DAMNet: Hyperspectral Image Classification Method Based on Dual Attention Mechanism and Mix Convolutional Neural Network

Deep learning approaches have been broadly used in the study of hyperspectral images (HSI) with promising results in recent years. However, improved network complexity and features extraction still have scope for improvement. As a result, we propose a hybrid convolutional neural network (DAMNet) based on a 3D-dual attention mechanism (spectral and spatial attention mechanism) to fully capture spectral and spatial features and improve classifying precision. This is a spectral-spatial 3D-CNN that is being followed. The 3D-CNN allows for the encoding of joined spatial-spectral features of a stacks of images. The 2D-CNN learns more extensive abstract spatial representative on top of the 3D-CNN. When compared to using a 3D-CNN alone, the usage of mixed CNN streamlines the model. By incorporating the 3D-dual attention mechanism module into the network, it can enhance the representation of image features in various dimensions, thereby increasing classification precision. Image features from multiple levels are represented in the image, improving classification accuracy. This article improves the classification accuracy of common datasets with the architecture DAMNet, while also validating the superiority of the attention mechanism on the Xuzhou dataset.