Hierarchical Analysis Method For High Resolution Remote Sensing Images

Due to the constant development of the remote sensing technology, aboundant information of earth surface observation is available for practical applications. Especially, the appearance of high resolution remote sensing (HRRS) image extends visual contents of the observation. However, how to effectively process the HRRS data for convenient use has continued to be a challenging task in recent research. Therefore, the thesis focuses on the theories and methods of HRRS image analysis based on the throughout cue of hierarchical processing.In the research of texture feature extraction for remote sensing images, a novel extension of local-binary-pattern (LBP) texture operator is proposed to overcome the main disadvantage of the original LBP operator which is unable to utilize the information in"non-uniform"patterns. The novel extension classifies and combines the"non-uniform"patterns according to two predefined measures of structure and similarity, and then achieves better discrimination ability and more robustness against noise.In the research of hierarchical classification for HRRS images, accurate classification for HRRS images is still a challenging problem due to inherent ambiguities caused by only using the appearance of the visual data. A cascaded classification algorithm is presented by exploiting multiple hierarchical contexts. Referring to object-oriented image classification, the proposed algorithm consists of three stages which make use of different information including simple local contexts as well as complex global contexts. Since the later stage uses a higher level context than that of the former stage, the proposed algorithm can gradually refine the classification results and finally yields higher classification accuracies of different land-covers.In the research of hierarchical segmentation for HRRS images, two new integration-based algorithms are presented to improve the traditional methods. One is based on the combination of some existing methods. It first produces the initial segmentation by using watershed transform in the way of embedded integration, and then merges and modifies the obtained result by using MRF model in the way of post-processing integration. The other aims for designing a novel integration strategy. The optimal cue is adaptively determined by analyzing the characteristics of the image data, and then detailed partition is implemented by selecting an appropriate segmentation method. The approach takes advantage of different segmentation cues and methods, and provides better segmentation results which are more consistent with human's vision.In the research of hierarchical object detection for HRRS images, a multi-level object description is presented based on Probabilistic Latent Semantic Analysis (PLSA) in order to solve the problem caused by the un-matching relationship between object and its features. This method introduces an additional level of latent aspects so that objects are not simply described by a feature set but the probability combination of latent aspects generated by PLSA. The distributions of latent aspects are much better than the distributions of features for the representation of the essential properties of object. Accordingly, the proposed multi-level object description method is effective to model the diversities of different objects. The satisfied experimental results are obtained while the proposed approach is applied to detection of ship objects.