A Per-Pixel Stratified Classification Methodology for Land Cover Mapping Based on Medium-Resolution Satellite Imagery

Classification is one of the most important procedures in the chain of extracting Land Use/Land Cover (LU/LC) information from remote sensing imagery. How to improve the classification accuracy is the key problem that has long bedeviled the researchers. Therefore, many new classification methods and technologies have been developed, such as the Artificial Neural Network classification, the Fuzzy classification, the Knowledge-Based classification, the Support Vector Machine classification and so on. This paper proposed a simple and flexible methodology for land cover mapping based on the knowledge rules for per-pixel judge who referred to many indexes, such as NDVI, NDBI, and some typical spectral characteristics of the land-objects. The Principal Components Analysis (PCA) was also employed to distinguish the city resident and the greenhouse of the agricultural land whose spectral signature was very approximate. In this study we selected the Xiangfan City in Hubei province as the study area and the multispectral bands of ETM+ data on Sep.2nd, 2002 as the study data. We also collected a higher resolution IKONOS image (2002) as the reference data for the visual interpretation and accuracy validate. During the study, we also set the supervised classification which based on maximum like hood classifier as the contrast test. Results from the study shows that the stratified Classification whose rules based on knowledge could give a accuracy as high as 78.2%, compared with the 70.5% of supervised classification. Because of the open framework in this classification methodology, other classification principle could be easily integrated with in order to improve the precision, such as the Fuzz Logic, GIS knowledge, etc.