Title: SCL: Advancing Domain Generalization in Remote Sensing Change Detection Through Single-Temporal Multimodal Contrastive Learning

Abstract: While CNN and transformer-based models for change and anomaly detection have shown significant success on paired datasets in recent years, they often struggle with cross-dataset generalization. This limitation stems from domain-specific architectures and a heavy reliance on extensive labeled paired data. To address these challenges, we propose SCL (Single-temporal multimodal Contrastive Learning), a foundation model for change detection that leverages visual-language pre-training and requires no training on target datasets. We introduce the Dynamic Text-vision Context Optimization (DTCO) module to enhance the model's capacity to capture the interplay between textual and visual contexts through prompt learning. Furthermore, to mitigate the dependency on paired data, we present a Single-temporal trAINing strategy (SAIN) combined with controllable generation. This approach enables training on a vast corpus of existing single-temporal images, eliminating the need for paired labels. Extensive evaluations on diverse real-world change detection benchmarks confirm that SCL achieves superior performance and generalization capabilities, surpassing current state-of-the-art methods under the tested conditions. The code is publicly accessible at https://github.com/Kane-Du/scl-cd.git.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC