Supervised Land Cover Classification An introduction to digital image classification using the Multispectral Image Data Analysis System (MultiSpec )

Transcription

1 Supervised Land Cover Classification An introduction to digital image classification using the Multispectral Image Data Analysis System (MultiSpec ) Level: Grades 9 to 12 Windows version With Teacher Notes Earth Observation Day Tutorial Series Developed by The University of Georgia Chapter of the American Society for Photogrammetry and Remote Sensing (ASPRS)

2 Supervised Land Cover Classification Earth Observation Day Tutorial Series An introduction to digital image classification using MultiSpec for Windows computers Purpose To display, classify and verify the quality of the land cover classification of a section of a satellite image, using the Multispectral Image Data Analysis System (MultiSpec ) software and a supervised classification method. Student Outcomes Students gain experience in the use of remotely sensed image data. Students learn how to use an image processing tool to classify a remotely sensed image. Students gain/improve spatial or landscape level perspective of the area around their school. Science Concepts Geography The characteristics and spatial distribution of land cover types and ground elements surrounding the students school. How humans change their environment and how the environment changes human patterns of land use. Physics Energy-object interaction and the use of remote sensors to detect the energy reflected from objects. Color formation and color systems. Life Sciences The characteristics of vegetation communities surrounding the students school. The distribution of habitat types surrounding the students school. Scientific Inquiry Abilities Identify answerable questions. Conduct scientific investigations. Develop descriptions using evidence. Communicate results. Other Concepts Mathematics and Computing Math concepts are reinforced, such as average and ranges. Use of algorithms for digital image analysis. Specific Learning Objectives Upon successful completion of this lesson students will be able to: Inspect a Landsat-Thematic Mapper (TM) image to identify land cover types and ground elements. Classify a Landsat-TM image using a supervised classification method. Verify the accuracy of a land cover classification using reference data. National Science Education Standards Content Standard A: Science as Inquiry Abilities necessary to do scientific inquiry Understanding about scientific inquiry Content Standard D: Earth and Space Science Standards Structure of the earth system Content Standard E: Science and Technology Standards Abilities of technological design Understanding about science and technology Content Standard F: Science in Personal and Social Perspectives Populations, resources, and environments Science and technology in society 2

3 National Educational Technology Standards (NETS-S) Creativity and Innovation Apply existing knowledge to generate new ideas, products, or processes Identify trends and forecast possibilities Research and Information Fluency Locate, organize, analyze, evaluate, synthesize, and ethically use information from a variety of sources and media Evaluate and select information sources and digital tools based on the appropriate ness to specific tasks Process data and report results Critical Thinking, Problem Solving, and Decision Making Identify and define authentic problems and significant questions for investigation Collect and analyze data to identify solutions and/or make informed decisions Use multiple processes and diverse perspectives to explore alternative solutions Lesson Description Level Grades 9 to 12 Time Two sessions of 60 minutes. Session 1: image classification. Session 2: classification quality analysis and accuracy assessment. Time includes the presentation of the activity and background information, as well as hands-on activities. Materials Computer(s) running Windows operating system (if students will be using Macintosh computers, please refer to the Mac version of this tutorial). Note: we suggest students work as a group, with ideally two students per computer. MultiSpec software: this tutorial was created using MultiSpec version bit for Windows computers. Pre-selected section of a remotely sensed image acquired by the Thematic Mapper (TM) sensor onboard the Landsat satellite. Color printer (optional). The classification accuracy assessment activity requires students to have a calculator or access to spreadsheet software. Preparation This lesson is designed to follow the Viewing Remote Sensing Imagery with MultiSpec tutorial, also from AmericaView. Before beginning this lesson, students should be familiar with opening and visualizing a satellite image using MultiSpec software. See the accompanying Teacher Preparation Plan document for details. The MultiSpec software should already be installed on each computer to be used by the students as a result of completing the Viewing Tutorial. If some time has passed since you performed the Viewing... tutorial, you may wish to take some time to re-familiarize yourself and students with MultiSpec. The Landsat-TM image previously downloaded and clipped for the Viewing... tutorial may also be used for this tutorial. If you no longer have access to the Landsat-TM image used for the Viewing... tutorial, follow the instructions in the Teacher Tutorial.. to download the Landsat-TM image, identify the area to be classified and clip a section of the image representing the area to be classified. 3

