| A new classification approach, referred to as supervised iterative classification, is described. TM and SPOT XS data are used to compare the performance of the iterative classification approach versus traditional single stage supervised, and unsupervised classifications. All the classifications are performed by a maximum likelihood classifier, with a first pass classification with a parallelepiped classifier to speed up the computation. Original information employed for production of the Forest Map of Spain is used as ground truth, in the training stage for the supervised classifications, and in the testing stage for all the classifications. Five types of classifications for both TM, and SPOT data are considered. The optimum number of bands for classification is determined by comparison of the overall classification accuracy obtained with sets of 2, 3, 4, 5, and 6 bands. Comparison of classification approaches (two at a time) for each type of data is performed by means of a Z-test. Two types of separability indices, Transformed Divergence (TD), and Jeffry-Matthusita distance (JM), and their weighted versions, Weighted Transformed Divergence (WTD), and Weighted Jeffry-Matthusita (WJM), are used in the band selection process. GIS analysis of the study area is performed in order to determine its potential for improving classification accuracy by means of image segmentation, and/or prior probabilities. Results show that, in the present study, GIS does not permit image segmentation based on information different than spectral. It is also shown that prior probabilities improve the band selection process for TM data, only when using the TD separability index. The use of prior probabilities does not improve the band selection process for SPOT data. The highest classification accuracy is obtained when using 6 bands for the single stage classifications, and 5 or 6 bands for the different iterations in the iterative classification approach. Comparison of the error matrices obtained from all the classifications shows that the iterative classification is superior to any other classification approach for both TM and SPOT XS data. No differences are found in the performance of supervised classifications when using spectral analysis or spectral separability indices for band selection. Supervised classifications are shown to be superior to unsupervised approaches for both TM, and SPOT data. Comparison of TM versus SPOT XS data shows that TM data proves to be superior to SPOT data only when the iterative classification approach is used. (Abstract shortened by UMI.)... |
