Integration of self-organizing maps with spatial indexing for efficient processing of multi-dimensional data

Abstract

This paper investigates the integration of a class of adaptive soft-computing techniques and architectures with helical hyperspatial codes (HHCode) - indexing technology developed at Canadian Hydrographic Services - and their use in developing automated systems for processing of complex, multi-dimensional geo-spatial information, mainly multi-spectral satellite imagery, in a broader context of knowledge extraction and representation. The soft-computing methods investigated here involve fusion of techniques used in self-organizing maps (SOM - a class of unsupervised neural networks) and fuzzy logic. The topological relationships between the features - automatically extracted by SOM from multi-spectral images - are formed into a neural network in a meaningful order. The ordered features can later be interpreted and labeled according to the specific requirements of the application. Two SOM/HHCode integration architectures are proposed and discussed in the paper: closely-coupled integration where HHCode queries control the size of the neural network - in other words, the generalization level; and an architecture where HHCode is used to encode the results of clustering by SOM as well as the topological ordering of heterogeneous data encoded in the network. Results of tests performed on multi-spectral 20-m SPOT satellite images are given.