Research For Classification Of High Spatial Resolution Remotely Sensed Imagery

Compared with medium and low resolution remote sensing images, those ones with high spatial resolution carry more and more abundant details, providing a better presentation of textures, shapes and geometry structures of spatial entities. Meanwhile, coupled with gradually improved spatial resolution, the increasing data size rises a need for automatic classification techniques with high efficiency to extract information from those images since traditional methods are not feasible. However, high resolution images with relatively low spectral resolution and rich details often possess complex spectral distributions, which results in the difficulties of classification based on spectral domain. In this situation, the intrinsic potential information in high resolution images should be fully considered together with the spectral features to solve this issue.This paper firstly takes advantages of support vector machine in handling problems of small samples, non-linearity and high-dimension pattern recognition to perform classification of high resolution images based on the spectral domain. Results indicate that spatial entities in images cannot be totally distinguished merely depend on their spectral differences since those with similar spectral features would be mistaken for each other.Then this paper seeks to extract pixel shape index(PSI) as a supplementary for the spectral features to carry out classifications. Results show that compared with spectral-based classification, psi could effectively separate spatial entities with similar spectral features but different geometry shapes. Moreover, PSI method also achieves better performance than classifications based on wavelet texture features or multi-scale regional features. However, PSI method is vulnerable to the disturbances of details in high resolution images, making it undesirable when classifying regions with rich details.Considering above situations, object-oriented analysis is introduced into the PSI method and a bandwidth self-adaptive mean shift method is proposed to extract spatial objects in high resolution remote sensing images. In the experiments, with small targets remained and excessive details ignored, large targets are smoothed enough, which indicates that the method is multi-scale unification. The final classification proves that compared with PSI method, the proposed object-oriented method achieves better performance both in visual effects and precision.