Research On Semi-supervised Classification Method Of Hyperspectral Image Based On Sample Pseudo-Labeling

Hyperspectral images usually contain hundreds of spectral bands,and provide a wealth of spectral information that can reflect different land covers.In the field of hyperspectral image processing,hyperspectral image classification is an important research topic.However,in the hyperspectral image,the imbalance between the limited training samples and the extreme spectral dimensions restricts the development of hyperspectral image classification methods.Since labeling sufficient samples of hyperspectral images to achieve adequate training is quite expensive and difficult,the problem of high-dimensional small sample classification is still the difficulty of hyperspectral image processing.In response to this problem,the pseudo-labeled samples are formed by assigning a pseudo-label to unlabeled samples in this paper,two different hyperspectral image classification methods based on sample pseudo-labeling are proposed,and the proposed method is verified using multiple hyperspectral data sets.(1)A hyperspectral image classification method based on superpixel and multi-classifier fusion is proposed.This method guarantees the confidence of pseudo-labeled samples based on the homogeneity of superpixels,and multi-classifier fusion is also proposed to improve classification accuracy according to the spatial distribution characteristics of land covers.The proposed method includes the following three steps.First,superpixels are used to generate pseudo-labeled samples to increase the number of training samples.Second,label propagation is used to classify the hyperspectral images,but for certain land covers with dispersed spatial distributions,the corresponding classification performance is poor.Thus,a support vector machine classifier is introduced to classify the hyperspectral images.Finally,the results of the support vector machine and label propagation classifiers are combined using our new weighted fusion algorithm.Three widely used real hyperspectral data sets were selected for evaluation.The experimental results show that the proposed method is superior to other common classification methods in the case of small samples.(2)A sparse representation-based sample pseudo-labeling method is proposed.This method draws on the idea of sparse representation and information entropy to select the purest samples to extend the training set.The proposed method consists of the following three steps.First,intrinsic image decomposition is used to obtain the spectral reflectance component of hyperspectral images.Second,hyperspectral pixels are sparsely represented using an overcomplete dictionary composed of all training samples.Then,information entropy is defined for the vectorized sparse representation,and the pixels with low information entropy are selected as pseudo-labeled samples to augment the training set.Finally,the quality of the generated pseudo-labeled samples is evaluated based on classification accuracy.Experimental results on real hyperspectral data sets show that the proposed method has excellent classification performance compared with other classification methods,which indicates that the pseudo-labeled samples generated have high confidence.