objects Button Run initiates vectorization Creating image using legend When displaying the image based on the slots values containing classification results using the legend the dialog looks as follows 40 Dasiicaoini C Plane Classiticatoin 1 C Gray Breen Legend Blue Used Entries Only Random Load Save True Color Select Haste f Primary Secondary 256 Values Vectonze Classification to MIF Select File Thematic MIF Style Field Field1 Topology Value Run Area 10000 Perim 20000 Preview Cancel ok Legend List box Plane enables to select slot containing the classification results Table Legend contains the list of all values within the legend indicating class code vectorization feature output color and comments for each value legend class Checkbox Used Entries Only if checked only codes of the classes present in the corresponding slot are put in the legend Button Random enables to select random colors
12. controls are described in details in the corresponding section of this user s manual Editing vector objects drawing mode To change the selected object drawing mode use Change Mode item in Pop Up menu of the map window MIF Display Mode dialog appears on the screen which controls are described in details in the corresponding section of this user s manual Raster data manipulations in NeRIS inner format The use of raster data in NeRIS inner format gives the best advantages in remote sensing data processing Creating primary raster To create a primary raster use Create Primary raster item in File menu of the program window A dialog will ap pear on the screen which controls are described in details in the corresponding section of this user s manual Loading working and secondary raster To load a working and secondary raster use Load Primary raster and Load Secondary Raster items respectively in File menu of the main program window Saving primary raster To save a primary raster use Save Primary raster item in File menu of the main program window Two options to save the primary raster are available using standard file selection dialog e Saving the whole raster Internal Raster Format item in the Save Dialog list of file types e Saving selected subregion Clip Subregion in STI Format in the list of file types of the Save Dialog standard window In the second option SubRegion selection dialog appears on the screen containing
13. dU M UNE 8 DEBE __ __ ME DM Mc 8 HESS D S qoo eee rien 8 MM Ma Me 8 E REEE Oi 8 NNNM NEN ONEEN EAE M 8 O eects nie 9 2 9 2 9 7 eR M M annie id MPH 9 KASTER DATA MANIPULATIONS IN NERIS INNER FORMA REM RU MEG 10 LIE vest eee eee 10 ices 10 MAT USUS AED os MR rm 10 Creating deleting cleaning and dubbing of raster 10 sei RHENUM UR EET dea 10 T ea 10 CHANGING PROCEDURE 13 Len eT Re Te eT een Tene ees eee 13 Principal CRISE chico ieee et eh ees 13 M 13 ie ee eR _ _ 13 CLASSIFICATION USING SOM 14 TTE ________ _ 4 6_6_6_
14. from STI into other formats For data exporting from STI format into one of the other formats GeoTiff ERDAS Imagine 8x ESRI BIL bil PCI Geomatica ENVI BMP PPM ArcInfo ASCII GRID ASCID use Export item in File menu Export from STI format dialog appears on the screen containing the following controls 11 Export from 5TI Format 2 Destination File JES amplene tif Save ESRI File SaveMaplnfao TABE Save WET File Source File JENS 20000906 sti 111021 02120000905 10 1_ 7111021 02120000905 20 chl_warp Remove 2 11 1021 02120000905 B3 chl warp Up 10 Lziiz1021 02120000905 40 chl warp Down 1 71171021 02120000905 50 chl warp 12 172171021_02120000905_ 70_ 1 warp Close Export Field Destination File specifies output file name and format File selection push button enables to choose a directory where the new file is created Checkbox Save ESRI World File If checked World file will be saved which is used for the creation of image spa tial coordinates of the ESRI program family ArcView ArcGIS Checkbox Save MapInfo TAB If checked TAB file will be saved which is used for the creation of spatial coordi nates and image projection in MapInfo Checkbox Save WKT PRJ File If checked text file with WKT projection parameters will be saved World known text Field Source File specifies input file name and forma
15. lower output values thresholds received after linear transformation from the current slot values Gamma correction Ch field located on the right enables to set gamma correction separately for each slot Gamma correction field located on the left enables to set gamma correction for the entire created image gamma correction curve dialog is shown the rectangular in right bottom corner Notably transformations are done in the order specified above for the controls Button Preview enables to preview the image without getting out of the dialog in case of a new image creation rather than modification of the existing this button 1 not available Button Load enables to load image parameters band and gamma correction sliders position from a file Button Save enables to save image parameters band and gamma correction sliders position to a file Button Auto performs image auto contrasting 38 Creating monochrome display Monochrome display dialog is shown below view Parameters HUE 12 172171021 02120000905 0 Fm Unlike the aforementioned dialog this dialog has only one slot therefore there is no control for the current slot selec tion 39 Creating image using neural net classification results When displaying the image based on the slot values containing the classification results the dialog looks as follows Plane Classification 1 Green
16. of built classification quality and compliance of neural network struc ture transformed into Sammon s space with the initial neural network structure Backbone tree 15 built for this purpose using both the entire distance matrix between classes and the distance within the neurons grid which is displayed over the built Sammon s map below is the Sammon s map for trained neural network and the backbone tree image allowing for a better understanding of the classes structure The absence of self intersection of the built tree indicates a successful self organization of the neural network and the possibility to use Sammon s coordinates for further processing 15 Hierarchical thematic classification NeRIS application provides for a wide range of remotely sensed data analysis using the multi layer image structure detected by the neural network First of all these are the unique means of palette creation to visualize the built classification based on the color gra dient display of the image structure To create such structural palette the application is using the color sprout metaphor Color sprouts are the col ors assigned to individual neurons of the neural net by the user To select these neurons the user can apply the calibra tion results select neural net classes matching better with thematic classes distribution of neural net classes throughout the territory visually assessed when building raster images or select
17. of these classes for a de tailed classification and study Classification post processing using local window As individual pixels classification being contextually independent does not allow for detection of a number of tradi tional decoding features e g texture NeRIS application provides for the post processing of the already built classifica tion using classes distribution inside a local window detected by the trained neural net Since such distribution is unidi mensional a standard approach of signature selection is used to operate it typical for the given thematic class of neural net classes distribution The signatures are automatically withdrawn from the trained and thematically calibrated neural network To verify the signatures against the required ones the NeRIS application uses 8 signature matching measures de scribed in details in the corresponding section of this user s manual As a result of such post processing the thematic class which signature 15 the closest to the one calculated within the local window is assigned to the central window pixel Thematic classification of vector objects Besides the aforementioned procedure of signature calculation within the local window the object by object classifi cation procedure becomes very popular lately when for each vector object on a certain digital vector map it 15 known in advance that it belongs to one thematic class In this case it s enough to calculate the cla
18. shifting it in feature space towards in put vector Unlike traditional classification methods when separate classes updating is made independently in SOM network all neurons consequently all classes are put into certain topological grid specifying neural network topology This topology has nothing to do with neuron s weight vectors position in the initial feature space and serves only to con trol neural network training for its self organization Topological grid is used to define the neighborhood of the winning neuron which includes all adjacent network neurons in the grid regardless of their weights distance similarity to the input vector The principal difference of SOM neural network from traditional classification systems 15 in the fact that besides the winning neuron all neighboring in the neural network grid neurons are also trained which leads to neural network self organization and determination of internal structure of multi layer image Determination of this very structure 1s the re sult of neural network training In this software version topological grid 1 in form of regular rectangular net containing not more than 255 neurons The example of a 7 by 7 set of neurons 7 rows with 7 neurons is shown below Neurons numeration in neural network corresponds to their position in gridtop topology bottom left neuron has the position 1 1 next one has the position 1 2 and so on when saving neural network to a file the neurons
19. the right mouse button on the hi erarchy tab field Hierarchy tab Pop Up menu Add new node to the hierarchy at the current level Child Add child node of the hierarchy at the current level Delete 44 uniform color legend for all neurons referred to the hierarchy node Save hierarchy tree Kohonen neural net training amp calibration Teach SOM Neural Net dialog This option 15 opened by selecting Teach SOM item in Classification menu and is designed for neural network train ing control This dialog is a multi tab interface containing all required tools for the creation training and calibration of Kohonen neural nets according to the specified raster slots of the primary raster Planes tab enables to select slots and their pre processing method Teach SOM Neural Net a 2 Planes Teach SOM Thematic Calibration Hierarchy Planes Classity athe 71171021 02120000905 B10 chi 1 00 8 171121021_02120000905 B20 chi 1 00 3 71171021 D2120000805 B30 chi 1 00 10 Lz1171021 o2120000905 840 c 1 00 11 Lz1171021 02120000805 850 cr 1 00 Remove Add Weight 00 1171021 02120000905 chl war This tab contains controls setting the number and order of primary raster slots used as well as the normalization method Table Planes to Classify contains the list of primary raster slots involved in training and having two columns each e Column describing prim
20. them after visual structure analysis presented as Sammon s map for better structure visualization The color for the visualization of the rest of the classes 1s generated automatically by interpolation of color sprouts using Sammon s coordinates As a result the closer are the classes in Sammon s map consequently within the source image space as well the closer 1 their color in palette However even the colors of neighboring classes are somewhat different which enables to visually structure the image and detect its spatial structure Whereas random color selection traditionally used to visualize classification results brings to a fractal picture with a big number of classes which losses spatial objects the approach used in the NeRIS application enables to present the results of classification into a big number of classes so that there are no visual distortions of spatial structure Notably there is a possibility of using different colors and its tones to build hierarchical structure created by the neu ral net classification as well as the possibility to build color legends using the neural net calibration results when neighboring color sprouts are used for the far off in the neural net but thematically close classes The possibility of building hierarchical classification using several neural networks should also be noted The first net is used for rough thematic distribution e g water objects plants etc with further selection