4 Preparation (cont.) Save copies of the selected section of the Landsat- TM image on the computers the students will be using. Make sure students will have read and write permissions to access the necessary folders (e.g. where the Landsat-TM image is stored) when following the tutorial. Identify land cover types in the selected subsection of the Landsat-TM image. Make sure students will have access to the United States Geological Survey (USGS) EarthExplorer website through a web browser. If a printer will be used (optional) verify that it can be accessed by every computer used in this exercise and that the printer has ink. Make sure students will have access to a calculator or a spreadsheet during the quality assessment analysis. Preparation Time Image preparation time: if the image to be used in the tutorial is part of USGS s image archive (i.e., the image covering your area has already been processed and is available for download), preparation time may take around one hour. If the image covering your area is not available for immediate download, you need to request USGS to process this image. Normally, you will receive a notification from USGS that the image is available for download two to three days after you place the image processing request. The teacher should also consider allocating time to understand the basics of remote sensing, including image acquisition by remote sensors and the characteristics and use of the acquired images. This lesson may be presented in conjunction with the other AmericaView Earth Observation Day MultiSpec tutorials. These tutorials guide students to classify the land cover around their schools using an unsupervised classification method and to conduct a land cover change analysis. For more information on these tutorials, see the AmericaView Earth Observation Day website MultiSpec software and data copy time (setup time): MultiSpec and the section of the Landsat-TM image must be available on every computer to be used by this exercise. Setup time per computer is less than five minutes if a computer account with read/write permissions already exists (make sure students will be able to read from/write to the folder where the section of the Landsat-TM image is stored). 4

5 Lesson Overview This lesson will guide students in the process of classifying a section of a Landsat-Thematic Mapper (Landsat-TM) image into multiple land use / land cover classes. For that, students will use the MultiSpec image processing software and a supervised image classification method. The classification process will be followed by the verification of the classification quality and an accuracy assessment. The main steps involved in this lesson are: Display and inspect a section of an image acquired by the Landsat-TM satellite Classify a section of the Landsat-TM image using a supervised classification method Display and inspect the classification results Analyze results and verify the quality of the classification, through a classification accuracy assessment Suggested Sequence of Activities Present an overview of the exercise Present background information regarding remote sensing: the Landsat satellite, spatial resolution (pixel size), spectral resolution (bands), spectral response of selected land cover types, including vegetation, soil, water and constructed areas Present the basics of supervised classification Inform students on where the image to be classified (tmimage.tif) is located and where to save the results of the analyses Conduct image analysis: image display and classification Present the students the basics of classification accuracy assessment Conduct classification results analysis through the inspection of results and classification accuracy assessment. 1. Why are the colors for the same land cover types so different when I compare my image with the Powerpoint samples provided? Some difference in color between images is normal. If the color difference is considerable and all other elements of your screen seem to present their normal colors, you may have loaded your image using a color composite different from the suggested color-channel association. Reload the image and in the Set Display Specification dialog box make sure you have3-channel Color for Type and the values 5, 4 and 3 associated with the Red, Green and Blue Channels. Frequently Asked Questions 2. How many classes should I use? This varies highly depending on your area. You may have as few as two, or up to ten. Try distinguishing as many classes as possible to increase your accuracy. 3. Why is MultiSpec grouping several classes together? You can repeat the classification, increasing the number of training sets, for example. It is possible that the spectral patterns of the two land cover types are close enough that the software cannot distinguish them. 5

6 4. I cannot differentiate roads from commercial areas. What should I do? Developed areas may be difficult to differentiate in a satellite image. All are composed of mineral sand often have very similar reflectance patterns. It may not be possible to separate them when classifying images. 5. What do we do if there is an area in our image for which nobody knows the land cover Remote sensing is a powerful tool, but it may be the case that remote sensing alone cannot provide all the information necessary to classify certain land cover types. Projects incorporating remote sensing usually include field work. During field activities, the area being analyzed is visited and data are gathered to support information extraction and land cover classification using images. 6. Why do I see errors in the classification and what can I do to increase classification quality? Sometimes, the use of multispectral images and supervised classification methods may not be enough to classify the high variability and complexity of some natural and man-made systems. In some cases, you may be able to increase the quality of your classification by changing your training fields and avoiding edges. Also include more examples of each land cover type and run the classification again. Depending on classification quality requirements and the ground elements being analyzed, data from specific sensors and with specific characteristics may be required. 7. What factors affect the quality of the classification? The list of factors affecting the quality of your classification is long and include the complexity of your ground elements and study area; your input data (which are, for instance, affected by the atmosphere, acquisition geometry, shadows, Frequently Asked Questions (cont.) sensor failure and image noise); as well as those factors associated with the classification method and settings chosen. You also need training sets that adequately represent your land cover classes for the software to be more accurate. 8. When analyzing classification quality, what land cover type should I choose, if the reference image is showing multiple land cover types around the coordinates of a classified point? Your reference image on the USGS EarthExplorer website will probably have finer spatial resolution than your classified image. Because of that, it is expected that EarthExplorer will have multiple ground elements (represented by multiple pixels) inside a single Landsat-TM pixel. In most cases you can choose the predominant land cover type around a given coordinate when evaluating the results of image classification. For example, coordinates from the classified image may direct you to a gap in a forest, where you may be able to see bare soil and shadow, surrounded by trees. If the gap is not very large, the predominant land cover type contributing to the final Landsat-TM pixel value may still be forest. 9. The location of ground objects in MultiSpec does not seem to match the location of the same objects in the USGS EarthExplorer image. Remote sensing products need to be geometrically adjusted, so coordinates can be used for image navigation and feature location. These geometrical adjustments may not be exact and positional discrepancies may occur. If these differences seem to be consistent throughout your image section, you can consider the difference in position of a selected point (image offset) when locating objects. We suggest using a color composite of the Landsat-TM image and locating features not only by using coordinates, but also by locating other ground elements in the area (for example, roads, building and streams) and considering their relative position to the features analyzed. 6

