Classification of multi-spectral remote sensing data with neural networks: a comparative study

Satellites or planes generate remote sensing images by simultaneously recording 'grey-level' images for a number of wave-bands. The resulting images are usually processed using statistical classifiers to extract features such as roads, built-up areas, vegetation or water. In the present study two types of neural networks, a multi-layer perceptron (MLP) and a Kohonen learning vector quantization (LVQ) network are tested as pattern classifiers. The results are compared with a nearest neighbour classifier (KNN). The aim of the study is to extract five classes: (1) roads, (2) buildings, (3) vegetation, (4) water and (5) derelict sites from data obtained using multi-spectral images of Stoke-on-Trent with a pixel resolution of roughly 4*4 m. The architecture and learning parameters of each network were optimised for 4005 training pixels selected randomly over the image (891*3989 pixels). Both network types and the statistical classifier were tested on 3552 test patterns. Standard back-propagation was used to train the MLPs while oLVQ1 and LVQ3 training were used for the LVQ networks.