The research on the role of several feature extraction methods in the landcover/landuse classification

摘要

A classifier of great capabilities and a good-selection of different features are two key and difficult keys answering for a high accuracy classification result. On the classifier, although there are all kinds of algorithms, most of them couldn't be used widely because of multifarious theoretical limitations. In this paper, based on the TM data, several representative interpretation features, including original bands, texture measurements and spatial metrics, are compared systemically for landcover/landuse classification test with the same classifier and the same training samples. The results show that different feature source has different relationship with the original band and they play the different roles. Summarily, the original bands are the most useful and essential feature source and play the important role and the others can only be seen as equivalent or enhanced feature source. Among which, the texture mean have equivalent capability as that of the original bands, and the spatial metrics and other texture measurements can be seen as compensatory source. For the combination of different features, the classification accuracy can be improved by using the texture measurements or the combination with original bands. As a sort of newly features, the classification accuracy was very poor if only landscape metrics were used, comparatively the accuracy can be greatly improved by combing with the original bands. So, the combination of original bands and texture measurements is the preference for TM dataset.