7 10. Where can I find more information about remote sensing and digital image processing? For an introductory approach, we suggest the following books: Remote Sensing of the Environment: An Earth Resource Perspective, by John R. Jensen. Remote Sensing and Image Interpretation, by Thomas Lillesand, Ralph W. Kiefer and Jonathan Chipman Frequently Asked Questions (cont.) 7

8 Supervised Land Cover Classification Earth Observation Day Tutorial Series An introduction to digital image classification using MultiSpec for Windows computers Lesson Overview In this exercise, you will classify a section of a Landsat-Thematic Mapper (Landsat-TM) satellite image, or group into common land use types, showing your school and surroundings. You will use the MultiSpec image processing software and a supervised classification method to classify this image into multiple land cover types occurring in the area. Supervised classification requires the user to supervise or determine land cover types. Training Fields (or areas) of specified land types are selected by the user to provide the software with some characteristics for each area. For this exercise, the imagery will come from Landsat-TM. The Landsat-TM is a special sensor onboard the Landsat Satellite platform and is capable of capturing multispectral imagery. Landsat multispectral images capture many bands of the electromagnetic spectrum, including visible, reflective infrared and thermal infrared energy. The supervised classification process will be followed by image inspection and the verification of the classification quality. The main steps involved in this lesson are: Display a section of an image acquired by the Landsat-TM orbital sensor. Classify the Landsat-TM image using a supervised classification method Display and inspect the classification results Analyze results and verify the quality of the classification. Classroom Instructions: 1) Start by launching MultiSpec using the icon on the desktop. Note to Teacher: Please refer to the MultiSpec Teacher Tutorial preparation guide of this Tutorial Series on how to create a desktop icon to launch MultiSpec. 2) Load a Landsat-TM image into MultiSpec in the same way you loaded the image in the Viewing Remote Sensing Imagery with MultiSpec Tutorial. From MultiSpec s File menu, choose Open Image. An Open dialog box will be displayed, allowing you to select the image file to be loaded. Note to Teacher: A detailed description on loading and viewing remote sensing imagery is provided in the Viewing Remote Sensing Imagery with MultiSpec tutorial, available from AmericaView. You may refer the students to this tutorial if you feel they need a refresher on how to load and view these images. 3) Using the Open dialog box, navigate to the folder where the tmimage.tif file is stored. (The teacher will provide instructions on where to find this file.) Select the tmimage.tif image and click Open. The Set Display Specifications for: dialog box will open (Figure 1). 4) In the Display group of the Set Display Specification for: dialog box, make sure you have 3- Channel Color for Type and the values 5, 4 and 3 associated with the Red, Green and Blue Channels. Enter the value 2 for Magnification. Click Ok. If this is the first time this image is displayed, the Set Histogram Specifications dialog box will open. 8

9 Figure 1: Set Display Specification dialog box for loading the tmimage.tif image. Your values for Line and Column in the Area to Display group may differ from the ones represented in the figure. 5) In the Set Histogram Specifications dialog box, accept the default options by pressing OK. The tmimage.tif image will load. Remember that the Text Output window will now be updated with statistics referring to the values associated with each pixel of an image (the digital numbers). These numbers will be used by MultiSpec to display the image in the resulting window. The numbers are stored in a.sta file so that they will not need to be computed again. Note to Teacher: Image statistics are calculated and presented for each spectral band and reflect the average, range and variability of digital numbers in the band. One can use the Selection Graph feature of MultiSpec for a more detailed numerical analysis of the digital numbers in the image and specific land cover types and/or ground elements. 6) Visually inspect the displayed image and think about how pixels in the image might be classified into categories or different land cover types. How many different land cover types can you identify here? Do you see roads and constructed areas? What about vegetation and water? What other features can be determined in the image? Note to Teacher: The following steps will deal with land cover classification of the Landsat-TM image section. Land cover classes that may be present in your image are: bare soil, vegetation, constructed areas, water and shadows. Some of these classes can be subdivided and detailed (for instance, light soil, dark soil and dense vegetation, sparse vegetation). We suggest preceding the following steps of the lesson with a review of the Earth Observation Day Powerpoint presentation showing examples of how land cover types and ground elements are depicted by color composites of images acquired by the Landsat-TM sensor. 9

