| Hyperspectral images with rich spectral information are widely used in applications such as disaster monitoring,military detection and fine agriculture,but the numerous spectral bands provide a large amount of information and also bring challenges to the subsequent image processing.Among them,hyperspectral image classification is a core task in the image processing process,and the accuracy of its classification of samples is crucial to the subsequent image analysis and applications.Machine learning is a commonly applied classification technique in the early development of hyperspectral image classification tasks,but only the spectral information of images is utilized,ignoring the spatial geometric information.Meanwhile,the feature engineering design process requires a large amount of a priori knowledge.Deep learning methods,which can automatically capture higher-order features from raw data,have gradually replaced machine learning hyperspectral classification methods.Convolutional neural networks,as one of the deep learning methods,are widely used in hyperspectral image classification tasks with the features of local connectivity and weight sharing.However,the convolution operation in convolutional neural network makes it confined to the Euclidean space and cannot effectively obtain the global relationship features between pixels.The graph neural network has a strong ability to capture object dependencies through the aggregation of node information,which can make up for the deficiency of the convolutional neural network approach.In view of this,the application of graph convolutional networks to hyperspectral image classification is investigated as follows.(1)A graph attention network SSGAT based on superpixel segmentation is constructed,which can adaptively acquire long-range contextual information and relational features of images based on the difference of importance of different bands,and solve the problem of insufficient utilization of relational features by convolutional neural networks.Based on the construction method of graph structure,the super-pixel segmented super-pixel blocks are considered as graph nodes in the graph structure,which effectively reduces the complexity of constructing the graph structure and reduces the classification graph noise.For the optimization problem of the network,the residual structure is introduced into the network,and it is proved through experiments that the method can effectively reduce the risk of network overfitting.The classification accuracies of SSGAT and the comparison algorithm are tested on three representative hyperspectral image datasets,and the overall classification accuracies of 94.11%,95.22%,and 96.37% are obtained on the three datasets by comparison tests and ablation experiments of SSGAT,respectively,which have excellent performance compared with other methods and have obvious advantages for the classification problems of large scale regions.(2)To address the problem that superpixel segmentation processing will make the features between neighboring pixels similar,resulting in a certain degree of detail loss.In order to obtain finer classification results,we propose a classification network FCGN that combines the graph convolutional branch and the convolutional branch,which contains two branches: the graph convolutional network based on superpixel segmentation and the convolutional network with added attention mechanism.By fusing the features of the two branches,the complementary features complement each other to complete the classification process and propose a solution to the problem of missing details caused by superpixel segmentation.The FCGN and the comparison algorithm are tested and ablated on three datasets,and the overall classification accuracies of 98.78%,98.99%,and 98.69% are still obtained despite the small number of training samples,which proves the effectiveness of the algorithm.The network can obtain excellent classification performance in both large scale and small scale regions. |
PDF Full Text Request