Tri branch attention enhanced 3DUNet for remote sensing based hyperspectral image classification

Several advanced strategies utilize deep learning to efficiently classify hyperspectral images (HSIs). UNet and other segmentation techniques are proficient with low-resolution 3D data; nevertheless, they encounter challenges with high-resolution hyperspectral images due to class imbalance and resolution degradation. Consequently, both local and global data are ignored. The HSI consists of diversified and diminished non-linear combinations. Enhancing classification accuracy is inherently difficult due to the limited availability of labeled data. To overcome the limitations of conventional U-Net architecture, we proposed a three-branched 3DUNet architecture. Each branch is designed to address a specific issue such as the spectral branch focuses on extracting the spectral dependencies, the spatial branch enhances spatial feature learning, and the combined branch integrates both spectral and spatial cues for robust joint representation. This design ensures complementary feature extraction and improved classification performance in HSI, Moreover each branch is equipped with the specific attention mechanism, which extracts the more relevant features. After this these features are then integrated and processed through a fully connected layer to capture all non-linear combinations. This improves the accuracy by augmenting the reduced and diverse information of small pixels. Experiments are performed on well known benchmark (HSI) datasets, including Indian Pines, Pavia University and Houston-2018, utilizing evaluation metrics such as average accuracy (AA) and overall accuracy (OA). The proposed model shown remarkable classification accuracy on the Pavia University dataset (99.67 %) Indian Pines dataset (98.02 %) and Houston-2018 (99.88 %). The proposed method demonstrates more robustness and superiority compared to existing models.