Hyperspectral And LiDAR Image Feature Extraction And Fusion Based On Mathematical Morphology

Hyperspectral image,as one kind of passive sensor data,which contains rich spectral information,can effectively identify the object according to the reflectivity of objects.With the improvement of spatial resolution of hyperspectral images,spatial information plays an increasingly important role in hyperspectral images classification.Therefor,it has become a research trend to integrate spectral-spatial information together for cooperative classification.LiDAR image,as one kind of active sensor data,which contains of abundant elevation and structure information,can effectively distinguish objects according to their shape and height.Hyperspectral and LiDAR images capture different information of objects,which contains great differences and compleme ntarities.Collaborative classification based on these two sources can achieve better recognition effect in some complex areas,such as urban areas.Therefore,in order to make full use of the information involved in these two sources,a morphlogy method is propose by extracting spatial information from hyperspectral images and elevation information from LiDAR image in this thesis.Then,different features are fused effectively and fed into Composite Kernel Support Vector Machine for classification.Main content of the thesis is divied into three parts as follows:Firstly,the existing mathematical morphology methods used in image processing,such as Extinction Profile and Attribute Profile,are vulnerable to external factors such as cloud occlusion and noise points.A new mathematical morphology method,Local Contain Profile(LCP),is proposed,LCP is based on the topology tree,taking advantage of relationships like containment,adjacency and separation between the shapes in the image,therefore LCP is insensitive to external factors.Moreover,due to the attributes(such as area,height,standard variance,boundary diagonal,etc.)in mathematical morphology are insufficient for remote sensing feature extraction,several new attributes are introduced,namely,compactness,elongation and sharpness.These new attributes can well assist the mathematical morphology technique to effectively extract the specific objects,providing supplementation for traditional attributes.Secondly,the existing feature extraction methods are difficult to extract the rich spatial information from hyperspectral images effectively,LCP is introduced into hyperspectral image processing.Experimental results indicate that LCP possesses superior capabilities for extracting spatial features in hyperspectral images.Thirdly,aiming at disposing the problem of single remote sensing data source has limitations in the classification task of complex scenes,a mathematical morphology method is proposed to achieve the collaborative classification of hyperspectral and LiDAR data.Based on the differences and complementarities of the information contained in the two types data,appropriate mathematical morphology methods and attributes are chosen to obtain the feature information from multi-source remote sensing data,and the extracted features are fused effectively to obtain the optimal classification performance of objects.