New supervised learning of neural networks for satellite image classification
Akito Ohkubo, Koichi Niijima · 2003
In this paper, we propose a new learning method of three-layered neural networks based on the concept of domains of recognition in the input space. This network is learnt by minimizing a cost function which is derived from geometric properties of the domain. Our learning process is to enlarge a domain of the hidden space associated with the domain of recognition in the input space. This theory is applied to a land cover classification problem for the satellite image data. The proposed method can classify input data into some categories which are mutually disjoint. In simulations, we process not only multi-band data observed by an optical sensor but also by a microwave radar.