10 Note to Student: The colors displayed by the loaded image section depend on how we choose to load an image on the computer monitor. When we loaded the tmimage.tif image, we chose to associate band 5 of the Landsat image with the color red in your computer s monitor, band 4 with the color green, and band 3 with the color blue. Different associations between bands and displayed colors will produce different visual results. This is desirable at times. 7) We will now classify the tmimage.tif image by grouping similar pixels into categories. This method of classification is supervised, meaning that the person using the computer has to inform MultiSpec what types of land cover exist in the area and where they are found before proceeding. MultiSpec finds other pixels of that land cover type based on how similar the values of each pixel is to the training fields, the user provided. MultiSpec looks at each pixel of the image and associates the pixel to the class defined by the user selected training areas which is the most similar to that pixel. To prepare MultiSpec for the classification, from the Processor menu, select Statistics. The Set Project Options dialog box will open (Figure 2). 8) In the Set Project Options dialog box, leave the default settings and select OK. The Project dialog box will open. Figure 2: Set Project Options dialog box. 9) Next you will select training fields (Figure 3). To do that, draw a rectangle around a land cover type (for instance, bare soil) by clicking over the image, holding the mouse button and dragging the cursor to encompass pixels that correspond to the selected land cover type. Release the button when you are done. If you draw another rectangle, the first one is erased. You can zoom in ( ) or out ( ) to help you with selecting your training field. When you are done drawing your first field, go to Step 10 to learn how to save it. 10

11 Note to Student: MultiSpec gives you two options for selecting areas within your image: rectangles or polygons (irregular shape). Rectangles are the default. If you want to use the polygons selection, in the Project dialog box, check the square to the left of the Polygon Enter. To create polygons around the training area click once at each vertex, then double-click to finish the polygon. Be sure to turn the polygon option off before trying to select rectangular fields again. Figure 3: Training Field selection (dashed rectangle inside the circle). Note: Training Field Selection can produce inaccurate results if done improperly. Avoid edges of boundaries between different classes; choose areas representing a single land cover that are sufficiently large and homogeneous enough. Mixed use areas are those which have different uses within the same area, such as a school next to a forest and a factory. All pixels selected will be used by MultiSpec to locate the land cover categories for the rest of the image. Note how features can widely vary in intensity and color within one class, such as forests. For this reason, the user should select more than one training field for each class type. The number of classes will vary depending on your specific area. Note to Student: Many of you will not be used to viewing objects from overhead, so neighborhoods and other features might not be immediately recognizable. Image interpreters use a variety of elements when trying to identify land cover including but not limited to: (1) Location, (2) Size, (3) Shape, (4) Shadow, (5) Tone/Color, (6) Texture, (7) Pattern, (8) Height and Depth, and (9) Site/Situation/Association. Use one or all of these to help you determine what land cover an area falls into. 11

12 For example, let s consider your school. (1) Schools are located in many different places across the country, but they tend to be found in centralized populated areas. (2) They contain several large buildings typically. (3) Shapes may vary, but usually the buildings are rectangular. (4) Buildings will cast shadows. (5) Since you will only be seeing the roof, think of common roofing material colors such as black or grey. (6) The texture of school buildings are most likely smooth. (7) Schools are made up of several buildings that are close to one another. (8) Depending on your school, the building might be one or two stories and should stand out easily. (9) Schools are commonly found around neighborhoods and highways. Look for these features nearby. You should also use your local knowledge of the area to determine what features look like. For example, find a forested area nearby that you are familiar with on a resource like EarthExplorer or Google Earth. This will provide you with more examples to base your land cover decision on. 10) Back in the Project dialog box, click on the Add To List button. Under the Define Class and/or Field Description box (Figure 4), change the name of the class under Enter Class Name: to something appropriate to identify the land cover type. For the first time you do this process, the class will always be New since you do not have any created yet. The number of pixels for that field is also displayed. When you are finished, select OK to close the dialog box. Figure 4: Define Class and / or Field Description. 11) If there are other areas in the image that represent the same land cover class well, you can add them to this class by repeating steps 9 and 10. Every time you add a training field, make sure to select the correct class from the Class drop-down menu in the Define Class and/or Field Description dialog. Adding multiple training fields provides the software with more examples of what each class looks like, and will increase the accuracy of the classification. 12

