Earth's surface space is a complex huge system and character with hierarchical structures. Entities, patterns and processes
all show inherent hierarchy structure in nature. The character of Scale-dependence is corresponded with hierarchy. Many
research works have demonstrated that scale-dependence is a basic characteristic of Geo-spatial space. Therefore, the
multi-scale or hierarchical approach needs to be introduced in the course of spatial information analysis, monitoring,
modeling and management. It is well know that image analyze result was influenced by the window size that was
selected. The original fixed window cannot suit with the object spatial character. In this letter, we first propose an
optimal window selection method, based on the spectral information in a local block region, for choosing the suitable
window size adaptively. Secondly, the object spatial information is learned based on the selected optimal window size.
Thirdly, both the spectral and spatial information were used in image classification. In this paper, the proposed algorithm
can obtain the multi-scale features effectively and the features we get at different scale level have an obvious stability
with property. In the experiment on the QuickBird image data, the proposed algorithm clearly improves the classification
accuracies than fixed window sizes and reduces the salt and pepper effect and error. It is suitable to form multi-scale
hierarchy image-sets and select the objects at different scale levels.