The Research On Image Preprocessing And Segmentation Of Hyperspectral Remote Sensing

Hyperspectral image segmentation is regarded as an important task in hyperspectral image analysis and understanding. Accurate image segmentation conducives subsequent classification, target detection, objective extraction and so on. Due to the high-dimensionality of hyperspectral image, the conventional segmentation techniques can not be used. The study of segmentation for hyperspectral image has practical meaning. Because of the huge amount of data, the high redundancy between different bands and the inefficiency of direct processing, the common strategies for conventional remote sensing could not be easily applied to hyperspectral imagery. Therefor, the dimensional reduction of hyperspectral image is essential for the following image analysis. The strip noise can distractingly and obstructively affects the interpretation and application of hyperspectral image. After the processing of feature extraction or dimensional reduction, the strip noise becomes more prominent, seriously affecting the subsequent hyperspectral image analysis and interpretation. It is significant and necessary to eliminate strip noise before dimensional reduction. In this article, the removal of strips and dimensional reduction are researched as the preprocessing for hyperspectral image. Then, the segmentation technique for hyperspectral image is studied.The standard moment matching method can effectively remove the strips in hyperspectral image, but there exist the "ribbon" effect in non-flat area. In this paper, a destriping method based on improved S-G filter moment matching is proposed. Several commonly used strip noise removal methods are tested and compared. Simulation results show that the proposed method has a better effect on striping removal, and the original information of image is better preserved.It can be observed for hyperspectral image in which the correlations between neighboring spectral bands are generally higher than bands further apart, with high correlations appearing in blocks. Initially, the entire bands are partitioned into several highly correlated subgroups. After the division of hyperspectral bands, the image fusion technique is applied to compress the spectral data. Taking advantages of the image fusion technique for reducing the dimension of hyperspectral data, not only achieves the goal of lower redundancy, but also fuses the complementary information from different bands, which is help for the subsequent segmentation. Applying the method of local energy maximization for high frequency decomposition image and the weighted average for low frequencey decomposition image, it successfully avoids the blurring problem caused by direct weighted fusion, and obtains smaller correlation between the reduced images.In this paper, a new multilevel thresholding method is introduced for the segmentaiotn of hyperspectral images. The new method is based on fractional-order Darwinian particle swarm optimization(FODPSO). After the parket dimensional reduction for hyperspectral data, the FODPSO is used for searching the thresholds that maximize the between-class variation to achieve multi-level image segmentation for each reduced band image. Then, the segmentation maps of different subgroups are merged, resulting in the overall segmentation map of hyperspectral image, which exhibits many over-segmentation regions appearing as blobs in the imagery. In the paper, a new region merging strategy is adopted to improve the over-segmentation phenomenon of initial segmentation results. According to the steps described previously, the segmentation for hyperspectral image is realized with the efficiency improved to an extent.