Development of a combined GIS, neural network and Bayesian classifier methodology for classifying remotely sensed data

Abstract

This research is aimed at the solution of two common but still largely unsolved problems in the classification of remotely sensed data: (1) Classification accuracy of remotely sensed data decreases significantly in mountainous terrain, where topography strongly influences the spectral response of the features on the ground; and (2) when attempting to obtain more detailed classifications, e.g. forest cover types or species, rather than just broad categories of forest such as coniferous or deciduous, the accuracy of the classification generally decreases significantly. The main objective of the study was to develop a widely applicable and efficient classification procedure for mapping forest and other cover types in mountainous terrain, using an integrated GIS / neural network / Bayesian classification approach. The performance of this new technique was compared to a standard supervised Maximum Likelihood classification technique, a “conventional” Bayesian / Maximum Likelihood classification, and to a “conventional” neural network classifier.

Results indicate a considerable improvement of the new technique over the standard Maximum Likelihood classification technique, as well as a better accuracy than the “conventional” Bayesian / Maximum Likelihood classifier (13.08 percent improvement in overall accuracy), but the “conventional” neural network classifiers outperformed all the techniques compared in this study, with an overall accuracy improvement of 15.94 percent as compared to the standard Maximum Likelihood classifier (from 46.77 percent to 62.71 percent). However, the overall accuracies of all the classification techniques compared in this study were relative low. It is believed that this was caused by problems related to the inadequacy of the reference data. On the other hand, the results also indicate the need to develop a different sampling design to more effectively cover the variability across all the parameters needed by the neural network classification technique, while being, at the same time, economically feasible.