Forest classification by principal component analyses of TM data

Abstract Differential insolation generally strongly influences the spectral responses of woody areas distributed over hills and mountains, making them difficult to discriminate by means of remote sensing techniques. In this paper, we examine the possibilities of a new method of variance decomposition of Thematic Mapper (TM) data using principal component analyses applied to this problem. The first principal components found for every class of ground truth are assumed as expressing the portions of spectral variance related to different insolation and are therefore not considered in discriminating the forested surfaces of two TM scenes by a minimum distance classifier. The low quality and quantity of the ground information used in the research have forced us to employ them with some restrictions. This leads to overestimation of the classification accuracy evaluated by a final confusion matrix; for this reason, the obtained results cannot be considered definitive.