Labs+11+and+12

Labs 11/12 - Classification and Change Detection

Below are the confusion matricies for 1998 (left) and 2010 (right) you can see that the lowest user accuracy in 1998 is for fallow land at 37.04% which is not acceptable. For 2010, it is still fallow land with 30.71%. Based on these numbers, my classification was a failure because those values are no where near what I would call acceptable accuracy values. However, the overall accuracy is in the 90s for both, which is a bit misleading...
 * Questions:**
 * 1.Describe which class has the lowest user accuracy. Is that accuracy still acceptable?**


 * 2. Compare the two classified images and state what you find. How do they compare?**

Well, you can see that the image on the right (2010) has a vastly greater area covered by sparse forests (green) and less of the barren rock (brown). It also has more salt deposits (yellow) and more agriculture (bright green). However, you can tell in both images that the fallow land (light purple) which should be contained to the area near the agricultural valley is often confused for barren rock, an issue I knew I'd have. Also you can tell that this problem is much worse in the second image (right, 2010).
 * [[image:1998_training.JPG width="460" height="481" caption="1998 Classified Image"]] || [[image:2010_classified.jpg width="516" height="480" caption="2010 Classified Image"]] ||

Based on the confusion matrix, the 1998 classification is more accurate. However, you still have to take into consideration the less than 1.9 separability index for fallow land/ barren rock and the extremely low user accuracy for fallow land.
 * 3. Which classification is more accurate?**
 * 4. What is the largest changing class in square kilometers?**

According to the change statistics (below), the barren rock has a change of 494.27 square kilometers and in the last row, it is a negative amount with means it has decreased by 436.74 square kilometers. This is the basis for me falsifying my classification - it is impossible for 500 square km to change into sparse forest or fallow land or anyting else - barren rock on the side of a mountain will always be that, unless it is covered in snow and then it might be misclassified. It is difficult to compare such a solid classification image and a hazy two colour multiview image. However, I think using some of the PC bands from 2010 (link below) that you can definitely see the changes in the mountain tops on the left side of the images -- they are stark white indicating big changes and that is where the classification shows growth in the sparse forests. Also, the agriculture-filled valley has vastly contrasting colours in it in the PC bands from 2010 which show the changes of crop rotation per the different years. Also in PC band 1 from 2010, you can see exactly where the salt deposits grew dramatically (deep black) in the classified image from 2010. Lab 9 PC Analysis
 * 5. Compare your results with the 'change detection lab' from report 2. Discuss how the changed classes are visible in your change detection.**

Other lab 11/12 goodies: the validation separabilitiy matricies!


 * || [[image:1998_validation.JPG width="563" height="630" caption="1998 Validation"]] ||
 * || [[image:2010_validation.JPG width="577" height="710" caption="2010 Validation"]] ||