Accuracy of a supervised classification of the artificial objects in thermal hyperspectral images

Thermal hyperspectral imagery introduces new possibilities in remote sensing. This paper deals with testing the accuracy of the supervised classification of the artificial objects in the thermal hyperspectral imagery, on a part of 79 channels, in three raster sets: (a) five thermal images (8.747, 9.648, 10.482, 11.266, 11.997 /spl mu/m), (b) five thermal images and one infrared reflective (0.807 /spl mu/m), (c) four thermal images (9.648, 10.482, 11.266, 11.997 /spl mu/m), one infrared reflective and two visible images (0.501, 0.587 /spl mu/m). The visible channels were used as a substitute for the real ground truth source and one infrared reflective channel was used for the enhanced discrimination of the vegetation. Global characteristics and principal components of thermal rasters were considered; unsupervised and supervised classifications have been done too. The result was difficulty in determining the training fields which are the best or, at least, good enough for desired accuracy of supervised classification.