HyperCLIP-CD: Hyperbolic Multimodal Representation Enhancement for Remote Sensing Change Detection

Remote sensing change detection (RSCD) is an important branch of remote sensing image analysis, aimed at identifying changes in land features from bitemporal images. In recent years, change detection methods have predominantly been based on deep learning approaches, some even integrating foundational vision-language models to provide multimodal data, achieving impressive results. However, existing multimodal methods merely utilize multimodal data superficially, without delving into the representation relationship between vision and language. Recent research has proposed that non-Euclidean spaces, such as hyperbolic space, offer advantages in modeling graph-text relationships. To this end, we introduce HyperCLIP-CD, a multimodal network specifically designed for the RSCD task. It leverages the property of hyperbolic space, which is suitable for embedding tree-like data, and combines CLIP to capture the hierarchical structure and non-Euclidean features of image and text data.Specifically, we have designed a hierarchical semantic textual prompting (HTP) generation method tailored for change regions that meet the requirements. Based on a dual-branch CLIP architecture, we extract multimodal features from bitemporal images and propose a hyperbolic space-based graph-text alignment module (HMSA) to obtain deeper semantic change information. HyperCLIP-CD has achieved higher detection accuracy than other advanced networks on both the binary change detection datasets LEVIR-CD and CDD, as well as the semantic change detection dataset SECOND.