PeftCD: Leveraging Vision Foundation Models With Parameter-Efficient Fine-Tuning for Remote Sensing Change Detection

Abstract

To tackle the prevalence of spurious changes, the scarcity of annotations, and the difficulty of cross-domain transfer in multitemporal and multisource remote sensing imagery, we propose Parameter-Efficient Fine-Tuning Change Detection (PeftCD), a change detection framework built upon Vision Foundation Models (VFMs) with Parameter-Efficient Fine-Tuning (PEFT). Specifically, PeftCD adopts a shared-weights Siamese encoder instantiated from a VFM, into which Low-Rank Adaptation (LoRA) and Adapter modules are injected as fine-tuning strategies, so that only a small number of additional parameters need to be trained for task adaptation. To better explore the potential of VFMs in change detection, we investigate two representative backbones: the Segment Anything Model v2 (SAM2), which provides strong segmentation priors, and DINOv3, a state-of-the-art (SOTA) self-supervised representation learner. Meanwhile, PeftCD employs a deliberately minimal and efficient decoder to highlight the representational capacity of the backbone models. Extensive experiments demonstrate that PeftCD achieves SOTA performance across multiple public datasets, including SYSU-CD (IoU 73.81%), WHUCD (92.05%), MSRSCD (64.07%), MLCD (76.89%), CDD (97.01%), S2Looking (52.25%), and LEVIR-CD (85.62%), with notably precise boundary delineation and strong suppression of pseudochanges. Overall, PeftCD achieves a favorable balance among accuracy, efficiency, and generalization, offering an efficient paradigm for adapting VFMs to practical remote sensing change detection.