Research On Semi-supervised Classification Method Of Hyperspectral Image Based On Superpixel

There are generally have more than 100 spectral bands in hyperspectral image that can provide detailed spectral information,in which,every pixel can be treated as a vector with higher dimension.Although spectral information of hundreds of dimensions can allow for precision landscape classification,the imbalance between high dimensions and the number of training samples also restricts the development of hyperspectral image classification methods.How to improve the classification of high-dimensional and small-sample remote sensing images and how to solve the accurate classification with phenomenon of homologue spectrum and hetero-spectral homologue are still the difficulties in hyperspectral image processing.In HSI thesis,the relationship between spatial context information and neighborhood pixels in hyperspectral remote sensing images is deeply analyzed,and two high performance classification techniques based on superpixel are proposed and the advantages of the proposed method are verified by several datasets.(1)Due to the influence of weather,sensor,shooting environment and other factors,there is a serious proximity effect between pixels in hyperspectral images.The classification performance based on the mixed radiance information is bad.However,superpixel can eliminate the partial proximity effect to some extent and smooth the difference between adjacent pixels in the homogeneous region.Therefore,a new extended random walk classification method based on superpixel is proposed in HSI thesis.By improving the latest extended random walk classification algorithm,the graph model is constructed with superpixel as node and the weight matrix between superpixel is also computed.Then the energy function containing the prior probability distribution of the superpixel is introduced to obtain the final probability distribution of each superpixel.Experiments show that the proposed method goes beyond the original pixel-by-pixel extended random walk method,especially with small samples,which can widen the accuracy of other classification methods.(2)In order to solve the problem of insufficient samples in hyperspectral image classification,some semi-supervised methods can generate pseudo-labeled samples for classification.However,error labeled samples used for model training will seriously affect the classification performance.In view of the fact that the homogeneity of superpixel ensures the confidence of pseudo-labeled samples set,a hyperspectral image classification technique based on superpixel and feature extraction is proposed to improve the traditional label propagation algorithm.Firstly,a two-step extended label propagation method is designed to interpret the data distribution of hyperspectral images,where the first step is neighborhood propagation,and the second step is superpixel propagation,which propagates the label within the superpixel.The strong edge feature extracted by rolling guidance filter is used as the input of support vector machine,and the pseudo-labeled samples set is added to the training set to obtain the final classification result.Experimental results demonstrate that HSI method solves the problems of traditional label propagation algorithm,such as propagation range,propagation efficiency and the confidence of propagation label,and obtains better classification results.The results of classification technique based on superpixel is presented in HSI thesis show that compared with a single pixel,the spectrum of superpixel is more stable and less affected by noise.Superpixel can effectively solve the problems of homologue spectrum and few training samples,so it has more advantages in dealing with limited sample classification.Finally,four experimental data sets are used to verify the performance of the two semi-supervised classification methods proposed in HSI thesis.Experimental results proved that the hyperspectral image classification method with superpixel as the minimum classification unit is reasonable.