Spatially Anisotropic Reasoning Network for Remote Sensing Scene Graph Generation

Abstract

Remote sensing scene graph generation (RS-SGG) aims to advance remote sensing image interpretation from primitive entity recognition to high-level holistic scene understanding. Due to the large spatial coverage of remote sensing images, objects are often organized into multiple functional subscenes, while predicate semantics are highly dependent on the spatial configurations of the subject and object. However, existing scene graph generation (SGG) methods typically perform unconstrained global interactions over the entire panoramic scene, which introduces substantial noise in multisubscene scenarios, and they rarely exploit the relative spatial structure underlying subject–object relations. To address these issues, we propose the spatially anisotropic reasoning network (SARNet) for RS-SGG. First, we design an anisotropic elliptical influence propagation (AEIP) module that models an adaptive spatial influence region for each object using anisotropic Gaussian ellipses derived from oriented bounding boxes (OBBs), focusing on instance-centered interactions while suppressing noisy long-range relations. Second, we introduce a relative spatial configuration-guided attention (RSCA) module that incorporates the relative geometric configuration between each relation and its corresponding subject–object pair, and reallocates channel-wise attention to enhance spatially discriminative features for predicate prediction. Extensive experiments on the STAR and AUG benchmarks demonstrate that the proposed method achieves state-of-the-art performances on RS-SGG tasks, highlighting the importance of spatially structured reasoning that jointly captures object-centered functional context and relative spatial configurations in large-area remote sensing images. The code is publicly available at https://github.com/Bamboo0216/SARNet