Abstract: ‍ ‍Hyperspectral images contain rich spatial and spectral information， and several methods exist for the accurate classification of hyperspectral images. To address the problems in hyperspectral remote sensing images， such as the difficulty of deep semantic feature extraction and the increase of network computational cost with increasing network layers， this paper proposes a hyperspectral image classification algorithm based on convolutional residual attention network. First， the spectral-spatial feature extraction module was designed to enable the network to capture the information in the hyperspectral image more comprehensively. Second， the attention enhancement backbone was designed to address important channels and specific areas through the interaction between the attention mechanism and the residual network， improving the discrimination ability of features and enhancing the perception of spatial information. Finally， attention contrastive learning was introduced to enhance the intra-class correlation and inter-class discrimination of samples. A fixed proportion of training and testing samples were randomly selected on three public hyperspectral datasets for experimental analysis. Indicators such as overall accuracy （OA）， average accuracy （AA）， and Kappa coefficient were used to evaluate the classification results. The dataset selected in this paper includes both high- and low-resolution hyperspectral data， which can fully test the generalization of the proposed method. Experimental results on three datasets show that compared with some representative hyperspectral image-classification algorithms， the proposed method has better classification performance. The proposed method plays an important role in promoting research in this field. However， the proposed method has limitations including the use of single data source and difficulty in determining the input scale. In future research， we will address the exploration from the perspective of multi-scale and multi-source data and consider irregular and regular inputs to improve the classification performance of hyperspectral images.