Remote Sensing Scene Classification with Multi-Granularity Perception and Residual Attention
Lists
Only you see this
Abstract
Currently, Vision Transformer has become the mainstream method for remote sensing image understanding, especially for remote sensing scene classification. However, it tends to focus on local features with high redundancy when the network is shallow, and it often falls into the problem of feature collapse when the network is deep, leading to the disappearance of some low-level features. It is likely to influence the scene classification performance. Therefore, we propose a novel transformer-based method for remote sensing scene classification. It consists of two key modules: the Multi-Granularity Perception Module (MGPM), which captures multi-granularity features and expands the receptive field in shallow layers, and the Residual Attention Module (RAM), which integrates low-level features while preserving global context in deep layers. By using these two modules, the proposed method effectively enhances feature diversity and preserves essential low-level information, addressing the issues of redundancy and feature collapse. Experimental results on the UCM and NWPU datasets have demonstrated the feasibility and effectiveness of the proposed method for remote sensing scene classification, providing a valuable reference for remote sensing scene classification applications.
BibTeX
@misc{liu2025remote,
author = {Bingrong Liu and Hongzhang Xu and Kangyuan Zhang and Zheng Dai and Zhengqi Dong and Jun Shi},
title = {{Remote Sensing Scene Classification with Multi-Granularity Perception and Residual Attention}},
year = {2025},
doi = {10.1109/igarss55030.2025.11243703},
}Loading related content...