The use of multilayered perceptrons for remote sensing classification with temporal data

The task-based multilayered perceptron (MLP) is a variant of a class of non-linear classifiers based on the neural net construct. This paper describes the use of MLPs in classifying a remotely sensed image of a farm property into discrete ground cover classes, using LANDSAT TM image data. The methodology derived removes the burden of net configuration from the user. Use of a priori information, derived from the data and their class separability, is made in the selection of the net variables and architecture, to assist in convergence towards a global error minimum during training. A node reduction technique known as task-based pruning is also used to reduce and optimise the MLP architecture. A generalized network based on multi-temporal data of the property is constructed and a comparison with maximum likelihood classification of the same property are made, the MLP approach producing equivalent, or better, classified images when validated against the available ground truth