A Method to Improve the Accuracy of Remote Sensing Data Classification by Exploiting the Multi-Scale Properties in the Scene

Land use mapping is one of the major applications of remote sensing. While most studies focus on the advanced remote sensing thematic classification algorithms for land use mapping, the scale factor in remote sensing data classification was less recognized. Previous studies showed that while the multi-scale characteristics exist in the remotely sensed data for land use classification, some classes are mostly accurately classified at a finer resolution, and others at coarser ones. Thus, it is helpful to improve the overall classification accuracy by mapping different land use classes at different scales. In this paper, a framework for improving the land use classification accuracy by exploiting the multi-scale properties of remotely sensed data is presented. Firstly, the remotely sensed data at original fine resolution was up-scaled to different coarser resolutions; Secondly, the up-scaled data were classified by independently trained Maximum Likelihood Classifier at every resolution, and the corresponding a Posteriori Probability of MLC classification was saved; Thirdly, the classification results at different resolutions were integrated by comparing the a Posteriori Probability of classification at every resolution. The final class of pixel was labeled as the class that has the maximum a Posteriori Probability. A case study on the land use mapping using Landsat TM data using this framework was conducted in the Dianchi Watershed in Yunnan Province of China. The land use was categorized into 6 classes. The classification accuracy was assessed using the Confusion Matrix. Comparison between the classification accuracy at multi-scale and that at original resolution showed an improvement of overall classification accuracy by about 10%. The study showed that by exploiting the multi-scale properties in the remotely sensed data, the accuracy the land use mapping can be improved significantly.