Research On Hyperspectral Image Classification With Small Samples

Hyperspectral image(HSI)is a three-dimensional data cube composed of tens or even hundreds of spectral bands with rich spectral information.With the development of sensor technology,the spatial resolution of hyperspectral images has been greatly improved.Deep learning has shown excellent performance in HSI classification by mining spectral features and spatial features,but deep learning methods often rely on the availability of a sufficiently large set of labeled samples for training.However,HSI labeling is a complicated and laborintensive task.Therefore,how to achieve the classification task of hyperspectral images with limited labeled samples is a hot issue in the field of HSI processing.In this paper,we propose three different classification methods for HSI with few labeled samples based on active learning,semi-supervised classification,meta-learning,domain adaption,and attentional feature extraction mechanisms.The main research of this paper is as follows:1.A local-global active learning semi-supervised network(semi-LG-AGCN)based on graph convolutional networks is proposed.The proposed method is an end-to-end active learning network that extracts local and global information by constructing local adjacency matrices and global adjacency matrices,and then measures the discriminative information of unlabeled samples by an uncertainty algorithm to filter out the unlabeled samples that are most informative for the classification task.Unlike the previous labeling performed manually by a human in active learning,the studied semi-supervised classification is achieved by expanding the training set of a supervised classifier with high-confidence pseudo-labeled samples identified by the preceding active learning step and retraining.Experimental results demonstrate that the proposed semi-LG-AGCN outperforms not only other approaches to semi-supervised classification but also several existing fully supervised methods which have demonstrated effectiveness on limited training data.2.A cross-domain few-shot learning method(GCM-FSL)based on graph convolution contrast learning is proposed.Since HSI is discontinuous graph structure data,the proposed method uses a graph convolution network to extract structural information of HSI.The proposed method designs a positive and negative sample alignment module,which establishes positive and negative sample correspondences for source and target domain data based on similarity and utilizes contrast learning to improve the similarity between positive samples.In addition,a domain distribution metric module is designed for mitigating the impact of the difference in distribution between source and target domain data by maximum mean discrepancy.The experimental results conducted on two different target domain datasets demonstrate that the proposed GCM-FSL outperforms not only other deep learning methods with limited labeled samples but also existing HSI few-shot learning methods.3.A deep cross-domain few-shot learning method(DAFSL-PC)based on residual attention module and prototype domain adaption is proposed.In dealing with the domain drift problem,a nonparametric prototype classifier is proposed to replace the shared classifier scheme commonly used in deep learning to alleviate the problem of the source domain classifier cannot be applied to the target domain classification when the data distribution is inconsistent.In addition,an attention-based residual network is designed for extracting more discriminative embedded features of HSI.The experimental results conducted on two different target domain datasets demonstrate that the proposed DAFSL-PC outperforms not only other deep learning methods with limited labeled samples but also existing HSI few-shot learning methods.