解耦(概率)
遥感
频率响应
声学
控制理论(社会学)
计算机科学
频域
信号处理
雷达探测
时频分析
频率调制
滤波理论
噪音(视频)
电子工程
物理
空间技术
雷达跟踪器
作者
Rong Wang,Tao Lei,Tongfei Liu,Xingwu Wang,Yingbo Wang,Maoguo Gong,Asoke Kumar Nandi
标识
DOI:10.1109/tgrs.2026.3706730
摘要
Remote sensing change detection (RSCD) plays a critical role in understanding surface dynamics. Recently, Mamba-based methods have demonstrated promising performance in visual tasks and RSCD due to their efficient modeling of long-range dependencies. However, these methods still face two major limitations. First, the state-space modeling mechanism of Mamba exhibits inherent low-pass characteristics, which suppress high-frequency variation information of remote sensing images, such as edges and textures, leading to inaccurate delineation of changed boundaries. Second, as RSCD relies heavily on modeling differences between bi-temporal images, feature representations based solely on either the spatial or frequency domain are insufficient to fully capture the semantic structure of really changed regions, often resulting in false alarms and missed detections in complex remote sensing. To address these issues, a novel framework termed Freq2MCD is proposed for RSCD, which integrates frequency-aware representation learning with dual-domain collaborative modeling strategies. Specifically, VMamba is adopted as the backbone to exploit its global modeling capability, while a frequency-aware separation strategy (FAS) is introduced to decouple and enhance frequency components, thereby alleviating the low-pass bias of Mamba. Furthermore, a dual-domain adaptive synergy strategy (DDAS) is designed to model bi-temporal differences from local structural and global frequency perspectives, improving the discriminability of changed regions in complex remote sensing scenes. Extensive experiments on three public datasets demonstrate that Freq2MCD consistently outperforms state-of-the-art methods, achieving IoU improvements of 2.06%, 2.88%, and 1.77% over the second-best Mamba-based approaches on the SYSU, CLCD, and WHU-CD datasets, respectively. The code will be available athttps://github.com/199-wh/Freq2MCD/
科研通智能强力驱动
Strongly Powered by AbleSci AI