Lab 5 - Multispectral Classification


In this lab you will experiment with different classification algorithms with Landsat TM imagery from the state of Amazonas in Brazil.


First, open the BZ92_276...sub file in the /2002/lab7 folder in the class_lab folder on the desktop.

The first step is to select Regions Of Interest (ROI's) to represent the different land cover classes that you want to differentiate using the image data. Remember that the classes must be spectrally distinct for the classification algorithm to distinguish them.

 
Before picking the ROIs, look at a few different false color representations of the image data to get some idea of what you will be classifying. In general, time spent examining your data is time well spent. If nothing else, it will minimize the low grade embarassment of having casual observers ask you about features in the image that you failed to notice because you didn't examine it in detail.

 

Compare RGB: 3/2/1, 5/4/3, 6/4/1 side by side in 3 different image windows.

 
Q 1. Why is the 3/2/1 image so difficult to interpret?

Q 2. How would you enhance this image to improve interpretability?

 
Link the images (Functions -> Link -> Link Displays) and use the dynamic overlay tool (mouse) to compare specific features.

Choose your favorite of the 3 RGB images, blow away the other two and proceed selecting ROI's (Tools -> Regions of Interest -> ROI Tool). The ROI tool is described in the ENVI User's Guide and in the Tutorial as well as in the online Help.

You might pan around and  look at 2D scatter plots for (say) the NIR and visible red bands to look at the spectral signatures of land cover types present in the image. 

Select a single ROI for each of the following landcover classes:

Forest

Lake/pond

River

Mud/Sand

 

Remember to select "New Region" to add a new ROI class or you will keep adding to the first class and you will have to begin again.

If you do this, just hit "Delete" and start the class over.

 

Save the ROIs to a file (remember to "select all" before saving) using your initials as the first 3 letters in the filename (e.g. CSSexample.roi). Make sure you save it to the Class folder or your desktop.

Q 3. What tool would you use now to assess the spectral separability of the classes qualitatively?

Try it. Are they spectrally distinct? If not, you might want to reselect your training areas.

ENVI provides two statistical measures of the spectral separability of the classes in the ROI tool. If you have time, run it and see if your ROIs are spectrally distinct.
Q. 4. Which of the 2 measures provides a more conservative estimate of class separability?

 
First, from the Classification -> Supervised menu, run the Parallelopiped classification.

In all classifications, Output Result to Memory and Do Not output Rule Images.

 
Display the classified image (which will appear in your Available Bands list) in a new window as a gray scale image. Nice gray scale, eh?

 
Link the classification window with the RGB image containing your ROIs and compare the results w/ the dynamic overlay.

 
Q 5. What are all the black areas?

 
Now, from the Classification -> Supervised menu,  do a Maximum Likelihood classification and display the result in a new image window.

 
Q 6. What happened to the black areas?


Now rerun the MaxLike classification with a threshold.

Q 7. What threshold did you choose?

Q 8. What didn't get classified?

 

Now select additional ROI's for each class and add two more classes (and save to your file).

 

Q 9. What classes did you add to improve your classification?

 

Run the Maximum Likelihood, Parallelopiped and Minimum Distance classifications again using the new classes.

 

Summarize the results.

Proceed to the "Post Classification Processing" section of the Classification menu.

 
Calculate the Confusion Matrix (for the ROI method only) for both of your Maximum Likelihood Classifications.

 

Q 10. Why are they different?

 

Q 12. Explain the "off diagonal" numbers.

 

Summarize your impression of this exercise in one sentence.

Q 13. What did you learn?