ABSTRACT Binary change detection (BCD) in remote sensing is crucial for environmental monitoring and urban planning. While vision foundation models like SAM2 offer new opportunities, adapting them to multi‐temporal imagery faces challenges from domain gaps, spectral variations, and pseudo‐changes. Existing approaches typically require full parameter fine‐tuning of large foundation models, leading to prohibitive computational costs and overfitting risks on limited remote sensing datasets. To address these limitations, we propose SAM2‐FreqCD, a frequency and spatiotemporal aware framework that introduces the Meta SAM2 Hiera encoder to BCD. Thus, we designed the SA (SAM2 with adapter) module and a decoder with a frequency‐aware block (FAB) and spatial‐targeted channel blocks (STCB). SA module integrates Hiera blocks with parameter‐efficient adapters for frozen‐parameter fine‐tuning. FAB employs mixture‐of‐experts for frequency decomposition, while STCB enhances multi‐temporal fusion through cross‐temporal attention to implicitly simulating spatiotemporal interactions, collectively improving discriminative feature extraction. Experiments on public datasets including SYSU‐CD, WHU‐CD and LEVIR‐CD+ demonstrate that our method outperforms strong baselines such as ChangeMamba and TransUNetCD, especially achieving an F1 score of 94.89% on WHU‐CD while requiring substantially fewer trainable parameters and exhibits strong robustness on degraded data.
Ren et al. (Thu,) studied this question.