21. using neural net SOM which will be used for pattern SOM neural net processing Field Window Size specifies the size of window ROI tab controls Group Define Region of Interest indicates the requirement to limit the training area by the regions contained in vec tor map file File selection push button enables to select the vector map file containing training area borders 53 Teach tab controls Teach Pattern SOM Neural Net Input Teach Pattern SOM Thematic Calibration earning Parameter Sampling Step il i Run Length 300000 Alpha 0 3 Court of Pixels 2990430 Use Distortion Mode Weights Use weights Sorted Field Sampling Step pixels sampling step during training and calibration value 2 indicates the use of each second pixel in row Field Radius neighborhood radius value used in training Field Run Length total number of used pixels if the region being trained contains fewer pixels the latter are used several times Field Alpha initial value of training rate Checkbox Use Distortion Mode enables to change training mode used if there was no self organization of neural net or classes ordination during regular training Pattern SOM neural net tab controls Input Teach Pattern SUM Thematic Calibration SOM Parameter Horis E H SOM Vert Size 3 Distance Euclidean Neig
22. 22 Cancel Projection 157 box contains the following projection types NonEarth projection plan e Longitude Latitude Longitude latitude Albers Equal Area Conic Albers Equal Area Conic Azimuthal Equidistant Azimuthally equal distant projection Equal Area Cylindrical Equal Area Cylindrical projection 20 Ekert Pseudocylindrical Number IV Ekert Pseudocylindrical Number IV Ekert Pseudocylindrical Number VI Ekert Pseudocylindrical Number e Equidistant Conic Simple Conic Equidistant Conic Simple Conic projection Gall Gall cylindrical projection Rectified Hotine Oblique Mercator Rectified Hotine Oblique Mercator projection e Lambert Azimuthal Equal Area Lambert Azimuthal Equal Area projection e Lambert Conformal Conic 2SP Lambert Conformal Conic 2SP projection e Lambert Conformal Conic 2SP Belgium Lambert Conformal Conic 2SP projection for Belgium Mercator Cylindrical Mercator Cylindrical projection e Miller Cylindrical Miller Cylindrical projection Mollweide Mollweide cylindrical projection e New Zealand National Grid New Zealand cartographic grid e Swiss Oblique Cylindrical Swiss Oblique Cylindrical projection e Robinson Cylindrical Robinson Cylindrical e Sinusoidal Sinusoidal projection e Universal Polar S
23. 6_ r E 14 ieee 14 14 USING SOM NEURAL NETWORKS FOR REMOTELY SENSED DATA CLASSIFICATION AND 15 Br ND DLNLSaussivethodipyR EDDIE MIO Masc HPM RR MMC MNA ices eee 15 LIH Buen ____ 16 Classification post processing using local window ic insincere vinta nano canine 16 16 Classification post processing using Markov s random fields 16 a DE ue eden AMAA DUREE MED 17 INERIS APPLICATIONCONTRODB 18 18 Creating new raster file in NeRIS inner format Create New Raster dialog esses 20 Manipulations with primary and secondary rasters Primary Raster 2 Principle Components calculation Principle Components Using filters for image changes Filter Plane 25 dic ct I ipd serra enter renee Zo LINE 26 features in local window 26 Can baare DAI ERO la arn AR Ra 28 2 Moving data from vector maps into raster slots Rasterize Vector Map ain Creating sampling text file from primary raster Raster 8 dialog PN i Creating recoding table
24. Correction ye Se 8 LAL Ko 15 Load Save 1 Auto v True Color Select Haste f Primary Secondary C 256 Values Color Select Raster group of radio buttons controls the selection of primary Primary or secondary Secondary raster as the source of slots for the view List boxes Red Green and Blue specify slots for the red green and blue planes of the view being created respec tively There are three rectangular in the middle of the dialog to select the current slot as well as to output the data about the selected slots the histogram shown in red maximum and minimum output values shown in blue to the left upper and lower thresholds of the input image shown in grey vertical lines and gamma correction curve dialog also shown in grey Current active slot is reflected in the option group to the right RGB group of radio buttons enables to select the current slot for its display control Horizontal slider bars Two horizontal controls below the histogram provide for setting up upper and lower input values thresholds used during the transformation of current slot values values below the threshold are shown 0 val ues over the threshold are shown as 255 other values have linear transformation in the interval within the specified val ues Vertical slider bars Two vertical controls to the left of the histogram provide for setting up upper and
25. Map file name selection button The file name is used to indicate the file containing borders of the region to be vec torized List box Code Field enables to select the vector map attributive table field containing the value recorded for each re gion into the final slot List box Plane to Fill enables to select the primary raster slot where rasterizing results will be recorded Button Run initiates rasterizing 29 Creating sampling text file from primary raster Raster Sampling dialog To move data from primary raster slots into text files use Raster Sampling item in Utilities menu Controls dialog appears on the screen as shown below Use MIF as Border JE Sample S ample mit m Use Code Field Class 2110 Save En File Planes To Sample Fal 6 Settings Sampling Step 8171171021_0212000090 m Write Digital Numbers Classes as Bit Set Write Coord 7 71171021_0212000090 _ 9 211 1021 021 20000305 Lz 1121021 0212000092 1 Output Forma Comma Delimited Blank Delimited One Column Fuzzy Close Sample Checkbox Use MIF as Border if checked the region outlined by the specified vector map 15 used for sampling if not checked sampling will be done for the entire raster 11 02120600030 Remove Add r 71171021 02120000905 1 Border file selection button defines the vector map name co
26. RIS application enables the user to save the created raster display in the map window together with the coordi nates Save display as item in View menu of the map window is used Display settings can be saved in the following formats e Using one channel BMP file with 256 gradations of grey color BMP file with 256 colors and True Color BMP file depending on the created image coordinates are not saved e Saving coordinates in TAB format of the MapInfo with the option to save the image as a one channel BMP file with 256 grey gradations or as a True Color BMP file e Saving the selected number of images in a standard file with bites interleaving BIL coordinates will be saved in standard files with HDR and BLW extensions e Saving the image and the coordinates in the inner ScanEx format STI with the option to save part of the image Save subregion dialog is used to select the subregion similar to primary raster saving e PPM Raster format images 13 Classification using SOM Basic methods of neural networks classification available in NeRIS application are illustrated below Their controls are described in details in the corresponding sections of this user s manual with the help of Teach SOM Neural Net SOM Classification SOM Classification Postprocessing and Contextual Postprocessing dialogs Classification elements Neural network training and self organization is illustrated in this section as well as the process of its themat
27. aining List box SOM Classes Plane specifies the primary raster plane classified using neural net SOM which will be used which will be used for pattern SOM neural net processing Field Window Size specifies the size of window ROI tab indicates the vector layer limiting the training area if Checkbox Use Vector Layer as Border is checked Pattern SOM tab contains the following controls Classify using Pattern SOM Neural Net xl Input Pattern SOM Load SUM Comments Result Plane 211 1021 02120000905 BTOU ch warp Y 56 Field Load SOM loads the neural net from file Button Comments enables to view parameters of loaded neural net List box Result Plane enables to select the plane for storing classification results Button Classify initiates the image classification Classification results postprocessing using local window SOM Classification Postprocessing dialog Classification results can be subject to postprocessing in local window in order to define thematic classes instead of neuron numbers Select Local Postprocessing item in Classify menu Postprocessing dialog 1s shown below SOM Classification Postprocessing Define Region of Interes Use Vector Layer as Border JE Sample ROI mif Input Parameters SOM ta Use JES ample tst som Plane to Process 12 1171021 02120000905 10 chl warp Postprocessing Made Postprocessing Parameters Measur
28. all types of objects e Adding new vertices and deleting existing ones lines polylines multiple lines regions e Adding and deleting polylines multiple lines e Adding and deleting borders regions Changing coordinates of individual vertices 15 applicable to active points lines and polylines all vertices are outlined with squares when displayed For multiple lines and regions it 15 required first to select an active arc it has big squares around vertices during drawing by clicking with left mouse button on any location of the arc to be made active To change coordinates of the active arc vertex press left mouse button within the outlined vertex square and drag the vertex to a new location To add a new vertex press left mouse button over the line connecting two vertices of the active arc created vertex can be already moved to the required location To delete a vertex on the current arc keep left Ctrl key depressed and click left mouse button on the vertex to be de leted To add a new line to a group of lines or a new border to a region use Add Arc Region item in Pop Up menu that appears if the current object is a group of lines or a region After that the program activates drawing of new line border to be added to object Changing object attributive data To change attributive data of the selected object use Change Attributes item in Pop Up menu of the map window Edit Object Data dialog appears on the screen which
29. are presented in rows starting with the bottom This numeration will be further on referenced as coordinates in neural network grid in particular such coordinates are used when calculating the neighborhood Distance functions Three functions are used to define the winning neuron in the NeRIS application names used in the application are indicated in brackets e Sum of absolute differences between vector components City Block or Manhattan distance e Euclidean distance between vectors Euclidean e Angle between vectors spectral angle calculated as relation of scalar vectors product to the product of their lengths Spectral Angle functions are selected to make neural network training stable up to a certain limit with respect to linear illumina tion changes of different image channels planes Data manipulations Data manipulations include both raster data and vector data manipulation assets Distance measurements on maps Relevant toolbar hot key is used to measure distances To measure a distance press left mouse button over the start point and drag the cursor to the destination point The distance 15 shown the status bar at the bottom of the main widow Vector data manipulations Creating vector objects layer To create a new vector object layer the data on its geographical projection is required therefore it can be created only if a pre loaded raster is available in the memory projection pa
30. ary raster slot e Column with slot weights multiplied by the values prior to vector input in the neural net thus changing the impact of the relevant slot on the training results To control normalization click left mouse button on the corresponding cell to change weights use the relevant field in the tab bottom Button Remove deletes current slot from the list for training Button Add adds a new slot in the list for training Field Weight changes the plane weight List box Plane below enables to change current slot in the list for one of the primary raster slots ROI tab enables to define the region of interest mask which pixels are used for neural net training 45 Teach 50M Neural Net 21 Planes Teach SOM Thematic Calibration Hierarchy Defne Region of Interes lise Vector Layer as Border JESS ample SAUL mit Legend Hierachy Class NONE The application provides for two selection options using vector map in MIF format as mask and using the slot pre viously classified by another neural net by specifying hierarchy nodes containing the required classes in this case mask control is done via the hierarchy tab Checkbox Use Vector Layer as Border indicates the requirement to limit the training area by the regions contained in vector layer file File selection push button enables to select the vector map file containing training area borders Slots names and hierarchy nodes
