Spatial Pattern Recognition in Remote Sensing by Neural Networks

Abstract

Neural networks are sophisticated pattern recognition tools. Over the last decade, following the early pioneering work of Benediktsson, Swain and Ersoy (1990) and Hepner et al. (1990), they have grown significantly in popularity as tools for the analysis of remote sensing data, primarily for the purpose of deriving spatial information used in environmental management. This popularity is due to several factors:

• Remote sensing applications are data rich [they rely increasingly on high dimensional imagery].

• Remote sensing data arise from complex physical and radiometric processes.

• There has been dissatisfaction with more conventional pattern recognition algorithms.

• There is a general desire to get maximum accuracy out of the data in remote sensing; high accuracy in some applications [e.g. agricultural subsidy fraud monitoring, mineral resource location] has high economic value.

• Neural networks are suitable for data for which statistical properties are unknown or poorly understood.

• Neural networks have potential scalability.

• Neural networks are potentially adaptable to parallel machine architectures.