Improved Active Deep Learning For Semi-supervised Classification Of Hyperspectral Image

The classification task of hyperspectral remote sensing image has been rapidly developed in the field of remote sensing research by taking advantage of the remote sensing data widely.However,in practical classification tasks,the distribution of different features is complex in remote sensing images,while requiring experts to spend a lot of manpower and material resources to label unlabeled samples.Then,It leads to the problem of the scarcity of the labeled samples with obtaining a large number of labeled samples difficultly.Therefore,it is not only necessary to consider the use of the labeled samples rarely and a large number of unlabeled samples,but also to consider solving the problem of the high cost of manual labeling for unlabeled samples in order to improve the classification performance of hyperspectral images with small samples.There is no doubt that the aforementioned issue is of great significance for the semi-supervised classification task of different features.Based on the theoretical research of hyperspectral feature extraction,the paper takes the semi-supervised classification method of active deep learning as the main line,and establishes an improved semi-supervised classification model of active deep learning.Aiming at the problem of expensive labeling costs when experts label unlabeled samples in active deep learning models,a semi-supervised classification model of dual-space probabilistic fusion with improved active deep learning is proposed.Firstly,the dual-space probabilistic fusion method is improved on the basis of the original Recursive Filtering algorithm,which utilizes the fusion of different aspects of spatial information.Then,the improved algorithm is used to simulate the action of human experts marking samples in the active learning process for achieving more accurate classification of unlabeled samples.Experimental results show that the improved active deep learning model can improve the classification accuracy of different features effectively.Aiming at the problem that the dual-space probability fusion algorithm can not label unlabeled samples more accurately with fewer labeled samples,an improved active deep learning semi-supervised classification model is proposed,which is the combination of the anchor graph construction.Compared with the double-space probability fusion algorithm,this method introducing random multi-graph algorithm considers the relationship between all samples in the process of predicting labels,and uses anchor graph learning to automatically label unlabeled samples.Then the labeled samples are continuously added to Convolutional Neural Networks for training,and the final semi-supervised classification model is obtained.Finally,the proposed model is applied to hyperspectral datasets in four different scenarios,and the experimental results prove the accuracy and generalization of the proposed classification model.