Land-cover mapping in the Arno basin, Italy: multispectral classification and neural networks

Two Landsat Thematic Mapper (TM) images covering the Arno basin, one of the major watersheds in Central Italy, have been classified using Neural Networks techniques. The main advantage in using Neural Network classifiers is that they do not require any a priori assumption in the statistical distribution of the class, since they are non-parametric classifiers. Furthermore, the ability of Neural Networks to “learn” and adapt to different situations makes them more flexible and capable of recognizing also the inputs with higher degree of noise. Different Network architectures have been trained and applied, and different levels of discrimination have been tested, i.e. various numbers of target classes. A two-layer feed-forward network with log-sigmoid transfer function has given the best performance. Results show that the recognition of some classes is excellent with Neural Networks, while for others there is still a large number of pixels incorrectly classified.