High-Level Semantic-Guided Lightweight Network for Remote Sensing Image Change Detection

Abstract

Existing remote sensing image change detection (RSCD) methods generally suffer from high computational overhead and insufficient utilization of high-level semantic information. To address these issues, this letter proposes a high-level semantic-guided lightweight network for RSCD, termed CGLNet. In particular, a dual-difference fusion module (DDFM) is designed to jointly model the magnitude and direction of changes, thereby enhancing the representation capability of change features. In addition, a high-level semantic guidance module (HSGM) is constructed to strengthen global semantic features and provide consistent global constraints for the decoding stage. Furthermore, a guided decoder is developed to achieve efficient multiscale feature fusion and feature reconstruction. Experimental results demonstrate that with only 0.62 GFLOPs and 0.99 M parameters, CGLNet outperforms 12 state-of-the-art (SOTA) methods on both the WHU-CD, NJDS, and LEVIR-CD datasets, achieving $F1$ -scores of 90.77%, 78.09%, and 91.47%, respectively. The proposed method achieves high detection accuracy with low network complexity, indicating strong potential for practical applications. The source code is available at https://github.com/bobo59/CGLNet