31. aster item in Utilities menu of the main program window Patch Primary Raster dialog appears of the screen which controls are described in details in the corresponding section of this user s manual Import export of files Raster data importing to STI format To work with raster data in GeoTiff ERDAS Imagine 8x ESRI BIL bil bsq PCI Geomatica ENVI BMP JPG and ArcInfo ASCII GRID ASCII formats use Import item in File menu Import to STI format dialog appears on the screen containing the following controls 10 Import to STI Format 21 Destination File 5 ample snew sti Source File Use Mult channel Source Channel Close Import Field Destination File specifies the output file name and format File selection push button enables to choose a directory where the new file is created Field Source File specifies input file name and format File selection push button enables to select a graphical file to be imported Checkbox Use Multi channel Mode activates the list of existing channels Table Source Channels displays the list of existing channels Button Add adds the selected file to the list of existing channels Button Up Down moves the selected channel in the list one row up or down Button Remove deletes the selected channel from the list of existing channels Button Clear clears the list of existing channels Button Import initiates importing operation Exporting data
32. aving them in new slots added to primary raster As only 8 bits are used to save numbers in slots the calculations are made in two steps At first maximum and minimum values are calculated which the principal component acquires then these values are used for the linear transformation to the 0 255 range of values 0 is the minimum principal component value 255 maximum and saving of received specified values The principal component dispersion data 1 stored in the slot description 24 Using filters for image changes Filter Plane dialog To apply filters to the primary raster slots use Filtration item in Utilities menu Filtration control dialog indicated below appears on the screen List box Plane to Filter selects the primary raster slot which the filter will be applied to List box Result Plane selects slot for the results Filtration tabs a set of fields to select filter type and its parameters Currently the following filter types are avail able statistic Median tab convolution filters Convolution tab calculation of image texture features in local window Texture tab and brightness classification Relief tab Using statistic filters Median tab Filter Plane _ Plane to Filter Result Plane 17 171171021 0212000090 gt Median Convolution Texture Relief Filter Mame Linear Averaging Window Size JKA Close Apply List box W
33. avoid photometric correction during the multi date images analysis Multi layer raster images segmentation using trained neural networks Thematic calibration of trained neural networks in order to correlate the structure and detect its possible thematic interpretation Thematic interpretation of multi layer raster images building thematic legends and automatic selection of the matic objects and saving them as vector maps Use of post processing of classification results both for verification and reliable thematic classification of preset vector objects The use of image structure detected by neural network to create invariant presentation of the image data 1s a unique function of the NeRIS software providing great possibilities for remote sensing data thematic interpretation not avail able in other software applications Data models Vector data visualization NeRIS application uses vector digital non topological maps designed to store geometrical and attributive data about vector map objects Data models visualization details are indicated below Vector cartographic information It is common practice in computer cartography to present vector data in layers The number of co resident layers should not exceed 256 NeRIS application has all traditional layers control tools change output order when drawing select drawing mode both for the whole layer and using thematic legend control view and editing properties as well as obje
34. ce SOM was calibrated using From SOM classes as thematic labels Transter Group Label Source Three options are possible e None deactivates thematic labels e Labels from Vector enables to select the attributive data field for thematic labels use Vector layer speci fied in the second tab of the dialog is used for calibration e Labels from Plane enables to use thematic labels from raster plane raster map or classification results Checkbox SOM Classes Ready enables to set new thematic labels on the previously classified layer the layer is specified in the list on the right 48 Button Set Labels initiates thematic calibration of the previously trained neural net Field Tranfer thematic labels from SOM specifies the neural net where the data 15 taken from and provides for the data transfer from one grid target grid to another source grid loaded in SOM tab Group Transfere mode Two options are possible e Source grid classes centers are used as data and the nearest neuron class is searched for them in the target grid used for calibration e The layer classified by the target grid is used this layer contains the numbers of this grid classes and it was used for the source grid calibration In this way each neuron of the source grid is matched with the target grid neurons distribution whereas each target grid neuron is matched in its turn with the thematic classes distribution Button T
35. cedure with Shift key pressed click left mouse button on the relevant neuron If left mouse button 15 clicked with pressed Alt key the class will be selected in yellow the map should already be visible SOM tab As it was already mentioned the main purpose of this control tab 1 to geometrically map the thematic data con tained in the neural net Field Alpha specifies the parameters value used in Sammon s map creation iteration Field Err specifies the current Sammon s map error Checkbox Span SOM enables to specify the requirement to build and view minimal backbone tree corresponding to the entire classes distance matrix Checkbox Span Tree indicates the requirement to build the minimal backbone tree on the distance matrix in neural net topological grid rather than on the entire classes distance matrix Button Load SOM loads the trained Kohonen neural net from file Button Comments enables to view loaded neural net parameters Button Pack SOM enables to execute one Sammon s map creation iteration Button Reset View enables to use neurons coordinates in topological grid instead of Sammon s map 43 Button Save View enables to save the built neural net image as a Windows meta file Palette tab This control tab is meant for the thematic palette selection using the following controls Checkbox Seeds View specifies the output color in the circles corresponding to neurons if checked color seeds of the future palette a
36. cts Button Legend enables to control thematic vector layers display see corresponding dialog description Button New MIF creates a new vector layer Checkbox View raster enables to deactivate raster drawing in the display window in this case only vector layers are drawn in all windows Vector layers thematic display control Legend dialog To control vector layers thematic display use Legend dialog which can be opened in vector layers control dialog It enables to select one of the three methods of object style display e Using values of one of the fields to specify style thematic legend e Using one style to display all layer objects e Using the style which is saved for each object in the vector layer Objects output style determination method Legend create thematic legend in accordance with the attribute values e Determine display all objects using one style e default use default objects style 32 As the controls depend on the selected display method we will describe thematic legend creation controls first Display according to thematic legend 21 Field Class Use Legend H _ NM Gradient Legend Uniform Style C Default Style Cancel List box Field provides for selection of vector layer attributive table field according to which a legend will be created In this case the simplest version is used when the number of different field value
37. cts selection for a specific layer Each layer 1 loaded from a separate file using MapInfo export format MIF MID This format enables to save both geometric objects description and attributive tables with different data for each of the objects layers are saved to files in the same format NeRIS application uses non topological item by item geo coding for objects geometry mapping The number of geometric primitives is conventional for non topological systems such as MapInfo or ArcView MapInfo terminology is also used in NeRIS application Below is the list of all geometric primitives used relevant MapInfo system terms are given in brackets e Point a unique point vertex e Line a line represented by the start and end points vertices having the start the end and the direction e Polyline one line represented by a sequence of points vertices in the application The line does not have a di rection the start and end points are not selected deliberately For the internal line presentation the word arc 1s used the word vertex is used for the line points e Multiple Line a group of lines regarded as one object having one record in the table of attributes Lines inside the group are not ordered and presented in the application as mentioned above e Region spatial object region The region is presented as a set of borders closed rings The border does have selected points or selected traverse direction F
38. culation List box Grey levels specifies the number of image intensity values intervals used in GLC matrix calculations List box Window Size specifies the local window size used in GLC matrix calculations Checkbox Set Symmetry if unchecked the obtained GLC matrix is symmetrized Checkbox Equalize Levels if unchecked the brightness intervals are built equal in sizes if checked approximately equal in number of pixels inside Group of radio buttons Texturial Features specifies the texture features being calculated Specifying the probability recorded in cell i j of GLC matrix through f i define average in rows average in columns average dispersions in rows average dispersions in columns and the texture features Angular Moment ANGULAR MOMENT gt f i Ti y i j 24 Entropy entropy gt fli j log faa i j Inv Diff Moment Inverse difference moment i Difference moment JY Fl j i j Cluster Shade i j Cluster Prominence Cluster prominence 2 0 4 J f i sJ i j k i j k E Al Difference variance k E y f i S Where 44 sum average Diff Entropy Classification of image brightness relief Relief tab This filter takes brightness values as relief vertices and classifies each point based on local quadratic model built using local window Below are the codes recorded by the pr
39. description can be edited using standard Win dows tools by making this field active using the mouse Selected slot in the list 1s the current one all operations are applied only to the current slot Button Add enables to create a blank slot filled with zero values and to add it to raster Button Remove enables to delete the current slot from the raster Button Duplicate enables to create a copy of the current slot and to add it to raster Button Clear enables to clear the current slot fill it with default value normally zero Button Histogram enables to renumber histogram of the current slot Button Import enables to import raster data into PPM PGM Surfer GRD ArcInfo ASCII BMP formats and to save it as slot slots 22 Creating mosaics from several images Patch Primary Raster dialog To control mosaics creation use Patch Primary Raster item in Utilities menu The dialog with controls indicated below appears on the screen Patch Primary Raster 2 Region to Patch 5 ample s5 ample mit From Slot To Raster Slot 13331005 610 b gt L71171021_021 20000 Regression Use Regression Close Apply Region to Patch push button 1s used to outline the region borders in the target coordinate system of the primary raster which should be replaced by the relevant region from the indicated slot of the secondary raster List box From Raster Slot specifies the secondary rast
