3D-CAN: A 3D Convolution Attention Network for Feature Extraction and Classification of Hyperspectral Images

In recent years, the deep learning method has gained widespread popularity in hyperspectral image classification (HSI) tasks because of remarkable feature extraction capability. However, conventional deep learning approaches often yield suboptimal classification performance due to the limited discrimination ability of the extracted features. Moreover, the challenge of obtaining high classification performance with a limited number of samples in hyperspectral images (HSIs) has emerged as a prominent research area of interest. To address this concern, this letter provides a novel 3D convolution attention network (3D-CAN) for feature extraction and HSIs classification. The attention mechanism employed in this context enables capturing both the long-range spatial dependencies of HSIs in the horizontal and vertical directions and the varying significance levels across distinct spectral channels. The experiments are executed on both the new data sets and the publicly accessible benchmark data sets to expose the efficiency and durability of the proposed model. Empirically, it has been demonstrated that the presented model succeeds over the remaining state-of-the-art approaches in terms of classification accuracy.