A way to improve an architecture of neural network classifier for remote sensing applications

Recent results in neural network research have demonstrated their utility in a variety of application areas. Neural networks are able to achieve a very high performance, and classification accuracy in real world applications such as handwritten character recognition, remote sensing images, vision, robotic. Network performance greatly depends not only on the input/output data, but also on its architecture. Most of neural network applications have been developed using anad hoc approach resulting in poor efficiency and performance. In this paper, a development method of neural network applications is presented, and illustrated with a neural classifier of remote sensing images. It is shown how to create in an iterative way a neural classifier architecture, and how to refine a network organization using performance evaluation criteria.