Hyperspectral image (HSI) classification is a vital remote sensing technique that finds numerous real-world applications ranging from environmental monitoring to medical diagnosis, food safety, and security. Recently, deep convolutional neural networks have shown impressive results due to their automatic local feature extraction capabilities. However, current deep networks suffer from severe overfitting on HSI datasets with limited training samples, leading to unsatisfactory classification accuracy. Additionally, most of the attention mechanisms adopted in these networks are twodimensional, neglecting the relationship between spatial and spectral dimensions. To address these problems, we propose a dual-branch, triple-attention mechanism network (DBTANet) for HSI classification. Specifically, DBTANet adopts two parallel branches that independently extract spatial and spectral features, effectively enhancing the divergence and independence of the features. Furthermore, we have developed a unique module with a triple-attention mechanism to automatically focus on the spatial and spectral features that are most helpful for classification, while allowing for information interaction across spatial positions and spectral channels. Additionally, we introduce dynamic convolution to allow for adaptive adjustment of the convolution kernel parameters based on different inputs, further enhancing the feature expression ability of DBTANet. Experiments on three real HSI datasets demonstrate that compared to other recent methods, DBTANet achieves significantly better classification performance, particularly with limited training samples.
|