Hyperspectral Image Classification Based on Multiscale Residual Hybrid Network With Attention Mechanism

In recent years, deep learning-based hyperspectral image (HSI) classification methods have shown remarkable effectiveness. Nevertheless, labeling process of HSIs is laborious and time-consuming, resulting in the limited number of labeled samples. It is still a challenging problem to obtain satisfactory classification accuracy from the limited training samples. The richness of spectral and spatial information helps to accurately classify all kinds of objects in the image. Effective use of spatial and spectral information in HSIs is essential to complete the classification task. To address the problem of low classification accuracy of existing algorithms on small samples of HSIs, this paper proposes a multi-scale residual hybrid network with attention mechanism, MSRHN-A. This method designs two multiscale feature extraction modules within a 3D-2D hybrid network architecture, aimed at separately extracting fused spatial-spectral features and high-level spatial features. In addition, a channel and spatial residual attention module is established by combining the residual connection mode with the attention mechanism. The attention module optimizes the feature extraction process so that the model can focus on important areas in the image. To improve speed and prevent overfitting, the model employs BN, Dropout, and early stopping strategies. We conducted experiments on three publicly available hyperspectral datasets: Indian Pines, University of Pavia, and Salinas Valley. The experimental results show that compared with existing methods, our proposed MSRHN-A classification model exhibits significant performance, and even in the case of insufficient training samples, it can still achieve satisfactory classification accuracy.