Lab 5

The difference between the overall accuracy and the kappa statistic is that the kappa statistic can tell you how much better, or worse, your classifier is than what would be expected by random chance. Since my kappa statistic was 85%, that tells me that my classifier is 85% better than a random assignment of cases to the various classes. Overall accuracy just tells us how well the pixels are sorted into classes. Overall accuracy is  computed by dividing the total number of correctly classified pixels (i.e., the sum of the elements along the major diagonal) by the total number of reference pixels.

Bare Soil and Developed High Intensity

Developed High Intensity and Bare Soil

I would try to include more training data for both developed high-intensity and bare soil, as there is lower user and producer accuracy for both when compared with the other classification classes. Also, noting the places where there is confusion, I would try to gather more diverse examples or focus on specific edge cases that the model struggles with. I could also create new features to better access the training data by removing irrelevant features or scaling existing features to improve classification accuracy.

For the maximum likelihood classification, I had an overall accuracy of 88%. For the minimum distance classification, I had an overall accuracy of 91%. The main difference between minimum distance and maximum likelihood classifications was that the maximum likelihood classification had a lot of unclassified pixels, which skews the results for producer accuracy and user accuracy. The classes that were most frequently confused were the developed high-intensity and bare soil, as roughly 700 pixels were confused between them. The difference in terms of characteristics is that each class has a different spectral pattern. For example, bare soil is distinctly different than water, and in the confusion table, we see that 0 pixels are confused among them. The spectral characteristics between each class are also significantly different, so we see that in the confusion table, there isn’t as much confusion between classes.

The changes that occurred between 2001 to 2016 through just visual inspection were that a lot of grassland/forest to the south of Houston was converted to developed housing. Also the west of Houston, there is a similar phenomenon occurring as well. There also appears to be a decrease in vegetation on the outer edges of Houston as more developed spaces took precedence. There also seems to be a decrease in pasture lands as more land is turned in developed land.

Wetlands, according to the data, are decreasing in size. The trajectory of change in wetlands is a decrease of 13.69% between the years of 2001 and 2016. The dominant drivers of these changes are mostly urbanization through habitat destruction, pollution, and altered hydrological regimes. Agriculture land use also is degrading wetlands and fragmenting habitats.

Two examples of change that seem to be unreliable are the increase in grassland and shrubs. Just looking at the change images visually, there seems to be a decrease in the amount of area for shrubs and grassland. However, in the change detection statistics, we see an increase in the image differences section. I can infer that urban/developing land is on the rise which causes a decrease in other land use types.

The change map points out specific areas that have changed. It gives a clear view of the areas that have changed from one class to another. The simple visual comparison doesn’t have that ability as we are just looking at two different maps side by side with no pointers at where the change is located. However, the change map is quite hard to read just by using the colors. Because there are so many change types, the colors on the map are hard to decipher. The spatial distribution of the changes in the region correlates with population growth. As more people are arriving, there needs to be a. increase in the amount of housing, which causes more land to be changed to developed land.