A Multi-Scale Additive Enhanced Network for Remote Sensing Scene Classification

In Remote Sensing Scene Classification (RSSC), the Convolutional Neural Network (CNN)-Transformer hybrid model shows great potential in extracting multi-scale features. However, RSSC still faces challenges. The first is the issue of disparate target sizes in photos obtained by remote sensing. Secondly, there is a significant degree of resemblance between the classes, which increases the difficulty of the classification task. In response to the above issues, a multi-scale additively enhanced CNN-Transformer hybrid network is proposed. The model extracts richer feature through multi-scale feature extraction to alleviate the problem of large differences in target size. Multi-Scale Additive Enhanced Block (MSAE Block) is designed, which extracts local and global features through multiple branches, supplementing Chunked Additive Attention Module (CAAM) with bi-dimensional features and alleviating the problem of high inter class similarity. A Multi-Scale Additive Enhanced Network (MSAENet) is proposed based on MSAE Block and Bridging Residual Module (BRM) and is validated on two datasets, WHU-SIRI and AID. Based on experimental data, the classification accuracy of MSAENet is better than the contrasting models based on CNN, Transformer, and CNN-Transformer hybrid architecture. It achieves a classification Top-1 accuracy of 97.71% on the WHU-SIRI dataset with just 0.74G of FLOPs, 1.67% higher than ResNet50 and 5.83% better than Vision Transformer (ViT).