STATISTICAL PROBLEMS IN THE DISCRIMINATION OF LAND COVER FROM SATELLITE IMAGES - A CASE-STUDY IN LOWLAND BRITAIN

This paper discusses underlying statistical problems in identifying land classes from satellite imagery. Problems in assessing accuracy of automatic classifiers are illustrated using results from analysis of two Landsat Thematic mapper (TM) subscenes collected for the same areas within Cambridgeshire in the summer and autumn of 1984. Problems arising from between-site variability of training sites are discussed. Use of rejection thresholds is also discussed. Recommendations are given for improving choice of training data to attempt to overcome some of these problems.