Goals and Background:
The goal of this lab was to gain experience working with an unsupervised classifier. Unsupervised classifiers do not take training samples or other spectral information in order to give a classified land output. With parameters and multispectral raster data provided, they run fully automatically. The unsupervised classifier used in this lab was the ISODATA algorithm. We ran this classifier from within ERDAS Imagine. The classifier was run twice with different amounts of clusters. After the processing was completed, the clusters were examined individually and identified as the class that the majority of the cluster happened to cover. Unsupervised classification can produce a less than perfect result with salt and peppering of incorrect land cover spread over sections of correct land cover. This is a reason not to use this type of classification and to instead choose a supervised classifier or an object based classifier. It is important nonetheless to have experience with this type of classifier to understand how it works and give context to data that may be encountered with this type of processing.
ISODATA stands for Iterative Self-Organizing Data Analysis. The process takes a user input of the maximum number of clusters (classes automatically found), maximum number of iterations (number of times the process is repeated and statistics are recalculated), convergence threshold (maximum percent of pixels whose class values will remain unchanged between iterations), and maximum standard deviation (if this is reached for a cluster, the cluster is split in two). The process ends after the maximum number of iterations or the convergence threshold is met.
Methods:
Part 1 Section 1:
In this section the ISODATA algorithm was run on a Landsat 7 (ETM+) image from June 9th, 2000 covering Eau Claire and Chippewa counties. The algorithm was run in this part using The parameters shown in Figure 1. Notably, this part of the lab used 10 classes, and a convergence threshold of 0.95.
 |
| Figure 1 |
Part 1 Section 2:
After running the algorithm in the last section, the classes automatically generated needed to be assigned to an classifications. The classes they were assigned to are shown in
Figure 2.
 |
| Figure 2 |
The process involved having the image in a viewer in ERDAS Imagine, and opening the attribute viewer. The classes could then individually have their colors changed to an easily identifiable one, and using google earth linked to the viewer in ERDAS the areas covered by each class could be assigned to the most appropriate classification. A view of the completed process is shown in
Figure 4.
Figure 3 shows the classes before assignment and color coding.
 |
| Figure 3 |
 |
| Figure 4 |
Part 2 Sections 1 and 2:
Part two followed the exact same procedure as Part 1, however this time 20 classes were generated and the convergence threshold was lowered to 0.92. The parameters are shown below in Figure 5.
 |
| Figure 5 |
After the algorithm was run the same process as in Part 1 Section 2 was followed. The next step was to reclassify so that pixels previously assigned the same classification would no longer be assigned a number between 1 and 20, but now have a value between 1 and 5, the number of classifications we assigned pixels to. The output was then brought into ArcMap and a cartographically pleasing map was created. The output of the reclassification process can be seen in
Figure 6.
 |
| Figure 6 |
Results:
The outputs of Part 1 and Part 2 can be seen below in Figures 7 and 8. Part 2 produced a more pleasing output, with more accurate classifications. There were more accurate boundaries and there was less salt and peppering. Figure 9 shows how with 20 classes or clusters instead of 10 the urban classification does not extend outside of the boundary of the road nearly as much. Red was assigned to designate urban or built up land cover. The area outside of the road in this instance only had vegetative and bare soil cover. Figure 10 shows a map created with the data from Part 2 after all processing was completed.
 |
| Figure 7 |
 |
| Figure 8 |
 |
| Figure 9 |
 |
| Figure 10 |
Sources:
Lab instruction from Dr. Cyril Wilson. Landsat imagery is from the Earth Resources Observation and Science Center of the US Geological Survey.
Comments
Post a Comment