Lithological Classification by Hyperspectral Remote Sensing Images Based on Double-Branch Multiscale Dual-Attention Network

Abstract

Although many scholars have realized that deep learning methods have great advantages in hyperspectral lithology classification now, most of them use simple convolutional neural networks for discussion, which are difficult to effectively extract spectral sequence information from hyperspectral image (HSI). To address this issue, we propose a double-branch multiscale dual-attention (DBMSDA) network using two HSIs multiple different rock types as data sources. This network utilizes the multiscale spectral residual self-attention structure and dense connections in the spectral branch to extract diagnostic multiscale spectral information from HSIs, improving the ability to reuse spectral information. In the spatial branch, dense connections are used to fully extract diagnostic spatial information, and finally, the two are fused for classification. We compare the DBMSDA network with several traditional machine learning and deep learning methods. Experimental results show that the DBMSDA network has the best lithology classification performance and robustness. Its OA on the ZCD1 and HCD2 datasets is 85.08% and 90.80%, respectively, representing an improvement of 5.95% and 2.4% over the lowest OA, confirming that the DBMSDA network can effectively improve accuracy and is superior to other methods. Meanwhile, the experiment also showed that the DBMSDA network has good classification performance on HSIs of various rock types and HSIs from different data sources. Finally, ablation experiments were conducted to demonstrate that the DBMSDA network relies more on large training samples compared with traditional machine learning models, and that the DBMSDA network has better applicability and stability under the same sufficient training samples.

Keywords

• Deep learning
• dual-attention mechanism
• hyperspectral image (HSI)
• lithology classification
• multiscale features