Abstract:
This paper illustrates problems and solutions for the design and configuration of neural network architectures used to classify remotely sensed multi-temporal imagery. A methodology is presented that avoids many of the traditional problems associated with neural network design, thus enabling the technique to be applied directly, without placing the burden of configuration on the user. The structure inherent within the data is used to derive both the network architecture and starting weights. Experiments with agricultural landuse classification show that the resulting networks classify at least as well as more traditional statistical approaches.