Research On Hyperspectral Image Classification Method Based On Multiple Kernel Learning

As the core method of remote sensing image analysis,hyperspectral classification has always received great attention and has been widely used in related fields of the national economy,such as natural resource surveys,urban land use planning,fine agriculture and forestry,and environmental protection.Therefore,it is very important to design a suitable and stable method for hyperspectral classification.In the framework of multiple kernel learning method,this paper takes fully mining the inherent structural information and discriminative information contained in the sample as the starting point,and introduces theories and methods such as local binary pattern(LBP),superpixels,and deep learning to solve problems that the edge of HSI and inaccurate pixel extraction and serious time-consuming pre-training phase.The main research work is as follows:(1)Aiming at the problem of inaccurate extraction of edge pixels of HSI,a multiple kernel learning classification method based on LBP and superpixels is proposed.First,the original hyperspectral image is divided into multiple superpixels through the entropy superpixel segmentation algorithm.Secondly,on the hyperspectral image with superpixel index,LBP and weighted average filtering are used to obtain the inter-superpixel,intra-superpixel and spectral features of the image,and use these features to generate the corresponding Gaussian kernel.Finally,the composite kernel method is used to select the base kernel weights corresponding to different features to determine the contribution of different features to the final classification performance of a given class,so as to achieve accurate classification of the class.(2)In response to the time-consuming and insufficient labeled samples in the deep learning-based hyperspectral image classification method,we combined the hierarchical deep convolutional neural network with LBP,and proposed a multiple kernel classification model based on LBP and random patches.The model uses the spectral features,local texture features and multi-layer convolution features to complete the HSI classification.Compared with traditional CNN classification methods,this method can improve the classification performance while reducing the time required to train the CNN model and the number of labeled samples.