13 12) When you are ready for a new class type, follow steps 9 to 11. Be sure to select New under the Class drop-down menu and name it appropriately. The new class name will appear in the Define Class and/or Field Description next time you draw a training field. Then you will simply select the appropriate class name from those on the list. Finish all of your training fields. Figure 5 shows a sample image with multiple training fields/areas outlined. Figure 5: Sample image with multiple training fields selected. 13) Under the Processor menu, choose Classify. The Set Classification Specifications window will appear (Figure 6). Under Write Classification results to: click the check box next to Disk file. Choose GeoTIFF as the output type. Click the check box next to Image Window Overlay. Select Add new overlay from the drop down menu. Unclick the Training (resubstitution) option under Classify:. Click the OK button in the Set Classification Specifications window. If prompted to update statistics, choose OK. Choose an appropriate place to save the file and be sure to name the file tmimage_supervised.tif. Click the OK button. After some processing, the results will be displayed on the screen. The Text Output window will present a list of the classes you created. 13

14 Figure 6: Set Classification Specifications dialog box 14) The classification results are overlaid on top of the TM image (Figure 7). Experiment toggling the overlay off and on a few times and inspecting how individual classes relate to the color composite image. To turn the overlay off, click on the red O drop-down menu button ( ) and select No overlays. To turn the overlay back on, click the red O button and select the all overlays option. After you finish analyzing the results, turn the overlay off. What happens to the image when you toggle the overlay on and off? Why do you think someone would want to have the ability to do this? Figure 7: Sample classification result of a Landsat-TM multispectral image overlaid on the Landsat-TM image. 14

15 Note to Student: Your classification results may display differently, as these results depend on the distribution of land cover classes around your school. 15) Now, open the classification image. This will be the same classification result you inspected as an overlay, but you will have more control over how classes are displayed. From the File menu, select Open Image to bring up the Open image dialog box. You may have to change the Files of Type drop-down menu to Thematic (*.gis; *.tif; *.clu). Then select tmimage_supervised.tif and click Open. The Set Thematic Display Specifications dialog box will open. Press OK, as the default settings are fine. A thematic image will be displayed. 16) You can control the magnification of the thematic image by using the and buttons in the menu bar. Use the Ctrl key with the mountain icons to adjust the zoom step factor and the magnification. Also, resize the tmimage_supervised.tif window by dragging its border. Arrange your windows side by side, to resemble Figure 8, below. Figure 8: Color composite of a Landsat-TM sensor image (left) and image classification result (right). Images are arranged side-by-side for comparison. The legend for the right image has been covered. 17) Note that the image classification result window is divided into two parts. On the left we have the legend for the cluster classification. The image with the result of our classification is on the right. Compare the classification results with the original image window. Note in Figure 9 (right image) how areas of bare soil, vegetation, constructed areas, and water were classified. (You may not have all these ground elements present in your scene; it will depend on the types of land cover around your school.) 18) Now it is time to inspect your results and identify your classes. There are several things you can do to make this task easier: a. In the legend, you can move the cursor over a color chip associated with a class, hold the Shift key down (the cursor will change to an open eye) and click the left mouse button down and up to cause the colors for that class to blink off and on, changing to white and back to the class color. 15

16 b. If you hold down both the Shift and Ctrl keys and then click the left mouse button down and up, then all of the other classes will blink off and on. c. Change the class names by double clicking on the name to the right of the color chip. An Edit Thematic Class Name dialog box will open. To change the name of the class, enter the new name and press OK. d. Also change the class color by double clicking on the color chip. A Color selector window will open. Select colors that suggest the type of land cover class you are representing (for instance, green for vegetation, brown for soil, blue for water you can be creative here). Select the new color and press OK. e. To save your class changes, go to the File menu and select the Save Thematic Class Info As option (this option is only visible when Classes is selected in the drop-down menu of the tmimage_supervised.tif window). The Save Thematic Class Information As: dialog window will open. Press Save. A confirmation window is presented. Click Yes to replace the original file. 19) Close the MultiSpec software by selecting File menu and then Exit MultiSpec. MultiSpec will alert you in case there is any unsaved change to your classification product and will allow you to save those changes. 20) Congratulations!! You just finished the image visualization and supervised classification tutorial, using a Landsat-TM image and the MultiSpec image analysis software. Note to Teacher: If a printer is available, following the inspection and editing of the classified image students may print their classification results, ideally in color. To print the classification results, give the following instructions to your students: a. Click anywhere inside the classified image window. The classified image window receives focus. b. Go to the File menu and select Print Preview. The classified image window then shows how the image will be printed (in case you change your mind and decide not to send your results to the printer, press the Close button on the print preview window and you will be taken back to the classified image). c. If the print preview looks ok, press the Print button on the classified image window to print your results. Note to Teacher: This is the suggested end of the first session of this activity, involving image visualization and supervised classification. An extension of this activity is proposed below, which can be developed with your students as a second session if time permits. This extension session will involve evaluating the accuracy of the land cover classification resulting from the first session, ideally a process which should be carried out after any classification project. 16

