F2 - Fire 2 module: Remote Sensing Data Classification F2.1 Task_1: Supervised and Unsupervised classification examples of a Landsat 5 TM image from the Center of Portugal, year 2005 F2.1 Task_2: Burnt area Maximum Likelihood Classification example of a Landsat 5 TM image from Acre, Brazil, year 2005
F2.1 - Task 1: Landsat 5 TM Portugal, 2005 Dataset: The image for this module is a Landsat 5 TM for Portugal (path 203 row 32) from 27 th of September, 2005 (six reflective channels and a thermal band). Exercise 1 (working directory: TASK_1\EX1\). objective: Compare the different combinations of Landsat TM channels (RGB color composites), and the spectral contrasts obtained for different land cover types. 1. Load all the bands of the Landsat 5 TM 20332 dataset from Portugal continental from the year 2005. Select the bands b1 b7 (or LS_PORTUGAL_2005.LYR) from the working directory; 2. A stretch can be applied to image. Right click on the layer, select properties, symbology and then histograms to manually adjust the histograms; 3. Select an area with different land cover types where you will extract a small window. We will use a shapefile as a mask. Load window_msk.shp; 4. Zoom to the extents defined by this vector file. Right click on the layer and select zoom to layer; 5. In the ArcToolbox select Composite Bands to group the bands into a single dataset in the area defined by the window shapefile. Load all the TM bands, set the output file name and in the environment settings, go to general settings and set output extent the same as the window_msk.shp file. The output grid file (wlsport) is shown in figure F2.1.1.; 6. Load the wlsport grid file and set different RGB color composites. For example, RGB 7,4,3; RGB 3,2,1; RGB 5,4,3, and RGB 6,4,3. You should observe contrasts similar to the ones in figure F2.1.2; 7. Identify different land cover types present in the image. What do you see? For the purpose of identification of burnt areas what are the best RGB band combinations? And the worst? Why? F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 1
Figure F2.1.1. Selection of the window area extracted from the Landsat 5 TM image (path 203 and row 32), from 27th of September 2005. F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 2
RGB 7,4,3 RGB 3,2,1 RGB 5,4,3 RGB 6,4,3 Figure F2.1.2. Different RGB color composites for the same window extracted from the Landsat 5 TM image (path 203 and row 32). F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 3
Exercise 2 (working directory: TASK_1\EX2\). objective: to analyze the spectral discrimination between different land cover types and produce a MAXIMUM LIKELIHOOD classification of five land cover types. 1. Load the grid file output from the previous exercise (wlsport) and create a new polygon shapefile in ArcCatalog, which will be used to digitize the training areas; 2. Digitize Regions Of Interest (ROI) according to the following six landcover classes: - agriculture (code 1) - forest (code 2) - shrublands (code 3) - urban areas (code 4) - water bodies (code 5) - burnt areas (code 6) You can add the example vector file to compare it with your own digitized ROI file (Portugal_2005_ROI. Shp) Figure F2.1.3. Example of the training areas on-screen digitized in each of the six pre-defined land cover/land use types. F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 4
3. Create the signature file for the ROIs digitized in the previous step using the ArcToolbox (Spatial Analyst Tools -> Multivariate -> Create Signatures). To create this file all the image (window) bands will be used as well as the ROIs file; 4. The output file will then be open in an Excel worksheet to create an average spectral profile for each land cover/land use class. According to the given training areas in the example, you can find the means and variance-covariance matrix for each land cover class using all the seven TM bands in the Portugal_2005_ROI.gsg file. Figures F2.1.4 and F2.1.5 show the analysis; 140 agriculture forest shrubland urban water burnt 120 100 digital number 80 60 40 20 0 TM1 TM2 TM3 TM4 TM5 TM7 TM6 Figure F2.1.4. Mean spectral signatures for each pre-defined land cover/land use type, using the seven TM bands and according to the digitized ROIs. Figure F2.1.5 is also presented to show the spectral variability of each land cover/land use class represented trough the bubbles size. F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 5
