摘要：The expert knowledge has been widely used to improve the remotely sensed classification accuracy. Generally, the expert classification system mainly depends on DEM and some thematic maps. The spatial relationship information in pixel level was commonly introduced into the expert classification. Because the geographic objects were found spatially dependent relationship to a certain degree, the commonly used basic unit of spatial relationship information in pixel greatly limited the efficiency of spatial information. A patch-based neighborhood searching algorithm was proposed to implement the expert classification. The homogeneous spectral unit, patch, was used as the basic unit in the spatial object granularity, and different types of patches’ relationship information were obtained through a spatial neighborhood searching algorithm. And then the neighborhood information and DEM data were added into the expert classification system and used to modify the primitive classification errors. In this case, the classification accuracies of wetland, grassland and cropland were obviously improved. In this work, water was used as base object, and different types of water extraction methods were tested to get a result in a high accuracy.