17 Supervised Land Cover Classification Earth Observation Day Tutorial Series Accuracy Assessment An introduction to digital image classification using MultiSpec for Windows computers Note to Teacher: This activity is a suggested extension of the Supervised Land Cover Classification exercise above. The following steps guide students in the evaluation of the quality of their classified product and include accuracy assessment procedures. Specific requirements associated with this activity include the per student/group of students availability of : a section of a Landsat-TM image (tmimage.tif file) a thematic image (tmimage_supervised.tif) representing the land cover classification of the tmimage.tif file a computer with the MultiSpec image processing software installed an internet connection and web browser a calculator or spreadsheet program Background Although classification algorithms do a great job grouping remote sensing satellite image data into classes, pixels can be misclassified or inappropriately grouped. This reduces the quality of the classification results and can impact products and decisions based on image processing and analysis. In order to better use products derived from remote sensing, it is necessary to understand and communicate their quality. This exercise will guide you through steps associated with quality analysis and accuracy assessment of your classification results. Accuracy Assessment provides a way to quantify the effectiveness of automatic classification efforts. Note to Teacher: Explain to your students that a detailed accuracy assessment, including an adequate definition of samples and the verification of the necessary number of sampled points, may take more time and resources than the group has available. For instance, accuracy assessment procedures often require that the points to be verified are chosen randomly and that sampled points are visited in the field. Often items are identified incorrectly from imagery but caught by field visits. The current activity will teach the students principles associated with accuracy assessment. The students will use a subset of the classified points to perform image verification by comparing the classification results of these points with high-resolution imagery. (This approach is considered a proxy for field work.) Lesson Overview We are going to evaluate the quality of our supervised classification by selecting classified points and by visually identifying their predominant land cover class. To do that, we will use high spatial resolution images as a reference. We will then use worksheets to record how well points were classified and will later calculate the accuracy of our classification. 17

18 Classroom Instructions: 1) If it is not already open, launch MultiSpec by using the icon in the desktop. 2) We will load into MultiSpec the same Landsat-TM image we used during the supervised classification procedure. From MultiSpec s File menu, choose Open Image. An Open dialog box will be displayed, allowing you to select the image file to be loaded. 3) Using the Open dialog box, navigate to the folder where the tmimage.tif file is stored. Select the tmimage.tif image section and click Open. The Set Display Specifications for: dialog will open. 4) In the Display group of the Set Display Specification for: dialog box, make sure you have 3- Channel Color for Type and the values 5, 4 and 3 associated with the Red, Green and Blue Channels. Enter the value 2 for Magnification. Click Ok. If the Set Histogram Specifications dialog box opens, follow step 5, other wise go to step 6. See activity 1 for a reminder about histograms. 5) In the Set Histogram Specifications dialog box, accept the default options by pressing OK. 6) After the tmimage.tif image loads, drag the border of the image window to enlarge it. 7) To be able to read latitude and longitude coordinates from the classified image, you need to know the Landsat-TM image map parameters. Map parameters include information about the image, such as coordinate system and coordinates for boundaries. We will read these parameters from the original Landsat-TM image and apply then to the classified image. Click anywhere inside the displayed tmimage.tif image and, from MultiSpec s Edit menu, choose Image Map Parameters The Set Map Coordinate Specifications window will open (Figure 9). 18

