FG-CROSS: Fine-Grained Spatial Alignment for Remote Sensing Image–Text Retrieval

Abstract

Remote sensing image–text retrieval (RSITR) requires matching not only scene semantics but also fine-grained spatial layouts. Existing cross-modal retrieval methods commonly align global image and sentence representations, which may confuse queries containing similar object vocabularies but different spatial relations. This letter proposes FG-CROSS, a fine-grained spatial alignment framework that explicitly models and aligns textual and visual relation structures. Specifically, FG-CROSS first converts descriptive text into spatial relation representations covering topology, direction, and distance cues, and then constructs object-level visual relation graphs from remote sensing images. To integrate heterogeneous information, a cross-attention fusion strategy adaptively combines global semantics, local object features, and structural relation cues. In addition, a Laplacian Spatial Alignment Loss (LSA-loss) is introduced to enforce cross-modal structural consistency by modeling neighborhood smoothness and relational dissimilarity. Extensive experiments on remote sensing benchmarks show that FG-CROSS consistently improves retrieval performance over representative RSITR baselines, demonstrating the effectiveness of explicit spatial relation alignment for fine-grained remote sensing image–text matching.