An evaluation of automated classification techniques utilizing landsat data for soils mapping in the Grand Rapids area, Manitoba

Abstract

The use of automated classification techniques utilizing LANDSAT data was evaluated for reconnaissance level soils mapping in the Boreal Forest Region near Grand Rapids, Manitoba. Supervised and unsupervised automated methods were employed rigourously to determine their potential with a single image LANDST data base. Both methods yielded similar results, as 8 to 10 soil mapping groups were successfully recognized. A tradeoff existed between the number of classes delineated and classification performance. Traditional methods resulted in the delineation of 57 map unit components. Automated methods provided several advantages, such as savings in expensive ground truth collection, uniformity of classification results, and ease of manipulation of the final map product. Automated methods of soil mapping in a Boreal Forest environment were dependent upon the indirest detection of soil conditions through vegetative cover. The poor coorelation between native vegetative cover, the major contributer to the spectral signature, and the underlying soil conditions was the main limitation in classification performance. Additonal limiting factors were the LANDSAT data base and the restriction of current automated classifiers to pixel-by-pixel analysis. Automated methods also do not provide map units comparable to those produced by conventional techniques. Several methods of enhancing the legibility of automated map products were suggested...