19 Figure 9: Set Map Coordinate Specifications 8) Use the space below to write down the parameters provided in the Grid Coordinate System and Zone options for your image, presented by the Set Map Coordinate Specifications window. Click OK when you are done. Grid Coordinate System: Zone: Note to Teacher: Landsat-TM images downloaded from the United States Geological Survey (USGS) use the Universal Transverse Mercator (UTM) projection and are therefore projected images not geographic coordinate system images. This projection uses coordinates that are in meters not geographic (latitude and longitude) coordinates. The map parameters above are necessary, so MultiSpec can convert from UTM coordinates to the geographic coordinates you will need to complete this exercise. 9) Then, display the results of the supervised classification you conducted during the previous session of this exercise (tmimage_supervised.tif file). From the File menu, select Open Image to bring up the Open image dialog box. You may have to change the Files of Type drop-down menu to Thematic (*.gis; *.tif; *.clu). Then select tmimage_supervised.tif and click Open. The Set Thematic Display Specifications dialog box will open. The default settings are fine, so press OK. The thematic image will be displayed. 10) After the tmimage_supervised.tif image loads, drag the border of the image window to enlarge it. 11) Now, we will apply the map parameters you retrieved before (from the Grid Coordinate System and Zone options) to the classified image. Click anywhere inside the displayed classified image to make sure it is the active window and from MultiSpec s Edit menu, choose Image Map Parameters The Set Map Coordinate Specifications window will open (See Figure 9). 12) Enter the parameters you wrote down to inform MultiSpec what grid coordinate system and zone to use with the classified image. After you enter the parameters, press OK to apply the changes and close this window. Note to Student: The changes to map parameters you just applied will be lost if you close the classified image. If that happens, you can reopen the classified image and repeat the procedures described above to reapply the parameters. Or save the image as a geotiff file using the Processor->Reformat-Change Image File Format menu item. 13) To enable the presentation of latitudes and longitudes for individual pixels, click anywhere inside the classified image window and go to the View menu. Make sure the Coordinates View option is selected. This will display the coordinate view within the window just above the image. 14) Use the drop-down menu on the top of the classification display window to select the Lat-Long (DMS) option, which stands for Latitude-Longitude Degrees, Minutes, and Seconds. Move your cursor over the image to see the coordinates change (coordinates are displayed to the right of the drop-down menu). These coordinates (latitude and longitude) will help you find classified points in the reference image. Note to Teacher: Before starting this activity, make sure your students will have Internet access and that the EarthExplorer website is not blocked and can be accessed by all computers used. 19

20 15) Launch a web browser and go to To visually identify the land cover of a given area, we will use the image display and navigation features of the United States Geological Survey (USGS) EarthExplorer website. 16) Using EarthExplorer (Figure 11), zoom in into the region you classified and get yourself acquainted with the tools available and the ways you can interact with the website. Examples of things you can do include: a. To the left of the image panel there s a text field where you can enter and locate an address. When you enter an address and press Show, a marker indicating the address is added to the image panel. b. You can also use the Add Coordinate button to the left of the image panel to directly enter latitude and longitude coordinates into EarthExplorer. A marker is added to the image panel indicating the location associated with the coordinates entered. c. On the image panel (right side of the screen) you can zoom in and out by using the vertical controls (the plus and minus signs located at the extremities of the vertical line). d. You can also hold down the left mouse button to pan the image. e. Important: notice that as you move the cursor over the image, the coordinates on the upper part of the image panel are updated. You can use these coordinates to locate sample pixels and compare your classification results with your visual interpretation of the land cover based on the image shown by EarthExplorer. Figure 10: The United States Geological Survey EarthExplorer website. 20

21 Note to Teacher: The following step involves the writing down of geographic coordinates of selected points, as well as image navigation based on these coordinates. We suggest preceding this part of the activity with a description of the geographic coordinate system and the representation of coordinates by using degrees, minutes and seconds (DMS). Try incorporating Cartesian coordinate systems from mathematics class. Geographic coordinate systems are a way of assigning a unique positional value to every location on Earth. DMS is one way of assigning those values, where degrees are the angular measure from some central point, such as the equator, to the location. Minutes and seconds are merely smaller divisions of degrees. Other associated concepts may be also presented, including cardinal points (North, East, South and West directions), parallels (lines of latitude) and meridians (lines of longitude). 17) Fill out TABLE 1 by selecting multiple sample points over the supervised classification image. Use the first two columns of this table to list the sample coordinates (in latitude and longitude). Enter also the land cover class resulting from your analysis of the supervised classification and the Landsat-TM image. Then, use the coordinates of each sample to display the same location in EarthExplorer. In the last column of TABLE 1, enter your interpretation of the land cover at each sample location, based on your visual interpretation of the EarthExplorer image. Remember to draw from the examples of land cover classes discussed in class. While working on this, make sure: a. You select approximately the same number of points for each class (for instance, if your classification has five classes, you can select five points to represent each of the classes, totaling 25 samples). b. You select points that are relatively well distributed over your classified image and not concentrated in a single particular location. Classification might be great in one section of the image, but poor in others. Note to Student: The EarthExplorer image will probably show more detail than your Landsat-TM image. Because of that, it is expected that EarthExplorer will have multiple ground elements (represented by multiple pixels) inside a single Landsat-TM pixel. In most cases you can choose the predominant land cover type around a given coordinate when evaluating the results of image classification. For example, coordinates from the classified image may direct you to a gap in a forest, where you may be able to see bare soil and shadow, surrounded by trees. If the gap is not very large, the predominant land cover type contributing to the final Landsat-TM pixel value may still be forest. Essentially all of the ground elements are averaged over the size of the pixel, which for Landsat-TM images is 30 meters. The dominant type wins in this averaging process. 18) After you finish filling out TABLE 1, you can calculate the accuracy of your classification. To do so, use TABLE 2 below (an Error Matrix) to compile the data from TABLE 1. Consider the following: 21

