Training of neural networks for classification of imbalanced remote-sensing data

The multilayer perceptron is currently one of the most widely used neural models for the classification of remote-sensing images. Unfortunately, training of multilayer perception using data with very different a-priori class probabilities (imbalanced data) is very slow. This paper describes a three-phase learning technique aimed at speeding up the training of multilayer perceptrons when applied to imbalanced data. The results, obtained on remote-sensing data acquired with a passive multispectral scanner, confirm the validity of the proposed technique.