Abstract: ‍ ‍Remote sensing change detection， as a crucial means to observe and analyze surface alterations， utilizes bi-temporal images as input to predict the “locations” of changes. The recent emergence of foundation models， such as the segment anything model （SAM）， has demonstrated powerful universality and generalization capabilities， offering more effective solutions for change detection tasks. However， their direct use in many remote sensing applications is often unsatisfactory due to the special imaging properties of remote sensing images. Additionally， SAM is initially designed for segmenting single-temporal images， relying on the manual addition of point or box prompts， which proves overly intuitive when handling bi-temporal image inputs. To address these challenges， this paper proposes temporal difference prompted SAM for remote sensing change detection （TDPS）， exploiting the robust visual recognition capabilities of the SAM to improve change detection in remote sensing images. Specifically， we first introduce low-rank fine-tuning parameters into the SAM backbone network to mitigate domain shifts from natural images to remote sensing images. Second， a temporal difference prompt generator is designed to optimize features of bi-temporal images together with query embeddings， yielding prompt vectors suitable for change detection tasks. Finally， extensive experiments demonstrate the effectiveness of the proposed approach， achieving state-of-the-art performance on two commonly used change detection datasets， LEVIR-CD and WHU-CD， with F1 scores improved by 1.4% and 2.5%， respectively， compared to state-of-the-art methods.