High-precision and automation are very important objectives of remote sensing information extraction research. This paper takes LANDSAT ETM as remote sensing data source

and uses spatial scaling （scale transformation） mechanism of "whole-local" based on Normalized Difference Water Index （NDWI）. Combining the segmentation and classification at whole and local scales respectively

this study fuses the hierarchical knowledge of water extraction and sets up an iterative algorithm to achieve the gradually approach of the water body’s optimal margin

so that we can extract the high-precise water body information. The experiment of plateau lake information extraction demonstrated that the complex and various information of water body could be extracted accurately

and the confusion between water body and shadow information could also be effectively avoided.