Research On 3D Spatial-Spectral Feature Extraction And Small Sample Size-based Classification Of Hyperspectral Remote Sensing

Hyperspectral remote sensing images contain rich spatial,radiological and spectral information on surface matter,and have become increasingly prominent in remote sensing information processing.They have been widely used in many fields such as urban classification,environmental monitoring and military target identification.However,traditional remote sensing information processing technology faces many problems and challenges in interpreting hyperspectral remote sensing images: how to effectively use the multiple information provided by hyperspectral images,and how to solve the bottlenecks of the performance of small-sample based classification.Based on the characteristics of hyperspectral remote sensing image,this paper studied how to obtain discriminative three-dimensional spectral-spatial feature of hyperspectral remote sensing image and further integrated various spectral-spatial features to achieve the effective classification of land cover on the basis of summarizing the existing remote sensing information processing technology.Most of the existing spatial-spectral feature extraction methods only mechanically connected the spatial features of different bands,and did not make full use of the context information in the spatial-spectral structure of hyperspectral remote sensing image.Aiming at this problem,this paper presented a spatial-spectral feature extraction method of hyperspectral remote sensing image based on three-dimensional local binary pattern(LBP).In this method,the three-dimensional LBP coding feature is grouped by using the statistical method,and the dense coding pattern for 3D topology is established instead of the traditional weighted coding.The three-dimensional LBP is invariant to rotation,which enhances the discrimination power of feature.In addition,the robustness of the three-dimensional LBP coding feature was improved by introducing the slack variable to realize the fuzzy processing of the threshold operation,eliminating the noise and the influence of the unevenness of the reflection value distribution.The experimental results showed that the spectral-spatial feature extraction method of hyperspectral remote sensing image based on three-dimensional LBP can effectively use the context semantics in spectral-spatial structure of hyperspectral image to significantly improve the performance of small-sample based classification.Different three-dimensional spatial-spectral features have different advantages.The threedimensional Gabor feature is well robust to the illumination and shadow in hyperspectral image.The three-dimensional morphological profile can effectively capture the shape of land cover in hyperspectral image,and the three-dimensional local binary pattern well characterizes the spatial feature in hyperspectral image.Since the three features represent different attributes of hyperspectral images,their discrimination power to different land cover type differs and complement each other.Based on this,this paper proposed a three-dimensional spatial-spectral feature classification method for hyperspectral remote sensing images based on sparse decision fusion.This method made a decision fusion of above three three-dimensional spatial-spectral features with different characteristics by sparse representation,and integrated the advantages of the three features to achieve complementary discrimination power.Experiments showed that the sparse decision fusion based three-dimensional spatial-spectral feature classification method effectively combined the advantages of the three kinds of three-dimensional spatialspectral features,and further improved the performance of classification with small sample.