Classification Of Hyperspectral Image Based On Graph Convolutional Neural Network And Reinforcement Learning

With the vigorous development of learning patterns such as deep learning,reinforcement learning,and meta-learning,general artificial intelligence has gradually become the goal of artificial intelligence.Typically,AlphaGo has achieved a revolutionary breakthrough in this area by using self-learning models.Intelligent feature extraction instead of complicated manual feature extraction has become a popular trend in the application of hyperspectral image classification.At present,many deep learning-based methods,such as SAE,DBN,CNN,etc.,have been demonstrated powerful feature extraction ability on hyperspectral images.Especially,CNN benefits from the particularity of convolution operators and achieves very competitive classification performance on hyperspectral data.In pixel-level HIC,each pixel is a discrete sample.CNN operators with fixed structures usually cause blurring of edges.While graph convolution operators with relatively flexible structures are more suitable for pixel-level HIC.In addition,thanks to the development of intelligence and autonomous learning,exploratory and continuous updating of the classifier can alleviate the burden of labeling and the problem of fewer labeled samples derived from it.Therefore,in this paper,from the perspective of image structural features and intelligent classification,GCN and RL are introduced into HIC.The main work is as follows:(1)A HIC method based on a novel similarity measure and graph attention convolutional neural network is proposed.In this method,which can enhance important information,suppress useless information,and improve reliability features during network iteration,a new similarity measure method KSAM-SID is combined with attention mechanism to choose samples flexibly,and to construct a graph shift operator capable of assigning different attention to the neighborhood samples.Experiments show that the method performs well in categories with a small size and classification boundaries.(2)Aiming at the over-smoothing problem of the classic GCN and the problem that the parameter k of KNN can’t be well adapted to different regional scales of the HSI during the construction of graph adjacency matrix,a method combined multi-scale adaptive attention mechanism with deep graph convolution network is proposed for HIC.The network model,which can effectively alleviate the gradient disappearance problem of deep GCN and implement multi-scale feature extraction to adapt to homogeneous regions of different structures,is implemented by combining skip-connection with dilated convolution.In addition,dynamic attention mechanism is introduced to solve the problem that the sample neighborhood influence is fixed in the iterative process,which can adaptively update the similarity between samples and aggregate more effective features.Compared with method(1),deep GCN can extract higher-order and more abstract features,and can perform better in different homogeneous regions.(3)From the perspective of intelligent learning,a HIC method based on RL and deep learning is proposed.In this method,HIC is regarded as a guessing game,in which the game state is set as the current classification situation of the training sample,the sample actions are set as the label set,and the delay reward is set according to the real label.Then,deep learning is used to calculate the instant reward.The network parameters are updated through several episodes of simulation game,so that the instant reward becomes reliable.This method can learn model parameters autonomously and shows great potential in HIC.The HIC method proposed in this paper is dedicated to mining the relationships between samples that conform to the data features,which is beneficial to training classifiers with relatively high robust performance using few or no labeled samples.The experiment proves that the application of graph convolutional neural network and reinforcement learning can learn hyperspectral image classifier from a new perspective,which helps to further improve the classification performance and reduce labor costs.