40. difies the field name List box Type enables to select one of the available field types Char Integer Short Float Decimal Date Logi cal for MIF MID files Field Field Length specifies the field length where length has no evident result from the field type Buttons Up and Down enables to change field sequence in table by moving the current field up or down Button Add Field adds new fields in table Button Delete Field deletes current field from table 36 Raster display control View parameters dialog One of the basic operations used by NeRIS application is making primary and secondary raster display Both source images and derived images after processing and classification can be used as initial data Another important operation using the controls of this dialog 1s to transform classification results into vectors and to store them in one of the loaded vector layers The application allows for several display methods RGB composite image using three primary or secondary slots e Black and white composite grey scale image using one primary or secondary raster slot e Image generation using the number of colors equal to the number of classes contained in one slot of the primary or secondary raster received after classification Display mode is specified using Select Raster View Type and Type controls of View Parameters dialog form View Type group of radio button specifies the number of slots used to make the display
41. e by moving mouse along the future line with depressed left mouse button and Shift key which enables to outline the required objects on the raster image After releasing the Shift key the program will mode back to separate points drawing It should be noted that in free drawing mode the polyline is simplified 8 down to half of pixel size preventing to draw polygrams inside the pixel however removing minor details to generalize stored data size reduction In any case to complete polyline creation double click with left mouse button or select Complete Editing item in Pop Up menu of the map window Multiple lines are originally created from one line like in the previous item whereas new lines are added in editing mode Region is originally created in form of one simple polygon similar to the polyline creation the border is locked auto matically Deleting vector objects Editing functions of the corresponding layer should be activated to delete vector objects with the active object to be deleted To delete a vector object press Del key or select Delete Object item in Vector menu of the map window The last deleted object can be restored using Undelete item in Vector menu of the map window Editing individual vector objects Edition of geometric properties of an individual vector object includes this type of editing 1s applicable to objects shown in brackets e Changing individual vertices coordinates of object
42. e Simpson Cut 0 30 Histogram Cut 0 01 Window Size 3 Close Apply SOM neural net classes distribution histograms in local window or areal object from the vector layer 15 used as input data vectors for postprocessing comparing with the reference histograms Reference histograms corresponding to the thematic classes are extracted from the neural net calibration For example if 4 thematic classes were used for the calibration of 3x3 neural net containing 9 neurons to each of them corresponds a histogram vector of 9 components representing to the probabilities of 9 SOM classes for pixels from this thematic class Store Measure id Group Define Region of Interest controls area to be postprocessed Checkbox Use Vector Layer as Border indicates the requirement to restrict the image area being classified with the areal objects contained in vector map file File selection push button enables to select vector map file containing region borders Group Postprocessing Mode controls postprocessing mode as shown below Check box Store Measure Postprocessing can be done in two modes depending on the status of the Store Measure check box if it is not checked each pixel value is replaced with the nearest in the sense of histograms similarity the matic class within the local window around this pixel If the control is checked the neural net classes distribution 15 calculated for each areal object in th
43. e vector map file and information on two closest thematic classes 1 stored in the selected attributive table field which should be symbolic List box Field below enables to select the attributive table field of the indicated vector map the field should symbolic with at least 15 symbols where region thematic classification results will be recorded to Group Input Parameters controls the input data Field SOM to use contains the name of the previously trained and thematically calibrated neural net List box Plane to Process specifies the primary raster slot containing classification results and subject for postproc essing Group Postprocessing Parameters controls the process 57 List box Measure enables to select one of the agreement measures to compare SOM classes composition using curent window or object classes histogram and reference histograms Simpson Simpson measure of agreement use classes presence absence only Fuzzy Simpson Simpson measure of agreement use relative proportions Chi Square using p value of Pearson Chi square distributions agreement criteria Dot Product using scalar vector product histograms are normalized in advance to unit length Euclidean len 1 using Euclidean distance histograms are normalized in advance to unit length Euclidean sum 1 using Euclidean distance histograms are normalized in advance to unit sum FuzzyART agreement measure based on the of adaptive r
44. eparated pixels having i and j brightness is recorded into the cell i j Therefore GLC matrix is square with vertical and horizontal dimensions equal to the number of used brightness values intervals 26 In texture classification the matrix derived features are normally used rather than the GLC matrices themselves Filter Plane Result Plane Plane to Filter 2 xX q 41021_027 2000030 gt 1 17111021 02120000905 Median Convolution Texture Relief Texturial Feature Angular Moment Propertie Distance C Entropy Grey Levels C Inv Diff Moment Diff Mament Set symmetry Correlation Cluster Shade Equalize Levels Cluster Prominence C Average C Entropy C Sum Variance C Diff Variance Diff Entropy HxB 5 5 ample mif Close Apply Checkbox Use MIF for Histogram Calc if unchecked the entire image 1s used for histogram calculation if checked only the region specified in the vector file is used for histogram calculation Region border file selection push button 1 used to outline the region in the target coordinate system of the primary raster where the histogram 1s calculated List box Distance specifies the distance used in GLC matrix calculation the bigger is the distance the bigger is the size of the used local window to get the sufficient number of pairs of pixels for statistic cal
45. er slot where the patch is taken from List box To Raster Slot specifies the primary raster slot where the patch is put to Button Apply initiates the operation Button Regression calculates the linear regression to equalize the patch and image brightness histograms Checkbox Use Regression If checked the previously built regression is used to modify the patch brightness Re gression can be built using one file with borders where slots images are partially overlapped whereas the patch itself will have different borders Principle Components calculation Principle Components dialog To calculate principal components of several primary raster slots eigenvectors of the corresponding covariance ma trix or the values of principal components for each raster point use Principle Components item in Utilities menu Dia log with the controls indicated below appears on the screen Principle Components x Use MIF as Border JESS ample 5 ample mit E Add Plane Delete Plane a LAT M1021 02120000905 220 chl warp 71171021 02120000905 B10 chl warp Set Plane 1 1021 0271 2000 Close Apply Checkbox Use MIF as Border enables to specify the requirement to set raster borders where the covariance matrix eigenvectors or principal components are calculated the entire raster 15 calculated by default Region border file selection push button 1s used to outl
46. esonance theory is used Correlation correlation coefficient is used measures have values within the range from 0 complete mismatch to 1 complete match of histograms Field Window size specifies the size of square local window Field Cut specifies the threshold for the composition match measure If measure values are beyond the threshold thematic class 1s regarded as indefinite Field Histogram Cut is used only for Pirson criteria and specifies the minimal probability value when several classes don t have to be combined in one for histogram calculations this 1s done to improve the algorithm stability 58 Classification results postprocessing using Markov random fields Contextual Postprocessing dialog Classification results can be subject to postprocessing in order to detect the thematic classes distribution best match ing with neural net calibration throughout the entire image being classified Select Contextual Postprocessing item in Classification menu Postprocessing dialog is illustrated below Contextual Postprocessing 2 Define Region of Interes Use Vector Layer as Border JE Sample ROI mif Input Parameters SOM to Use 5 ample tst som P Plane Classified using SOM Comments 8 71171021 02120000905 20 chl_warp Result Plane Classification 3 Pastpracessing Parameters Change cut x 10 Contextual weight 1 5 Iterations Equal Prior Probabili
47. ets The principal difference 1 in the fact that the raster slot with regular neural net classes is used as initial data Ly NeRIS application controls NeRIS application main dialog Main ScanEx NeRIS application dialog appears on the screen after loading the program File Utilities Classify Window Help RSs A 335919 81 6201621 69 S23 E As the application interface uses multi document style of the main dialog all created raster and vector data views are auxiliary to the main dialog and appear only inside it Status bar 15 shown at the bottom of the main dialog indicating the current operation mode mouse coordinates and the class number of the corresponding pixel during classification results visualization measured distance in distance measurement mode and the current input scale If at least one window map is created the application menu expands as compared to the one shown above by add ing new controls 5canEx NeRIS ViewA 5 4 2 E Sample 20000905 ml x 4 View Vector Utilities Classify Window Help 8 gin ye Zsa _ nU 1335162 34 6209146 93 E 19 Creating new raster file in the NeRIS inner format Create New Raster dialog The dialog is used to create a blank raster file containing one clean plane at the time of creation with specified di mensions and coordinate system
48. f this hierarchy SOM Classification 21 Planes SOM Hierarchy Plane 02120000905 10 SOM Parameters Hors Size 722 Vert Size 772 Dum Load Hierachy d SUM List box Classified Plane indicates the primary raster slot containing neural net classification results located on hi erarchy node Checkbox Use Curent Node to Define ROI indicates that the classes from current hierarchy node will be used as mask during neural net training Button Load Hierarchy loads hierarchy from disk Button Get Node SOM enables to get the neural net from the hierarchy node and make it current available for train ing and calibration the same way as if it were loaded from disk or created in the SOM tab Notably the current neural net can be saved into a separate file 52 Contextual Kohonen neural network training and calibration Teach Pattern SOM Neural Net dialog This dialog can be opened by selecting Teach Pattern SOM item in Classification menu and is designed to control contextual neural net training process Teach Pattern 50 Neural Net x Input Teach 50 Thematic Calibration Input Parameter SOM used to create classes JESS ample som A Use Classes Calibration Contrast Palette Sammons Distance SOM Classes Plane Classification Window Size 7 Inp
49. for classes display Button Load enables to load previously saved legend from file Button Save enables to save prepared legend into file Field Legend enables to add comments to the corresponding legend class Vectorize Classification to MIF group enables to transform classes selected for vectorization into vector objects in the same way as during neural net classes vectorization in the relevant dialog described above 4 Creating maps based on neural net classification results Create SOM Palette dialog Create SOM Palette aM ajx S ample tst som SOM Er 0 000 Alpha fn 25 Span SOM Span Tree 110 10 0 46 49 ChE Che 43 Che Lhx 83 Ch Ch Chis Them Prapartian Them Class 1 100 0 v ide Classification results display control The central part of the dialog is taken by a square displaying the loaded trained Kohonen neural net Zoom in and zoom out buttons in the top right corner enable to make the view larger or smaller The area of tabs controls is located to the right of the square containing data to select the required classes and palette for their display according to the thematic task to be resolved SUM Palette WEN Hierarchy 0 000 Tree Distance Scale Meurons Alpha 0 25 ian 35 85 H Span SOM HH Gen Palette EL Ferne quan Iw Span Tree Clear Seeds __ Rete