22 Supervised Classification a. Each cell of TABLE 2 should have the number of times a class from your classification corresponded to a certain class from your reference data (EarthExplorer image). b. Notice that diagonal cells indicate a match between your classification and the reference data. c. Although space for seven classes is provided, you may not have all these classes around your school and can leave cells blank. d. After you have finished transferring the counts resulting from your classification comparison into TABLE 2, calculate totals for rows and columns by: i. Adding all cell values for the first column and enter this value at the bottommost cell of that column. Repeat the same procedure for the other columns, with the exception of the last column. ii. Adding all cell values for the first row and enter the total in the rightmost column. Repeat the same procedure for all rows. e. To calculate the overall accuracy of your classification, add the values for the cells in the diagonal (bold cells), excluding the lower-right cell, and divide the total by the value in the lower-right cell of the table. If you multiply your result by 100, you will get an estimate of the percentage of the points that were correctly classified in relation to the total number of classified points. Sample Error Matrix Reference/Validation Data Dark soil Light Soil Green Vegetation Non-green vegetation Constructed Water Row totals Dark soil Light soil Green vegetation Non-green vegetation Constructed Water 4 4 Column totals Sum of the diagonal = 14 Total points sampled = 24 Overall accuracy: 14/24 = = 58.3% 22

23 f. Calculate also the accuracy of the classification for each class, in relation to the total number of samples for the class. This accuracy calculation is called User s Accuracy, and provides a quantitative measure of how many pixels are actually what they say they are. To do that, use the Error Matrix and Fill in Table 3 with the numbers of correct classifications (the diagonal numbers) and the number of total samples (the column totals in the bottom row). To calculate the accuracy column, divide the diagonal value for the each row by the row total, as shown below. Note to Teacher: Although other ways to estimate classification accuracy exist, for simplicity, we are restricting this tutorial to the calculation of the User s Accuracy only. Other examples include Producer s Accuracy. g. Do all your classes show the same User s Accuracy values? Think about potential reasons for the classification accuracy to be different when considering different classes. Sample User s Accuracy Calculation Class Correct classification Total samples Accuracy Dark soil 2 4 2/4 = 50% Light soil 1 4 1/4 = 25% Green vegetation 3 4 3/4 = 75% Non-green vegetation 2 4 2/4 = 50% Constructed 2 4 2/4 = 50% Water 4 4 4/4 = 100% 19) Be sure you have answered all the discussion questions at the end of this tutorial. Following your teacher s instructions, you may wish to print your results or discuss them with the class. Congratulations!! You just finished the accuracy assessment part of the unsupervised classification tutorial, using a Landsat-TM image and the MultiSpec image analysis software. 23

24 Note to Teacher: The following pages contain student questions and the tables used in this tutorial. We suggest printing these pages and providing them to your students as worksheets. 24

25 Student name/group: Date: / / Supervised Land Cover Classification with MultiSpec : Discussion Questions Answer the following questions in the space provided below. If you need more space, continue on a spare sheet of paper. 1) What other ways would it be possible to use this technology for the benefit of your community? 2) Were you surprised by the types of land cover you found nearby or the amounts? Why or why not? 3) What human impacts on the land are apparent from observing the imagery? 4) Why might people use satellite imagery instead of visiting a place in person to find out about it? 5) What were you most surprised to learn about your community? 25

26 Student name/group: Date: / / TABLE 1 Accuracy Assessment Tutorial: Classification results and reference data comparison for sample points # Latitude Longitude Land Cover Land Cover (Supervised Classification) (EarthExplorer image) 1 Example: Green vegetation Constructed

27 Supervised Classification Student name/group: Date: / / TABLE 2 Error matrix for supervised classification Reference/Validation Data Row totals Column totals Diagonal total = Overall accuracy = = % Total number of points sampled = 27

28 Student name/group: Date: / / TABLE 3- User s Accuracy Calculation Class Correct classification Total samples Accuracy 28