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

Date Published: 26 July 2007
PDF: 10 pages
Proc. SPIE 6752, Geoinformatics 2007: Remotely Sensed Data and Information, 67523H (26 July 2007); doi: 10.1117/12.761233

Published in SPIE Proceedings Vol. 6752:
Geoinformatics 2007: Remotely Sensed Data and Information
Weimin Ju; Shuhe Zhao, Editor(s)