A Lightweight CNN-Transformer Parallel Fusion Architecture for Accurate Remote Sensing Image Classification

Remote sensing image classification plays a significant role in urban planning, agricultural monitoring and other areas. However, achieving a lightweight and accuracy classification method remains a major challenge. To address this issue, we propose a lightweight parallel feature fusion network, PMNet. First, we design a parallel encoder structure for global-local feature learning to solve the problem of insufficient global modeling in CNN structure. Second, to tackle the issue of insufficient interaction between features and information loss caused by feature concatenation in existing methods, we propose a multi-directional cross-attention fusion module. Next, to accurately model the channel information generated at different stages of the encoder, we design a novel channel attention module. Finally, the model is trained on the WHDLD dataset using a joint loss function. Experimental results show that PMNet achieves the highest scores in MIoU, FWIoU and MPA, reaching $\mathbf{6 4. 4 1 4 \%}$, 75.673% and 77.028%, respectively, with only 10.154 M parameters, outperforming other advanced networks.