Multidate SAR/TM synergism for crop classification in western Canada

Abstract

Multidate synthetic aperture radar (SAR) and Thematic Mapper (TM) visible and near-infrared (VNIR) data were evaluated for classifying crops frequently grown in western Canada. The VNIR data were superior to the SAR data for single date classifications due to the multispectral information content. Multidate classifications with SAR data improved classification accuracy from 30 to 74 percent although multidate VNIR produced the highest single sensor result of 90 percent correct classification. This was slightly improved to 92 percent by including the SAR data with the VNIR data. However, transformed divergence statistics show that the SAR and VNIR channels are both found in the top eight channels, and, indeed, the best two SAR channels and the best two VNIR channels, based on their transformed divergence statistics, produced an overall classification accuracy of 85 percent. Furthermore, the May TM data combined with the SAR data yielded an 87 percent correct classification because the grain and alfalfa classes were much better separated when VNIR data was combined with SAR data. These results demonstrate significant synergism between the two sensors and suggest the need for a feature selection approach, or at least a knowledge based system incorporating the synergism effect, once multidate, multisensor data become available on a regular basis. The substitution of SAR data beneath cloud covered terrain in TM data is used to demonstrate another aspect of SAR/VNIR synergism.