50. hbour Bubble m Alpha Made Linear Mew Load Save Comments 54 This tab contains neural net creation parameters Field SOM Horis Size specifies neural net grid horizontal size Field SOM Vert Size specifies neural net grid vertical size List box Distance specifies distance calculation method e CityBlock L1 metrics e Euclidean Euclidean distance e Spectral Angle spectral angle calculated as relation of scalar vectors product to the product of their lengths List box Neighbour specifies the way to define neuron s neighbors in the grid e Gaussian Gaussian neighborhood e Bubble Bubble neighborhood List box Alpha Mode specifies the way to change the training rate e Invers Time in inverse proportion to the iteration number e Linear linear Thematic Calibration tab controls Teach Pattern 50M Neural Net 1 Input ROI Teach Pattern SOM Thematic Calibration Use Thematic Labels fram vector Code Field Set Labels Checkbox Use Thematic Labels from Vector enables to select the vector map attributive data field from Code Field ist box to use thematic labels Vector layer specified in the second tab of the dialog is used for calibration Group of training and calibration control buttons This group contains the following control buttons Button Create creates a new neural net with specified parameters Button Load loads neural net from file
51. hts contains the list of all attributive table fields of the used vector map and enables to select the one which contains weights the field must be in integer or floating point format with positive values Checkbox Sorted indicates the requirement to sort out vector map objects based on weights values used for training thematic attitude SOM tab provides for the neural net parameters setting Teach SOM Neural Net 2 Planes Teach SOM Thematic Calibration Hierarchy SOM Parameters SOM Haris Size 3 E SOMVertSize Distance E uclidean Neighbour Bubble m Alpha Made Linear Legend Hierachy Class NONE Mew Load Save Comments Field SOM Horis Size specifies neural net grid horizontal size Field SOM Vert Size specifies neural net grid vertical size List box Distance specifies distance calculation method e CityBlock L1 metrics e Euclidean Euclidean distance e Spectral Angle spectral angle calculated as relation of scalar vectors product to the product of their lengths List box Neighbour specifies the way to define neuron s neighbors in the grid e Gaussian Gaussian neighborhood e Bubble Bubble neighborhood List box Alpha Mode specifies the way to change the training rate e Invers Time in inverse proportion to the iteration number e Linear linear Button New creates a new neural net with the specified parameters Button Load l
52. ic cali bration Neural network training The purpose of neural network training 1s to detect the structure of the existing set of raster layers using three spaces e p dimensional space of the source image created by the vectors of values taken from p of equally located pixels of one raster which we call source image space e 2 dimensional neural network coordinates space or the coordinates within a neural network grid acquired in Sammon mapping described below which we call Sammon s space e Thematic space assigning pixels to thematic classes having the dimension equal to the number of thematic classes The detection of the source image structure will be considered here as a replacement of each source image point by a certain class represented by the vector in the same space and the description of these classes configuration since a cer tain distance function is used in training therefore only the classes neighborhood data can be regarded as the training result As we are trying to get not just a number of classes but their structure during the training self organization process is used which locates the neighboring classes in the source image space within the adjacent cells of the neural network topological grids and vice versa To achieve self organization during the training the neighborhood is used within the neural network grid after in putting the next vector of values to the neural net and defining the win
53. in the neural net file List box Plane Classified using SOM enables to select the raster slot in the field containing neural net classification for which the selected neural net was used this slot contains codes of neural net classes List box Result Plane enables to select the raster slot in the field where the classification results have to be stored codes of thematic classes Field Change Cut specifies the threshold value percent in classification changes between iterations which enables to assess the classification as stable and to suspend iterations Field Max Iterations specifies the maximum number of iterations Field Contextual weight specifies the degree of the given pixel thematic class compliance with the thematic classes of its neighbors if the weight is 0 the neighbors have no impact on the final classification of this pixel if the weight 15 big only the initial distribution and neighbors can have an effect on the final classification 59 Checkbox Equal Prior Probability enables to control the selection of initial values of classes based on their prob ability distribution If checked relative probability rather than the absolute one is used during the thematic classes ini tialization relation of this thematic class probability for this neuron to the prior probability of thematic class Group of radio buttons MRF Order specifies the neighborhood order for the Markov random field 1f the order 1s second the pr
54. indow Size specifies the size of local window used during filtration List box Filter Name specifies the type of the used filter Linear Averaging Use average value in window Median Filter Use value median in window Majority Filter Use the value most encountered in window 25 Using convolution filters Convolution tab Plane to Filter 1 7 Lz1171021 0212000080 gt Median Convolution Texture Relief Filter Diagonal Gradient Close Apply List box Filter Name specifies the type of the used filter Diagonal Gradient 3x3 filter to separate diagonal gradients angle 45 Back Diagonal Gradient 3x3 filter to separate diagonal gradients 135 Laplacian 3x3 Laplacian transformation filter Circular Marr 11 11 Marr filter Gaussian and Laplacian transformation superposition Laplacian of Gaussian Sharpening 3x3 filter to separate details Laplacian and original image sum Laplacian original Vertical edge Sobel 3x3 Sobel filter to separate vertical lines Horizontal edge Sobel 3x3 Sobel filter to separate horizontal lines Calculating texture features in local window Texture tab Texture features are calculated based on the analysis of the image brightness level co occurrence in the local rectan gular window recorded as Grey Level Co occurrence Matrix GLCM first suggested in the Haralik Haralik et al 1973 study Co occurrence 15 defined based o
55. indows dialog for color selection the filling lining color Checkbox Background provides for background fill control 1f unchecked background 1s not filled Button Background enables to select lining background using standard MS Windows dialog for color selection List box Border enables to select of the six standard MS Windows line styles for drawing borders Button Color under enables to select border color using standard MS Windows dialog for color selection Field Width specifies border drawing line thickness Rectangular display preview using the selected style 1s available in the corresponding field for control Group Line Style of linear objects display style control List box Line style enables to select one of the six standard MS Windows style for line drawing Button Color enables to select the line color using standard MS Windows dialog for color selection Field Width specifies line thickness when drawing Line display preview using the selected style is available in the corresponding field for control Group Symbol Style of point objects display style control Button Font controls font selection for point objects drawing Button Color enables to select the symbol color using standard MS Windows dialog for color selection Field Symbol specifies the displayed symbol number Font name and selected symbol are displayed in the corresponding field for control Vector object data modification Edit Object Data dialog 34
56. ine the region in the target coordinate system where the ei genvectors and principal components are calculated Table of raster planes contains two columns for each primary raster slot used in calculation e Left ID column of the table displays the internal slot identification and after colons the number of views where this slot was used the use of slot during the creation of views inhibits all its updating activities e Right Raster Planes column of the table contains this slot description 23 List box Set Plane changes the current slot to the specified primary raster slot Calculating covariance matrix and eigenvectors Button Add Plane adds one slot to the list for covariance matrix and its eigenvectors calculation Button Delete Plane deletes the current slot from the list for covariance matrix and its eigenvectors calculation Field PCI contains the number of eigenvectors being calculated corresponding to the specified number of maxi mum eigenvalues of covariance matrix Button Calc initiates the calculation of covariance matrix and its eigenvectors Button Save enables to save the calculated covariance matrix and its eigenvectors to a file on the disk Button Load enables to load previously calculated covariance matrix and its eigenvectors from a file on the disk Principal components calculation and saving Button Apply initiates the calculation of the specified number of raster planes with principal components and s
57. ing to correlate the detected structure with the pixels relation to certain thematic classes Thematic calibration of the trained neural network is used for such correlation by inputting vectors corresponding to the pixels with the known thematic interpretation to the neural network Neural net classifies each of such vectors and remembers which thematic classes and in which proportion are represented in each neural network class Basically if the calibration selection is enough complete these proportions are very close to the possibility of finding this thematic object among the pixels related to this neural network class consequently the vector of such probabilities for each neural net work class can be used for further creation of the classification procedures both based on the probability theory and on the fuzzy logics 14 Using SOM neural networks for remotely sensed data classification and interpretation This section illustrates the principals of using trained and thematically calibrated networks for remote sensing data classification and thematic interpretation Individual pixels classification using SOM Classification using neural networks can be done as follows e To find the corresponding neural network class for each pixel using the same standard algorithm of distance functions as during neural network training and to use the number of the corresponding class as the classifica tion result e Touse the calibration results
58. jects the top one 1 selected 1n the order of drawing In case it 1s shading the required object the drawing order in the layer can be changed Select Move down item in the Pop Up menu the selected object will now be the first in the drawing order list and will not be shading the previously drawn objects Only one object can be selected at a time Hereinafter it will be called an active object Drawing mode used for the active object 1s specified in item Set selection mode in Vector menu of the map win dow Creating new vector objects Editing functions should be activated in the corresponding vector layer to create a new vector object To create a new vector object use Create Object item in Vector menu or Create Object item in PopUp menu of the map window After selecting the type of the new geometric object point line polyline multiple line region the application activates the object creation mode current input mode used during object creation 15 specified in Set draw ing mode item in Vector menu of the map window Point object is created by pressing left mouse button on the required map window location Straight line is defined by one at a time clicking with left mouse button on the start and end points of the line being cre ated Polyline can be entered in two ways e Separate points mode by clicking left mouse button on the map window locations where polygonal line vertices should be placed e Free drawing mod
