Research On Hyperspectral Image Classification Methods Based On Superpixel Merging

With people's in-depth research in the field of optoelectronics,people's exploration of electromagnetic spectrum is no longer limited to a few bands in visible light and near-infrared.Hyperspectral image(HSI)has gradually entered the scientific field of vision.With the effective integration of spatial structure and spectral information,hyperspectral remote sensing has been widely used in all aspects of the national economy.Although the hundreds of spectral bands carried by hyperspectral images make it possible for accurate ground object recognition,the high-dimensional nature of hyperspectral images and the difficulty of collecting sample labels make the high-precision classification of ground objects face challenges.Therefore,the object-based hyperspectral image classification method ushers in the opportunity and becomes the mainstream.Exploring a new object-based hyperspectral image classification method,while improving the operation efficiency of the algorithm and the classification accuracy under the condition of small samples,will have a far-reaching significance for the interpretation of hyperspectral images.In order to solve the above problems,based on the existing work,this thesis proposes two very effective spectral-spatial classification methods from the perspective of traditional machine learning and deep learning,which are not only suitable for the classification of hyperspectral images,but also obtain satisfactory classification results on high-resolution multispectral images.(1)An effective spectral-spatial HSI classification method MS-KNN is proposed,which uses merged superpixels.For over segmented superpixel mapping,spatially adjacent superpixels are continuously merged by local modularization acting on the weighted superpixel map(superpixels are used as nodes rather than pixels).Then,a simple but effective conversion method is proposed to map superpixels to new samples.Finally,the k-nearest neighbor(KNN)classification algorithm,which is popular in machine learning,is used for classification.(2)A semi supervised spectral-spatial HSI classification method SSG~2CN based on sparse superpixel graph(SSG)is proposed.In the constructed sparse superpixel graph,each vertex represents a superpixel instead of a pixel,which greatly reduces the size of the graph.At the same time,the merged superpixel,local spatial connection and global spectral connection are used to consider the spectral information and spatial structure.Finally,graph convolutional networks(GCNs),which can be applied to the representation and analysis of irregular(or non grid)data,is used for classification.(3)To illustrate the effectiveness of the proposed methods on hyperspectral imagery and multispectral imagery,respectively,we conducted extensive experiments on three publicly available benchmark hyperspectral datasets from different sensors,Indian Pines,Pavia University,and Salinas,as well as the WHU-Hi-Han Chuan dataset and the WHU-Hi-Hong Hu dataset collected and shared by the RSIDEA research group of Wuhan University,and Gaofen image dataset(GID)is selected to verify the accuracy of the proposed classification methods.Both the experimental and comparative results show that the proposed algorithm can not only effectively classify hyperspectral images with limited training labels,but also has practical application significance on high-resolution multispectral images.