Community Research and Development Information Service - CORDIS

Abstract

There has been a rapid growth in the use of artificial neural networks in satellite remote sensing over the last five years. They are now used extensively for image classification in thematic mapping, though their comparative advantage to statistical classifers is difficult to assess. Moreover, neural networks require careful regularization of input data to ensure good learning performance, and to avoid problems such as over fitting or falling into chaotic regimes due to their nonlinearity. Recent experience has shown, however, that progress in the improvement of fixed category or hard classification is slow suggesting that soft classification approaches are needed in Earth observation to deal with the continuum nature of terrestrial land cover. Recent work has focussed on exploiting neural network algorithms in the soft classification context.
Apart from classification problems, neural network algorithms have potential as general data transformation systems which can be exploited in a number of ways in remote sensing (eg in signal inversion and geometrical operations for image geocoding or stereo matching). A general reluctance to use neural networks more widely in Earth observation is attributed to lack of user oriented systems, and to lack of fast training approaches for many network architectures. Recently developed neural network hardware systems are just beginning to be applied in remote sensing and may hold considerable potential especially if combined with appropriately engineered software systems for geographical users.

Additional information

Authors: WILKINSON G G, JRC Ispra (IT)
Bibliographic Reference: Article: Recent Developments in Spatial Analysis, Springer-Verlag (1996)
Record Number: 199610876 / Last updated on: 1996-09-16
Category: PUBLICATION
Original language: en
Available languages: en