Last week I conducted my first supervised CART classification of a subset of my study area in eCognition. I based it on a set of training points, and allowed the classification to develop “signatures” based on the object’s mean reflectances, mean slope, standard deviation of NDVI (as a measure of homogeneity), asymmetry, border index (how jagged an object border is) and how well an object resembles a rectangle. I also tested using border to built up/urban as a way to determine which class an object belonged to.
The classification is OK, not perfect, but not completely wrong either. It’s a start. What I can see so far is that I might have more built up areas (red) than I should have, and some strangely placed agricultural fields on a mountain.
To improve the classification, I can change the features I choose to include in the training of the classifier, and maybe add some that are more important. For example, the “border to built up” feature might be what gives me so many built up objects. Another improvement could be in the segmentation, to reduce the number of “mixed objects” (compare mixed pixels). A third option that could improve my classification would be to add, edit or change my training data. And also, using a subset means that we’re only using the points within that subset, so when doing the real classification, with 300 training points, I might get a better result.