Skip to content
Book Open access

ASF-Tune: Adaptive Spatial-Frequency Tuning for Change Detection

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 35 references

Abstract

Vision Foundation Models (VFMs), benefiting from large-scale pre-training, provide generic and highly transferable visual priors, making them a promising foundation for Remote Sensing Change Detection (RSCD). However, adapting VFMs to RSCD remains challenging due to the domain gap between natural and remote sensing imagery and the lack of bi-temporal joint adaptation in existing Parameter-Efficient Fine-Tuning (PEFT) strategies. In this paper, we propose a robust and elegant new paradigm for adapting VFMs to multi-temporal RSCD. Specifically, an Adaptive Spatial-Frequency Adapter (ASF-Adapter) is designed to jointly adapt bi-temporal representations with minimal trainable parameters by decoupling spatial and frequency modulation in the feature space. Notably, in the frequency branch, low-frequency components are leveraged for cross-temporal global style alignment, while high-frequency components are selectively refined to preserve informative boundaries and suppress non-semantic noise. Together with strengthened multi-scale structural context modeling, this design improves change-aware representation learning. In addition, a Text-Anchored Semantic Contrastive (TASC) loss is introduced to enhance the semantic consistency of changed regions by transferring semantic priors from vision-language models during training, without incurring additional inference overhead. Extensive experiments on two multi-semantic and three binary RSCD datasets, along with cross-dataset evaluation, demonstrate that our framework achieves state-of-the-art (SOTA) and exceptional generalization.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.