Research On Hyperspectral Images Classification Based On Spatial-spectral Features And Semisupervised Learning

In recent years,remote sensing technology has been deeply applied to many research fields,and hyperspectral image classification has attracted extensive attention of scientific researchers,which exhibits impressive advantages in remote sensing.In fact,the number of labeled samples has a critical impact on the traditional supervised classification methods,which is not consistent with the practical problems faced by hyperspectral images,because the collection process of labeled samples in hyperspectral images is complex,time-consuming and difficult to achieve.Therefore,how to realize hyperspectral image classification in small sample scenario has become an urgent problem to be solved.To solve this problem,based on the integration of image and spectral characteristics of hyperspectral images,this paper takes strong focus on spatial-spectral information to realize the semisupervised learning of hyperspectral images,which utilize the potential discrimination information of unlabeled samples in hyperspectral images.The research contents of this paper are as follows:1.A re-verification procedure of pseudolabels based on superpixels segmentation algorithm is proposed.The algorithm combines active learning and semisupervised learning,which complements and links each other.Firstly,two different active learning screening strategies are used to select unlabeled samples with large amount of information,which are marked by the artificial expert supervision system,and then expanded to the initial labeled samples set.In this way,three verification classifiers with strong differences are formed to verify the predicted labels of unlabeled samples for the first time.Then,the unlabeled samples passing the initial verification are put into re-verification candidate pool and wait for the reverification procedure based on the spatial consistency of superpixels.After above two pseudolabels verifications,a pseudolabeled sample set will be formed,and the pseudolabeled samples set will be added to the base classifier for the next iterative training.When the iteration stop condition is reached,the base classifier and predicted labels are output.Experimental results show that the algorithm can achieve more efficient hyperspectral image classification in the extreme small sample scenario where the number of initial labeled samples is very scarce.2.A 3D-Gabor filter and multi-graph fusion semisupervised algorithm is proposed.Firstly,the algorithm extracts the features of the original hyperspectral image with the help of a group of 3D-Gabor filters to realize the joint representation of the spatial spectral information of the original hyperspectral image.Then,the multi-view data generated by 3DGabor filter is double filtered to ensure the sufficiency and diversity of view data.The relational graph connection is generated on each view data respectively when view filtering is completed.Next,the label propagation results of each layer are fused by using the multi graph fusion algorithm,so as to alleviate the problem of low reliability in the propagation of singlelayer icon labels.Finally,the prediction category after multiple graph label fusion is weighted and re fused with the prediction category of SVM classifier to produce the final classification result.Experimental results reflect the importance of spatial information extracted by the proposed algorithm,which combines multi-layer label propagation algorithm to realize the effective classification of hyperspectral images in small sample scenarios.