agriculture forest shrubland urban water burnt 160 140 120 digital number 100 80 60 40 20 0 0 1 2 3 4 5 67 76 8 TM channel Figure F.2.1.5. Spectral variability and discrimination between the six land cover/land use classes. 5. Based on the mean spectral profile obtained for each of the six land cover/land use types, what are the most relevant aspects detected in the figures above? Why there is a pronounced decrease in the digital numbers from band 1 to band 2? What do you think about the spectral detectability of the different bands? F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 6
6. The signature file just created from the ROIs in step 3 will now be used to perform a supervised classification using a maximum likelihood algorithm. In the ArcToolbox you will choose Spatial Analyst Tools -> Multivariate -> Maximum Likelihood Classification. Please note that you will have to load all the seven bands from the Landsat TM windowed image to perform the classification. The output raster grid example was named as port_class_ml; 7. The classified raster will have six classes. In order to remove some noise from the classified image you may pass a majority filter, in order to replace salt and pepper class observations by the majority of class of their contiguous neighboring cells. You may use a kernel with four or eight neighboring cells (Spatial Analyst Tool -> Generalization -> Majority Filter). The output file was named as Port_Filt. The classification and the filtered grid files are shown in Figure F2.1.6; a) b) 1 - agriculture 2 - forest 3 - shrublands 4 - urban areas 5 - water bodies 6 - burnt areas Figure F2.1.6. Maximum likelihood classification into six land cover classes. a) output classified grid image; b) classification result from the application of a majority filter with a kernel size of 8x8. 8. Compare the results of the Maximum Likelihood classification with a RGB color composite chosen in exercise 1. What classes do you think are best classified? Does it make sense with the previous analysis of the spectral profiles of the different land cover classes? Which classes you found more difficult to classify and spectrally discriminate? F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 7
Exercise 3 (working directory: TASK_1\EX3\). objective: Apply the ISODATA algorithm to produce an unsupervised classification and compare the results with the ones from the supervised classification 1. Load the grid file output from exercise 1 (wlsport). In the ArcToolbox go to the Spatial Analyst Tools -> Multivariate -> Iso Cluster, load all the seven bands of the TM data study area and set the number of output classes, as well as the optional fields as shown in Figure F2.1.7. Here as an example we set the number of output classes equal to 10, 100 of iterations, the minimum class size equal to 50; Figure F2.1.7. Iso Cluster analysis with an example of optional settings, to produce the signature files from the selected study area (wlsport). 2. The output signature file (named as portugal_2005_class_iso.gsg) produced in the previous step will then be used to generate a probability map for each one of the 10 proposed land cover classes; 3. In the Spatial Analyst Tools -> Multivariate -> Class Probability module you will create the probability map for each of the 10 land cover classes, giving again as input data the TM bands of the windowed Landsat image, and the signature file created in the previous step. The output 10 layer-dataset was named as class_is; F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 8
4. Look at each of the 10 classes probability rasters and try to identify what is the land cover/land use correspondence to each class map. What are the main spectral confusions that you may detect between land cover/land use classes? Could you, based on this classification, be able to produce a binary burnt/unburnt map? Evaluate the spectral variability of burnt areas looking at the classes where burnt areas have high values of probability. See Figure F2.1.8 as an example. a) b) Class 1 - probability >= 90 Figure F2.1.8. a) RGB 7,4,3 color composite of the windowed study area of Landsat TM data (wlsport), with the class 1 observations with probability above 90; b) probability map of class 1. F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 9