59. lect one of the supported spheroids Field a specifies semi major axis of the spheroid in meters Field f specifies flattening factor Created raster can be saved to a file according to standard procedure using File Save Primary Raster menu or us ing relevant hot key button 21 Manipulations with primary and secondary raster Primary Raster dialog After selecting raster controls from Utilities menu Primary Raster Secondary Raster items the controls appear on the screen as shown below File 5 0000905 sti Size Pixels 2007 1490 Size World 325500 0 27 56920 0 385680 0 6211590 0 Pinel Size 30 0 30 0 Projection Universal Transverse Mercator UTM 29 ID Haster Flane Comments 1717021 02120000905 10 chil warp i E m DAS TEE pet m 9171021 02120000905 830 chl warp Fiemove 10 L 11 1021 02120000905 ch warp Duplicate 1 021 200003905 850 ch warp Clear 721 1021 02120000505 BU chl warp Histogram Import Information fields displaying sizes and raster projection are in the top of the dialog Primary Raster dialog contains two columns for each raster slot used e Column ID specifies the internal slot identification and after colons the number of views where this slot was used the use of slot during the creation of views inhibits all its updating activities e Column Raster Plane Comments contains this slot description
60. ll thematic classes used during the trained neural net calibration and enables to select the one which the thematic legend is being built for Selection of the current thematic class defines the image view built in the rectangular It displays the total number of pixels blue curve allocated to the classes during calibration and thematic proportion red histogram for all neural net classes arranged according to the share of the thematic class selected List box Them Proportion located on the left contains the proportion of all thematic classes within the class corre sponding to the current neuron current neuron can be selected using left mouse button Slider bar in the bottom of the dialog enables to select the threshold class all classes matching with neurons and located to the right of the threshold are considered as belonging to the selected thematic class whereas the ones located to the left as not belonging Color rectangular to the right of it enables to create the color of neurons referred to the selected thematic classes other neurons receive colors corresponding to the RGB composition of the first three vector elements of relevant classes Thematic palette composition 1s done by clicking left mouse button on the color rectangular neurons referred to the selected thematic class become outlined and can be used for vectorization as a result Neuron selection mark can be changes using standard MS Windows pro
61. matic information and its vectoriza tion for further use in other geo information systems NeRIS allows for joint processing of vector and raster information rasterization and vice versa vectorization for joint processing use of vector maps to control thematic interpretation processes NeRIS has a big toolkit for spatial data thematic interpretation and for the creation and updating of vector digital the matic maps Multi window interface for output raster layers visualization and for the creation of views black white and color images composition based on the data in the NeRIS internal format Saving created views together with geo references in formats compatible with other geo information systems Combination of raster images with vector maps including vector maps output in accordance with thematic leg end Interactive vector maps editing creation of new digital layers creation and deletion of objects objects geometric and attributive properties editing Interactive mosaicking and multi date raster creation with visual control of the results Traditional images processing use of different filters principal components calculation Training of neural networks to detect spatial structure of multi layer raster images visualizing the structure suit able for analysis Getting new images based on the image structure invariable with regard to a whole class of brightness transfor mations which sometimes allows the user to
62. n a pair of pixels separated a given direction and distance Thus for example if the distance equals one pixel and direction is horizontal all pairs of pixels in the window located next to each other horizontally are taken for co occurrence analysis As almost all the cells of the built matrix will be blank if the full range of possible brightness levels 256 1s used and the window is small the entire brightness values range of the selected image is divided into intervals inside which all brightness values are considered similar to build a GLC matrix Intervals can have even brightness and number of pix els found in each interval The first method can be used if the image brightness histogram does not have evident peaks and covers the whole range of possible values If the brightness histogram has for example only one narrow peak this method will bring to the creation of a big number of blank matrix cells The second method is used in this case based on the calculation of such a set of intervals using the image brightness histogram that the histogram area over each of them is approximately the same this will result in a better GLC matrix filling rate Moreover using the brightness histogram of only a required part of the image it 1s possible to improve the GL C matrix filling and consequently the quality of the image texture features calculation Own GLC matrix is created for each direction and distance where the probability of the s
63. ned neural net and primary raster slots This dialog has several tabs and contains all the required tools for the classification of the specified set of primary raster slots Planes tab enables to select slots and their pre processing method SOM Classification i 2 Planes SOM Hierarchy Planes To Classify Ww Result Plane L711 71021_ 02120000905 610 chi 1 00 FLAT 1021 02120000905 chl Comments This dialog contains controls specifying raster slots used in classification Table Planes to Classify contains the list of primary raster slots used in training Each of them has two columns e Column describing primary raster slot e Column with slot weights multiplied by the values prior to vector input in the neural net thus changing the impact of the relevant slot on the training results List box Plane below enables to change the current slot in the list for one of the primary raster slots Number of slots normalization and weight parameters are taken from the comments file of the trained neural net List box Result Plane enables to select the slot for the classification results input ROI ab enables to select the part of the area mask which pixels are used for neural net training 50 SUM Classification RES Planes 50 Hierarchy Defne Region of Interes Use Vector Layer as Border JENS ample SAUL rit Legend Hierachy Class
64. ning neuron which 1 the closest to the input vec tor it will be trained together with all the neighboring neurons NeRIS application is using two strategic options to choose the neighborhood terms used in the application as indi cated in brackets e All neurons located not further than at the specified distance from the winning neuron fall within the neighbor hood in neural network grid coordinates The training rate is similar for all neurons Bubble e All neurons of the neural network are in the neighborhood however their training rate drops as distance from the winning neuron increases Gaussian To bring the self organization process to the required results the radius of the neighborhood decreases throughout the training for the Bubble mode or the training rate drop expedites as the distance from the winning neuron in the grid increases for the Gaussian mode Finally for the same purpose the training rate 1s reduced down to zero throughout the training with the initial rate settings specified by the user NeRIS application uses two strategies of training rate reduction terms used in the applica tion as indicated in brackets e In inverse proportion to iteration number Inverse Time e Linear Linear Neural network calibration After self organization 15 complete the trained neural network detects the source image structure which can be used for multi layer images segmentation However it is more interest
65. ntaining one or several regions outlining the region used for sampling Checkbox Use Code Field if checked data from the specified vector map attributive table field 1s moved into the file in form of a separate column for sampling List box Use Code Field specifies the vector map attributive table field where the value for sampling is taken from File selection button Save to File specifies the text file name where the sampling will be recorded to Table Planes to Sample contains the list of primary raster slots involved in sampling There are two columns for each of them e Primary raster slot description column e Column indicating the requirement to normalize created file rows vector length made of the pixel values in the slots marked with will equal to one Button Remove removes the current slot from the list for sampling Button Add adds a new slot into the list for sampling List box Planes below changes the current slot in the sampling list for one of the primary raster slots Field Sampling Step specifies the step for raster pixel selection for sampling if 1 all pixels are selected if 2 every second in the row and so on Checkbox Write Digital Numbers 0 255 if checked the output values are scaled within 0 255 range otherwise values from 0 to 1 with the step of 1 255 are displayed Checkbox Classes as Bit Set if checked for each recorded sampling row the region code from the vector map at trib
66. oads neural net from file Button Save saves neural net to file 47 Button Comments enables to create and edit comments stored for the neural net In particular parameters used dur ing the neural net training and calibration are stored here including the slots composition and normalization method Standard multi line edit window appears on the screen after pressing this button as shown below The user can add his own comments which are stored into file when neural network is recorded 5 Comments E x AunLength 300000 Step 3 Hadius 5 000000 Alpha 0 300000 NeighbMode 1 1 Sum 519 Slot Wl bs SlotDiff 0 1 1 00 8 111021 02120000905 20 chl warp 2 1 00 9 7171021 02120000905 0 chl warp 3 1 00 10 71171021 02120000905 B4 chl warp Information about the reference of the current neural net to a certain hierarchy node 15 available in the bottom of the tab neural net can be taken from the hierarchy node for additional training or saved in the hierarchy node Thematic Calibration tab provides for the neural net thematic calibration Planes ROI Teach SOM Thematic Calibration Hierarchy Label Source None c Tabelsiiom Vector ic Labels 171171021 02120000 SOM Classes Ready ele _ 211 1021_02 Set Labels Tranter thematic labels From SUM m Transfere made Direct map of the SOM classes using distan
67. ode and make it current available for train ing and calibration the same way as if it were loaded from disk or created in the SOM tab Notably the current neural net can be saved into a separate file Button Set Node SOM enables to save current neural net created or loaded from a separate file using SOM tab con trols into the relevant hierarchy node Hierarchy node corresponding to the current neural net where the neural net was taken from or saved to 15 highlighted with the proper sign Below are the possible painting options for hierarchy nodes SOM 1 stored in hierarchy node which 1 not currently used No SOM in hierarchy node however classes assigned to the node are used for masking SOM 1 saved in hierarchy node which 1 not currently used however classes assigned to the node are used for masking SOM 1 saved in hierarchy node which was either made current taken from the hierarchy node or saved in hierarchy node SOM 1 saved in hierarchy node which was either made current taken from the hierarchy node or saved in hierarchy node however classes assigned to the node are used for masking No neural net in hierarchy node and classes in the node are not used for masking Classification using Kohonen neural networks SOM Classification dialog This dialog can be opened by selecting SOM Classification item in Classification menu and is designed for the clas sification control using previously trai
