A sampling strategy for a single step land cover change classification

摘要: In this study, we proposed a sampling strategy for single step land cover change detection method. The strategy facilitates the derivation of samples detailed "from-to" and no-change classes from images of multiple dates. It consists two steps. Firstly, interest will be defined their training derived separately date data sets. Secondly, sets class or signatures combined in pair artificially as one set both classes. As result, full list possible changes and no-changes are effectively trained. is simple able to eliminate those impossible directions considered by expert knowledge. Our case study on Drayton Coal Mine surrounding area demonstrated that the sampling when used together with single-step classification method yielded much meaningful cleaner land map than traditional two-step post-classification addition, one-step classification also provided higher overall testing accuracy (e.g., 82.3% vs 78.8%). On other hand, resultant more fragment, area of clearly over-estimated close 50%). One disturbing fact two-step post-classification generated large proportion not existent area. This problem can overcome developed strategy.