Land-Oriented Scene Graph Generation for High-Resolution Remote Sensing Imagery: A Specialized Dataset and Semantic–Visual Collaborative Method
TL;DR
This study constructs the first land-oriented remote sensing SGG dataset by integrating and refining land-scene samples from ReCon1M and satellite-based terrain and relationship and proposes a semantic–visual collaborative SGG framework, which combines oriented object detection, global contextual modeling, and semanticprior fusion to alleviate semantic–visual conflicts in remote sensing scenes.
Abstract
High-resolution remote sensing image interpretation is evolving from object-level perception toward semantic cognition. However, existing scene graph generation (SGG) methods are difficult to adapt to remote sensing imagery due to the lack of dedicated land-oriented benchmarks, semantic–visual inconsistency, and severe long-tailed relationship distributions. To address these issues, this study constructs the first land-oriented remote sensing SGG dataset by integrating and refining land-scene samples from ReCon1M and satellite-based terrain and relationship, containing 19 658 images, 50 object categories, and 45 relationship categories. Furthermore, a semantic–visual collaborative SGG framework is proposed, which combines oriented object detection, global contextual modeling, and semanticprior fusion to alleviate semantic–visual conflicts in remote sensing scenes. In addition, a dual-level prototype-constrained relation learning strategy is introduced to improve rare relationship recognition under long-tailed distributions. Experimental results show that, compared with PE-Net, the proposed method improves $R@100$/$mR@100$ from 43.59/22.75 to 52.81/36.51 under scene graph classification and from 15.09/5.10 to 17.60/8.94 under scene graph detection. The proposed dataset and framework provide an effective benchmark and methodological paradigm for high-level semantic understanding of remote sensing imagery.