F2.2 - Task 2: Landsat 5 TM Acre, Brazil, 2005 Dataset: The image for this module is a Landsat 5 TM for Brazil (with 6 bands) from 13 th October 2005 (TASK_2\EX1\LS_BRAZIL_2005). Exercise 1 (working directory: TASK_2\EX1\). objective: Perform a supervised classification to the Landsat TM only for Burnt and Unburnt classes. 1. Load raster dataset LS_Brazil_2005. Please try some different RGB colors composites with different bands and different stretches; 2. Create a new shapefile and collect several Regions of Interest (ROI) for Burnt zones and Unburnt zones (vegetation). Give different codes to Burnt and Unburnt classes (for instance 0 for Unburnt and 1 for Burnt). Please note that in this exercise we are including in the same Burnt ROI all kinds of vegetation types that could possible burn; F2.2.1. Examples of ROI for Burnt and Unburnt classes (0 for Unburnt and 1 for burnt classes). 3. Create signatures files for the ROI s collected in the previous step (Spatial Analyst Tool- Multivariate-). Save the file in the User Directory for EX1 with the extension gsg and export to Excel to create an average spectral profile for all the Landsat bands for each ROI class; F2.2 - Task 2: Landsat 5 TM Acre, Brazil, 2005 F2.2/pg. 1
40 35 30 Percent Reflectance 25 20 15 10 5 0 1 2 3 4 5 7 TM Bands BURNT UNBURNT F2.2.2. Examples of Average Spectral Profile for Burnt and Unburnt classes for all 6 Landsat Bands in Percentage Reflectances. 4. With the ROI data you will now perform a supervised classification to the image. Please choose the Maximum Likelihood classification algorithm (Spatial Analyst Tool- Multivariate-) and input the signature file (gsg file) to create a classified raster with 0 and 1; 5. Apply a filter to the classified image to create a new raster. You can choose the majority Filter (Spatial Analyst Tool- Generalization-) which replaces cells on the majority of their contiguous neighboring cells (use kernel with four or eight cells); F2.2.3. Example of the application of a majority filter to a classified image (kernel 4x4). The color red is the burnt area for the classified image. F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 2
6. Now you will extract only the burnt class. From previous filtered raster convert to new feature (shapefile) only for the burnt class. You are now able to compare the classification results with the original image; F2.2.4. Example of the Burnt class shapefile applied to the original image. F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 3
Exercise 2 (working directory: TASK_2\EX2\). objective: Perform a supervised classification to the Landsat TM but now for two types of Burnt Classes (Forest and agriculture) and one Unburnt class. Compare the two classification methods and the results. 1. Repeat steps 1 and 2 from the previous exercise but now add a new class to the ROI shapefile (you could take the ROI`s from the last exercise and reclassify the areas of burned agriculture to a new class with code for instance 2). You could add a new field to the ROI table to reclassify the ROI s (Figure 4).Save the file in the User Directory for EX2; F2.2.5. ROI table with a new field for the new class Burned Agriculture (code 2). 2. Create signatures for the previous ROI and export to Excel as you did in the previous step 3. Save the signature file in the User Directory for EX2 with the extension gsg; 3. With the ROI data perform a supervised classification to the image. Please choose again the Maximum Likelihood classification algorithm (Spatial Analyst Tool- Multivariate-) and input the new signature file (gsg file) to create a classified raster with 0,1 and 2; 4. Apply again a majority filter and extract now the two burnt class (compare with the original images and with the results of the first classification; F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 4
brazil_2005_classification2_burned_agriculture brazil_2005_classification2_burned_forest F2.2.6. Two types of burnt vectors extracted from the classified image. In yellow the burned agriculture class and in black the forest burned class. 5. See and compare also the spectral profiles for the two classifications regarding the burnt classes; F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 5
30 25 20 Percentage Reflectance 15 10 5 0 1 2 3 4 5 7 TM Bands BURNED FOREST BURNED AGRICULTURE BURNED CLASS (Class. 1) F2.2.7. Example of Average Spectral Profile for Burnt classes of the two classifications for all 6 Landsat Bands. The green line expresses the average spectral profile for the Burnt class of the first classification (Exercise 1).. F2.1 - Task 1: Landsat 5 TM Portugal, 2005 F2.1/pg. 6