Feature Evolution for Classification of Remotely Sensed Data

In a number of remote-sensing applications, it is critical to decrease the dimensionality of the input in order to reduce the complexity and, hence, the processing time and possibly improve classification accuracy. In this letter, the application of genetic algorithms as a means of feature selection is explored. A genetic algorithm is used to select a near-optimal subset of input dimensions using a feed-forward multilayer perceptron trained by backpropagation as the classifier. Feature and topology evolution are performed simultaneously based on actual classification results (wrapper approach).