Comparison of Neural Classification Algorithms Applied to Land Cover Mapping

We compared the performance of several supervised classification algorithms on multi-source remotely sensed images. Apart from the Multi-Layer Perceptron, K-Nearest-Neighbour and Radial Basis Function network approaches, we looked more in detail at the Support Vector Machine classifier, which recently showed promising results in our setting. In particular, it is able to provide meaningful answers for the analysis of mixed pixels. They correspond to areas on the ground that comprise more than one distinct class, representing a major challenge for the interpretability of the final land-cover maps. To assess their impact, we performed a rejection-based analysis, allowing classifiers to refuse answers on pixels they can not associate mainly with one class. The experimental results lead to the conclusion that the 1vs1 SVM approach with a linear kernel (using Bradley-Terry coupling) has to be preferred over all other classification algorithms examined, both in terms of accuracy as well as ease of visual interpretation.