68. ogram as a result processing results should be displayed using legend the biggest part of the results 15 usually taken by codes 1 and 2 Stream flow river bed Flat place plain Peaks vertices Col selliform Depression water basin bed It is possible to control the scale of classified morphological objects by changing the window size 28 Plane to Filter Result Plane q L71171021_021 2000090 gt 1 17111021 02120000905 Median Convolution Texture Relief 0 001 Convex 0 001 Determinant 0 001 Window Size x7 Mi Close Apply List box Window Size specifies the window size to build the model Field Slope specifies slope threshold value in radians below which the parameter 15 considered as zero Field Convex specifies surface curvature threshold value in meters below which the parameter is considered as Zero Field Determinant specifies internal thresholds to consider slope in radians surface curvature in meters and model determinant quadratic form values as zero Moving data from vector maps into raster slots Rasterize Vector Map dialog To move data from vector maps into the primary raster slots use Rasterize Vector Map item in Utilities menu Rasterizing control dialog appear on the screen Rasterize ector x File Mame JESS ample 5 ample mif Code Field Class Flane to Fill Classification 3 Hun Close
69. ogram runs a bit longer giving more spatially concurrent results of thematic classification Field Pow contains probability contrasting index during the initialization if its value is 0 the initial classes distri bution 15 independent of the probability distribution in the relevant net neuron If the value is big around 100 the the matic class which probability 1 the maximum for this neuron is used as initial value Value 1 corresponds to the use of the existing probability distribution for the neuron initial or relative without modification Checkbox Relative enables to mandatory consider all prior thematic classes probability as equal regardless of cali bration data in the neural net this may be required if thematic classes distribution used in calibration does not corre spond to their actual distribution for this area 60 NeRIS application menu description Create primary raster Create a new primary raster Load primary raster Load primary raster from file Open secondary raster Load secondary raster from file Save primary raster to file Import raster files to NeRIS Exit NeRIS Display parameters Change display parameters Save image as Save primary raster Import Export Exit View Image scale Set scale gp Fits all the image into the program window A 00 Vector layers control Undo remove _
70. or the internal border presentation the word closed arc is used the word vertex is used for the border points Region can be in form of a random multilinked complex poly gon the number of connected components and islands in each of them is not restricted Haster data visualization NeRIS application allows the user to work with many types of raster images e Raster images in MapInfo GeoTiff ERDAS Imagine 8x ESRI BIL bil bsq PCI Geomatica ENVI BMP JPG formats as well as ArcInfo ASCII GRID e Raster images in the internal STI sti format of NeRIS created inside NeRIS application after raster data proc essing and importing External raster data NeRIS application enables to simultaneously open a random number of images having their own windows for inde pendent location and scale control but with a vector data setting and output method general for all windows Such raster images preview options and their use to resolve different tasks are described in details in appropriate sec tions of the manual Raster data in NeRIS internal format Slots are used to store raster data images inside the application Each of the slots contains a single image plane with 8 bit value of each pixel Therefore three slots are required to store one input image displayed in True Color format according to the number of planes in the source file Slots are combined in two groups called primary and secondary raster currently with
71. rameters are taken from raster for the layer being created Primary raster control item or Secondary raster control item are used to create new layers in Utilities menu of the NeRIS master form After selecting this menu item a STI form appears described in the corresponding section of this user s manual Loading vector objects layer To load a previously created vector object layer use Layers control item in Vector menu of the map window or a relevant toolbar hot key After selecting this menu item Layers control form appears described in corresponding sec tion of this user s manual Layers control Layer control includes layers loading deletion changing layers output order and mode and layer properties visuali zation editing factors and objects selection possibilities and saving layers to files Layers control form is used for all these actions as described in the corresponding section of this user s manual Selecting individual objects for updating To select individual objects for updating use Select Object item in Vector menu of the map window or a relevant toolbar hot key Objects selection mode activation 1 indicated by changes in cursor shape and button appearance Click left mouse button on the object to select it objects selection properties of the relevant layer should be acti vated If there are several objects within the mouse capture area the top layer object will be selected If there are several top layer ob
72. ransfer initiates thematic labels transfer Hierarchy tab provides for the creation and operation with neural net hierarchy corresponding to the previously cre ated hierarchy of classes Neural net 1s on the node of this hierarchy Subsets of these neural net classes are below with all classes of subsidiary hierarchy nodes always on the nodes of the higher level Neural net can be saved for each hier archy node Such neural net is stored in the hierarchy file and can be used for further classification Therefore hierarchy is a convenience means of storing data about the hierarchy legend of classification and corresponding neural networks Teach SOM Neural Net E 2 Planes Teach SOM Thematic Calibration Hierarchy Classified Plane _ 21121021_02120000905_ 10 SOM Parameters Load Hierachy Horis Size 722 Vert Size 777 Save Hierachy 777 Made 5 Hode SIM List box Classified Plane indicates the primary raster slot containing neural net classification results located on hi erarchy node Checkbox Use Curent Node to Define ROI indicates that the classes from current hierarchy node will be used as mask during neural net training only pixels related to the mask are used in neural net training Button Load Hierarchy loads hierarchy from disk Button Save Hierarchy saves hierarchy to file Button Get Node SOM enables to get the neural net from the hierarchy n
73. raster plane which needs to be recoded Button Run initiates recoding Layers controls Layers control dialog To control vector layers use Layers Control item in Vector menu of the main display window menu Vector layers control dialog appears on the screen 3l Remove Clear U own Legend Mew _ awe Ws P Dea ELLEN Nn ok View raster Table Vector Layers Control contains the list of loaded vector layers with information and controls for each There are five columns available for each layer Column with the loaded vector layer type MIF in this version MapInfo program exchange format e Column with the layer name file name without extension or path where this layer 15 loaded from e Column with the layer editing feature e Column with the layer visibility feature Bui e Column with the layer objects selection possibility feature x Note All layer features are changed by clicking left mouse button on the corresponding cell Button Load loads vector layers from file After loading the layer will be stored in RAM so one layer can be loaded several times and changed differently then saved into different files Button Save saves current vector layer to file Button Remove removes current vector layer from RAM Button Clear clears all objects of current layer Buttons Up Down enable to change vector layers drawing sequence and their preview when searching for obje
74. re displayed unchecked the palette itself Button Gen Palette enables to build the palette using color seeds by interpolation according to the neurons coordi nates within the SOM area Button Clear Seeds deletes all color seeds Button Set Color changes the color of all currently selected neurons Button Preview enables to preview the palette results in the map window Button Save Palette saves the built palette including the neurons position within the SOM window to a disk file Button Load Palette loads previously built palette including neurons position within the SOM window from a file Button Grad enables to create the gradient palette using Color controls located in the SOM window corners as color seeds Checkbox Tree Distance if checked colors are interpolated using bilinear and linear methods depending on the seeds number according to the neurons position in the SOM window Button Rand enables to create random palette Button Thematic Palette enables to create the palette where each thematic class has its color thematic classes are numbered starting 1 in the same order as they are presented in the thematic classes list located to the right of the slider bar in the bottom of the dialog This palette type 1 meant for remote sensing data thematic interpretation visualization View tab Checkbox Scale Neurons enables to change the neurons display scale depending on the number of pixels related
75. s Recode Thematic Plane Midi di patie eit iene Vector layers thematic display control Legend dins oe k Von F dialog g 7 N New MIF 29 209 A MEN 36 um View 1 s dialog MITTUNT hale maps based o on neural net nds enis SOM Pali dines n PopUp mE AMET MM sex 5 Clasificar n dialog OM Neural Net dialog 08 53 OM E n Net dialog a seria DO indow SOM cation E ist Yd ASSL cation ui ae MESE AR sing randomfields Contextual Postprocessing dialog eee 59 NERIS APPLICATION MENU DESCRIPTION 61 NeRIS software key features NeRIS software application neural raster interpretation system is designed for remote sensing data thematic processing and interpretation NeRIS enables to use both conventional visualization filtration calculation to get new images and modern meth ods using adaptive algorithms based on the artificial neural networks for remote sensing input data ordination and inter pretation The latter are used for more comprehensive spatial data classification and segmentation The main advantage of the software 1s the use of artificial neural networks during the selection of the
76. s should not exceed 256 Table Field Value Legend contains the list of the selected field values one per row and for each of them display style for point linear and areal objects To change the style click left mouse button on the style cell to open vector ob Jects style change dialog Button Gradient Legend enables to obtain gradient color in objects display styles initial and final colors are se lected by clicking left mouse button on the corresponding rectangular Button Random Legend enables generate display style at random Display of all objects with one style Field C Use Legend Button Uniform Style opens the vector objects display style selection dialog which enables to select the uniform style of point linear and areal objects display 33 Vector objects display mode control MIF Display Mode dialog This dialog enables to control vector objects display mode MIF Display Mode Region Styl Color width H Forit Aria Color Symbol 32 E Cancel Width m This dialog contains three groups of controls for the point linear and areal objects respectively Group Region Style of the areal objects display style control List box Fill enables to select one of the eight standard MS Windows filling and lining styles Button Color enables to select using standard MS W
77. sh HH Set Color gt Preview Load Palette Save Palette Thematic Palette Transfer Palette 42 e SOM tab loads the trained Kohonen neural net builds two options of minimal backbone tree as well as displays Sammon s classes represented by the net neurons their visualization and export in form of graphic files e Palette tab creates thematic palette according to the task to be resolved e View tab saves the data contained in the Kohonen neural net compressed in two primary raster slots e Hierarchy tab creates classes hierarchy in neural net Buttons with lt and gt symbols enable to move from one tab to other Color gradient palette creation controls are placed in all four corners of the rectangular Each control has a color cir cle Standard MS Windows color selection dialog can be opened by clicking mouse button on each of them as well as the checkbox indicating the requirement to use this color for the palette creation On the right under the Tabs controls area there 1s a data field displaying the current net neuron position in its topo logical grid and the values of the first eight elements of the class center vector corresponding to this neuron Building thematic palette and vectorization preparation Controls specifying thematic data contained in the calibrated Kohonen neural net are located in the bottom of the dialog List box Them Class located on the right contains the list of a
