Objective
This project is a more focussed extension of the previous exploratory research project on Neural Networks for Remote Sensing Images which covered the application of neural networks to remote sensing in general. The purpose of this proposal is to develop to a very high level the possibility of operationally using neural networks in mapping land cover from space over large regions (eg approximately 10 000 square kilometres) and to make optimal use of fused multi-source data eg, from optical/infra-red satellite sensors and mircowave systems.
Multilayer preceptron networks have been developed to classify 2 date multispectral Systeme Probatoire d'Observation de la Terre (SPOT) imagery over 2 test areas of approximately 100 square km in France. The accuracy achieved has been of the order of 85% for 20 land cover classes, exceeding the accuracy obtained in tests with conventional maximum likelihood classifiers by about 30%. Quite large networks have been successfully developed for this purpose with about 100 neurons and about 2000 interconnections. Initial problems with training such networks have been successfully overcome.
The work will be concerned with the development of an operational system for routine use of neural network classifers in land cover mapping. The main tasks which are planned are as follows:
Development of the existing experimental neural network software into an operationally-usable package with a graphical interactive interface.
Development of networks to make use of merged satellite datasets combining Landsat Thematic Mapper optical/infra-red images and new ERS-1 Synthetic Aperture Radar images over various European test areas.
Experimentation with video techniques as a means of recording the time-development of network learning in image classification.
Gathering of ground data over large geographic areas for training networks on representative land cover types including man-made features, natural vegetation types, various types of forests and agricultural crops for large area mapping.
Further experimentation to find network training techniques with optimal convergence properties for inclusion in the operational package.
Comparison of the backpropagation method with the Learning Vector Qauntisation (LVQ) method.
The funds requested are primarily intended to cover satellite data (Landsat and ERS-1), collection of field data for training and performance evaluation of networks (mainly by contract), miscellaneous expenses concerning the video application and small-scale computer items. No major hardware expenditure is foreseen on account of investments made in previous years.
It is intended that the results of this final development stage of the neural network technique will lead to its routine use within the application projects of the Institute for Remote Sensing Applications. These could include: the CORINE land cover mapping project involving support to the new European Environmental Agency, the TREES project, and the MARS (Agriculture monitoring) project.
Topic(s)
Data not availableCall for proposal
Data not availableFunding Scheme
Data not availableCoordinator
21020 Ispra
Italy