An optimized approach to application of neural networks to classification of multispectral, remote sensing data

Satellite image processing is one of the key research areas in the area of remote sensing. Remote sensing derives immense applications from this field like terrain analysis and generation, topographic mapping. Traditional statistical approaches provide reasonable success in this field, but the efficiency is limited with respect to the robustness of results. The statistical approaches are parametric, based on an assumed statistical distribution and hence the efficiency and correctness of results closely correlates to the proximity of data to the assumed distribution. Feed-forward neural networks can be trained to learn pixel classes and hence can be applied to the area of satellite image segmentation. This paper describes a technique developed to select training parameters and collection of training sets. An algorithm to accelerate the training process and reduce the time for classification is also explored. This paper provides a suitably developed neural network architecture with high accuracy. We obtained accuracy and efficiency in terms of standard parameters, and were able to achieve accurate image segmentation with kappa coefficient of 0.97. The time for classification was reduced by more than 70%.