Research On Deep Learning Method Of Hyperspectral Remote Sensing Image Under Small Sample Conditions

Marking hyperspectral remote sensing images is a very difficult task,so a small number of labeled samples become a hindrance to deep learning methods for hyperspectral image classification.In this regard,this paper proposes a supervised and semi-supervised deep learning method suitable for small sample image classification.The main work is as follows:(1)A supervised deep learning model suitable for small sample hyperspectral image classification is proposed,which is a multiscale multilevel spectral spatial feature fusion(Multiscale Multistage Spectral Spatial Features Fusion,M~2S~2F~2)framework.In order to enable the model to extract more representative and discriminative features in small samples,the input is set to two spatial scales of 3×3 and 15×15 and the primary,intermedia and advanced features are combined to form multistage features.This multistage feature from different spatial scale inputs and fused through different complexity features is highly discriminative.Finally,the spectral and spatial features are fused,and the two complement each other and correct each other to improve the classification accuracy.Experimental results on the IN,UP and KSC datasets show that the framework has good classification performance.(2)A semi-supervised deep learning method(Multiscale Generative Adversarial Networks,MGAN)suitable for small sample hyperspectral image classification is proposed,which is suitable for small sample classification by increasing the number of samples.The generator in MGAN trains a new sample that is as similar as possible to the original sample with only a small number of labeled samples,and trains the discriminator with the original sample.In order to extract better features,the discriminator is set to two scale inputs of 33×33and 64×64 and gives the discriminator multi-class discriminating ability and not only the true and false samples can be judged.Experimental results on the IN and UP datasets show that this network has good classification performance.Through theoretical and experimental comparison analysis,the application of the above two methods is summarized.M~2S~2F~2is classified by extracting sample features.In the case of particularly small samples,feature extraction may not be sufficient,so the supervised deep learning(M~2S~2F~2)framework is more suitable for more training samples.MGAN can extract features after relatively increasing the number of samples,so the semi-supervised depth(MGAN)network is suitable for relatively few training samples.