Research On Semi-supervised Classification Method Of Little Sample Hyperspectral Remote Sensing Image Based On Generative Adversarial Networks

Hyperspectral remote sensing image classification is a research hotspot in the field of remote sensing image processing,and has been widely concerned by scholars at home and abroad for a long time.In recent years,the development of deep learning has brought new ideas and methods to the classification of hyperspectral remote sensing images.Based on the massive training samples,the deep learning method refreshes the accuracy of remote sensing image classification again and again,and has made remarkable progress.However,the acquisition of hyperspectral remote sensing data is difficult,the manual labeling is complex,and the lack of training samples has made the development of classification algorithms for hyperspectral remote sensing images limited.In this context,this paper focuses on the theme of "semi-supervised classification method of hyperspectral remote sensing image based on small training samples",and carries out related research,and designs a new classification scheme of hyperspectral remote sensing image.The main results of the research are as follows.(1)Considering the problem of insufficient samples of hyperspectral remote sensing classification,the idea of generative adversarial networks is applied to the hyperspectral remote sensing classification,and the ideas of semi-supervised classification are used to play the role of many unlabeled samples.The mutual adversarial of the neural networks to learning the distribution of the data from the unmarked samples,so that higher classification accuracy can be achieved with fewer labeled samples.(2)There are two problems in the current hyperspectral remote sensing classification method based on the generative adversarial networks:①the function of the discriminator to identify the generated sample and the predicted sample label is incompatible,which causes the generator and the discriminator to not reach the best at the same time.②The generator only receives the relevant data distribution without receiving the tag information returned by the discriminator,resulting in a problem that the sample semantics are not clear.So the function of the discriminator is separated into two independent neural networks:a classifier and a discriminator,a generator and a classifier to characterize the conditional distribution between the image and the label,and the discriminator only recognizes the real or fake of the image pair.Ensure that both the classifier and the generator can converge to their respective bests.(3)A classification model HS-TGAN based on spectral features of hyperspectral remote sensing images is designed.The model makes full use of the data distribution information provided by unlabeled samples.Only 10 samples can be used for each category to achieve higher precision classification results.Compared with the current published deep learning methods,it has certain advancement.(4)Aiming at the spatial characteristics of hyperspectral remote sensing images,a deep learning model S2TGAN suitable for 3D feature extraction and classification was designed.The improved bilateral filter is used to perform spatial-spectral information filtering.The model is experimentally verified on the original image,principal component analysis image and filtered image of multiple public data sets.The experimental verification shows that under the same experimental conditions,the classification effect is better than that based on spectral features.In this paper,some shortcomings of the hyperspectral classification model based on the generation-oriented network are improved.Several advanced classification algorithms are proposed,and the high-spectral classification model based on the generation-based confrontation network is constructed.The idea is novel,the theory is sufficient,and it has a good promotion.Space and development prospects deserve further research and exploration.