Research On Hyperspectral Image Classification Method Based On Deep Transfer Learning

In recent years,with the rapid development of remote sensing technology,the classification task of hyperspectral images has aroused widespread concern in society.In order to solve the classification problem of hyperspectral images,a large number of labeled samples are required.However,in practice,there are very few labeled samples or no labeled sample in some hyperspectral image scenes,and manual labeling of these samples has limitations.In this case,the classification task of hyperspectral images will become very difficult.But at the same time there are more labeled samples in similar scenes,therefore,hyperspectral scenes(source domains)with enough labeled samples can be used to help other hyperspectral scenes(target domains)with obviously insufficient labeled samples for classification.However,in two different hyperspectral scenes,due to the different atmospheric and lighting conditions at the time of image acquisition,and the different composition of the same feature category in different time and space,spectral shifts often occur between scenes.The existence of the spectral shift brings serious data distribution differences between domains,resulting in the classification model trained with labeled samples in the source domain cannot be adapted to the target domain.(1)In view of the problem of differences in the distribution of hyperspectral image data in different scenes,this paper studies from the perspective of supervised learning and proposes a deep domain adaptation hyperspectral image classification method based on weight adaptation.First,the method makes full use of a large number of labeled samples in the source domain and a small number of labeled samples in the target domain,and extracts more class-discriminatory features in the target domain through the loss of similarity between domains.Secondly,the method obtains the invariant features between domains by adaptively adjusting the weight of the similarity loss between domains and the cross-entropy loss of the source domain.Finally,the method fine-tuning the network by using a small number of labeled samples in the target domain.(2)In view of the problem of differences in the distribution of hyperspectral image data in different scenes,this paper studies from the perspective of semi-supervised learning and proposes a two-stage deep domain adaptation hyperspectral image classification method based on semi-supervised learning.This method can minimize the distribution difference between the two domains,and use a few target labeled samples to learn more discriminative deep embedding space.First,the method learns the deep embedding space by minimizing the distribution distance between the source domain and the target according to the MaximumMean Discrepancy.Second,the method is based on pairwise loss to minimize the distribution distance between samples from different domains but the same category,and to maximize the distribution distance between samples from different domains from different categories,while using Spatial-Spectral Siamese Network to reduce data distribution differences And learn more discriminative deep embedding space.Finally,for the classification task,the softmax layer is replaced with a linear support vector machine to minimize the margin-based loss and classify the target domain samples.