A Multiple Self-Organizing Map Scheme for Remote Sensing Classification

Abstract

This paper presents a multiple classifier scheme, known as Multiple Self-Organizing Maps (MSOM), for remote sensing classification problems. Based on the Kohonen SOM, multiple maps are fused, in either unsupervised, supervised or hybrid manners, so as to explore discrimination information from the data itself. The MSOM has the capability to extract and represent high-order statistics of high dimensional data from disparate sources in a nonparametric, vector-quantization fashion. The computation cost is linear in relation to the dimensionality and the operation complexity is simple and equivalent to a minimum-distance classifier. Thus, MSOM is very suitable for remote sensing applications under various data and design-sample conditions. We also demonstrate that the MSOM can be used for hyperspectral data clustering and joint spatio-temporal classification.