| Data obtained through remote sensing is a powerful tool for wide-scale land cover classification, but the process is complicated by several factors. Most land cover classification techniques are purely spectral, and classify an image pixel-by pixel. These methods ignore possible spatial correlation effects and must try to fit land cover into semantic categories that may not fit the physical characteristics of the area of interest. Many conventional classification methods also assume that input data is normally distributed and at a ratio scale, which greatly limits the type of data that can be used. I present a method of land cover classification that accounts for spatial correlation patterns in land cover. The characteristics of all pixels in a neighborhood are studied to find a class for each pixel in the image. The rules that use this spatial correlation information are evolved using adaptive computing techniques based on what is found to be useful over many iterations. The method I present can also freely use ancillary data in addition to data at a ratio scale. I find that for two study areas, my method matches the overall accuracy of conventional classification techniques. Overall accuracies for both conventional techniques and my adaptive computing technique were 65% and 80% for each study area respectively, with a similar distribution of error. |