78. sses signature for classification within the limits of the object and to find the closest thematic class in the calibrated neural network This very approach is used in the NeRIS application Classification post processing using Markov s random fields To get the thematic classification consistent with the entire image the application uses the Markov s random field tool enabling to receive generalized thematic classification with the preset extent of details based on the thematic cali bration of neural network The received classification will be coordinated not only with thematic distribution of probabilities kept within the neural net but also with thematic classes of this pixel neighbors using Gibbs energy minimization when applying Markov s random field of classes Consistence in this particular case indicates that the pixel thematic class is inclined to the neighboring thematic classes as a result of local changes in probability distribution of thematic classes for the given neural net class with the change rate controlled by the user 16 Contextual neural networks Besides the aforementioned post processing method using local window the application also uses contextual neural nets which apply classes distribution within the local window as initial data for neural network training Further use of such neural networks calibration classification and thematic classification 1 the same as for regular neural n
79. t File selection push button enables to select a graphical file to be imported Table STI channels list located in the center of the dialog displays the list of existing channels Button Up Down moves the selected channel in list one row up or down Button Remove deletes the selected channel from the list of existing channels Button Export initiates exporting operation 12 Raster data changing procedures Filtration Filtration item in Utilities menu of the main program window is used to apply different filters to raster layers Filter Plane dialog appears on the screen which controls are described in details in the corresponding section of this user s manual Principal components calculation To calculate principal components of the given raster planes use Principle Components item in Utilities menu of the main program window Principle Components dialog appears on the screen which controls are described in details in the corresponding section of this user s manual Raster data view merge To create a new map window use Create Working Window item in Window menu of the main program window or the relevant hot key button of the menu To merge raster data display in the map window use View Parameters item in View menu of the map or a relevant hot key View Parameters dialog appears on the screen which controls are described in details in the corresponding section of this user s manual Saving and exporting raster data Ne
80. tereographic Universal Polar Stereographic projection e Transverse Mercator Transverse Mercator projection e Gauss Kruger Gauss Kruger projection Pulkovo 1942 e Universal Transverse Mercator Universal Transverse Mercator projection UTM e Cassini Cassini cylindrical projection Polyconic Polyconic projection List box Zone enables to select one of the zones used for Gauss Kruger and UTM projections Fields Standard parallels North BN South BS and Scale indicate corresponding projection parameters main standard parallels and scale Field Azimuth enables to set oblique azimuth used in Rectified hotine oblique Mercator projection set in de grees Group Origin includes Longitude Latitude False Easting and False Northing fields and sets the starting point of the target coordinate system List box Datum enables to select parameters for one of the supported reference ellipsoids Shift DX specifies the shift in X direction Shift DY specifies the shift in Y direction Shift DZ specifies the shift in Z direction Rotation EX specifies rotation around X axis in seconds Rotation EY specifies rotation around Y axis in seconds Rotation EZ specifies rotation around Y axis in seconds Scale factor e specifies scale correction coefficient in parts per million Prime Meridian specifies prime meridian longitude in degrees east of Greenwich List box Spheroid enables to se
81. the following controls Select Subregion for Save n x E SScanexsTosha dok nenssz 00009051 sti Cale Subregion fram JESS ample 5 ample mit SubHegian AI Ae 2006 TT Hs 1489 Checkbox Cale Subregion from MIF If this control is active the region to be saved is specified using vector map as indicated below 5 us Pixels C world Saved region borders file selection push button enables to select the vector map file for which the smallest rectangu lar containing all objects 1 calculated This rectangular defines the raster region to be saved CoordSys radio buttons group specifies the method of saved region coordinates recording in raster pixels or in pro jection coordinates Fields X1 Y1 specify the left bottom angle of the saved rectangular and can be set and updated by the user Fields X2 Y2 specify the right top angle of the saved rectangular and can be set and updated by the user Creating deleting cleaning and dubbing of raster layers To create delete clean or back up raster layers use Primary raster properties and Secondary raster properties items in Utilities menu of the main program window A dialog appears on the screen which controls are described in details the corresponding section of this user s manual Creating mosaics To create mosaics from several raster layers or to patch a multi date remotely sensed data one raster use Patch Primary R
82. to the specific neuron Minimal and maximal size the value corresponding to maximum and minimum of pixel related to this class is specified in percents Hierarchy tab Neural net classes hierarchy provides for storage of data about thematic inclusion of one group of classes into an other and for example can be the basis for the creation of hierarchical legend Besides the hierarchy 1 the basis for the creation of the corresponding neural nets hierarchy enabling to the make a detailed research of this or that group of classes Current neural net is recorded in the node of hierarchy and cannot be changed When loading the hierarchy tree from a file the neural net stored in the hierarchy node becomes current as if it is loaded from a file into the SOM window If current neural net size and the replacing neural net size from the hierarchy node are different the error message is popped up and the hierarchy tree would not load To refer neural net neurons to this hierarchy node they should be selected by clicking right mouse button on them with pressed Shift key and then dragged to the corresponding hierarchy node using the drag and drop Windows tool The same mechanism is used to drag neural net neurons from one hierarchy node to another by clicking left mouse button on the field of the relevant initial node and dragging it to the new node To build and change the hierarchy tree itself use Pop Up menu opened by clicking
83. ty v MAF Orde First C Second Initialization parameters Pow ho Relative rut Thematic data contained in the previously trained and calibrated Kohonen neural net is used in this postprocessing mode Based on this data thematic classes probability distribution is calculated for each pixel of the region being classi fied This distribution is used both to specify the initial values of classes and to build the final classification Notably already built classification can also be used as initial thematic classes values e g vector map thematic rasterization results if this map is updated Received initial thematic classes distribution is verified by iteration using contextual data to assign a thematic class to each pixel best matching with both distribution of probability received from neural nets and with classification of this pixel neighbors with the help of Gibbs energy minimization while applying Markov random field Checkbox Use Vector Layer as Border indicates the requirement to restrict the image area being classified with the regions contained in vector map file File selection push button enables to select vector map file containing region borders Field SOM to use contains the name of the previously trained and thematically calibrated neural net File selection push button enables to select file containing this neural net Button Comments enables to view the comments contained
84. used for mask creation are indicated below If both vector map and hierarchy mask are applied only pixels within both masks are used for training Teach tab enables to specify training parameters Teach SOM Neural Net a Tl xl Planes SOM Thematic Calibration Hierarchy Learning Parameters Sampling Step 3 Radius Ex Run Length 300000 02 4 Count of Pixels 2990430 Use Distortion Made Sorted Field Sampling Step pixels sampling step during training and calibration value 2 indicates the use of each second pixel in row Field Radius neighborhood radius value used in training Field Run Length total number of used pixels if the region being trained contains fewer pixels the latter are used several times 46 Field Alpha initial value of training rate Checkbox Use Distortion Mode enables to change training mode used if there was no self organization of neural net or classes ordination during regular training Button Teach initiates neural net training Button Tune initiates tuning of reference elements of neural net classes Checkbox Use Weights indicates the requirement to use weights during neural net training with reference to the vec tor map objects The bigger 1 the object weight the stronger 15 the influence of the corresponding image area on the neural net training results neural net thematic orientation List box Use Weig
85. ut tab controls Field SOM used to create classes specifies the neural net for classification Group of radio buttons Use e Classes radio button enables to use histogram of classes from slot indicated in the list below as input data for pattern SOM Resulting pattern SOM input data vector length equals to the number of classes in SOM used for classification of the slot indicated in the list and contains counts for corresponding classes e Calibration radio button If checked histogram of thematic classes is used as input data corresponding to the classes numbers saved in the layer pixel from the list below used SOM must be thematically cali brated Created pattern SOM input data vector length equals to the number of thematic classes and con tains counts for corresponding thematic classes extracted from the SOM used to classify layer e Palette radio button If checked histogram of distances within RGB space from the given class color in the slot indicated in the list below to color seeds in the loaded palette is used as input data Created pattern SOM input data vector length equals to the number of color seeds in the palette and contains contributions of seeds into the class color Checkbox Sammon s Distance if checked distance to neurons with color seeds in the packed ordination plane 15 used instead of RGB distance in Sammon s map List box SOM Classes Plane specifies the primary raster plane classified
86. utive table is presented as a sequence of zeros and units containing one unit on the place corresponding to integer 30 code if code value equals to 2 and the maximum code value in the file equals to 5 the sequence 0 1 0 0 0 is produced in the sampling row on the place of the code Checkbox Write Coord if checked center coordinates of the relevant pixel are added to each output file row Radio button group Output Format specifies the output file format Comma Delimited In rows delimited by commas Blank Delimited In rows delimited by blanks Fuzzy ARTMap In ARTGallery program format Button Sample initiates sampling Transformation of values taken from the relevant slots is the same normaliza tion as during neural networks training Creating recoding tables Recode Thematic Plane dialog Recode Thematic Plane B ajx Use MIF as Border 5 ample 5 ample Hecode T able Create Thematic plane to recade 7 LAM ecu 02120000905 610 chl warp Hun Close Checkbox Use MIF as Border specifies the requirement to limit working area with regions contained within vector map file Border file selection button enables to select vector map file containing working area borders Field Recode Table enables to select the name of table containing recode values Button Create opens dialog enabling to create recode table List box Thematic plane to recode specifies primary

Общее описание программы NeRIS

Contents

Download Pdf Manuals

Related Search

Related Contents