Considerations for using sampling polygons in supervised satellite image classifications

Abstract

The dynamics of geographical or natural space can be studied through different methods and techniques. One way is to use remote sensing techniques to generate thematic maps from the classification of satellite images. In general, supervised classifications are used, since their accuracy will depend on the user's perception and experience, as long as some basic rules are obeyed. The concern of this study was to define the quantity of samples referring to the area worked for its classification. The sampling given by the training areas should be related to their spectral signature, i.e., the portions previously chosen as models for obtaining land use or land cover classes. The idea was to compare the results obtained in this sample, which included about 2% of the area used in the previous procedure, with those of that simulation. The hypothesis assumed that there should be a stabilization trend in the data, according to the increment of training areas given by sampling polygons, lower than indicated by previous experimentation. The simulations performed, however, did not confirm the hypothesis presented, and the results obtained indicated a balance very close to that found in a much larger area.