Visual-Grounded Hierarchical Perception for Remote Sensing Visual Question Answering

Remote sensing visual question answering (RSVQA) leverages vision and natural language for intelligent Earth Observation. While multi-modal large language models (LLMs) demonstrate broad potential in RS multi-task frameworks, their heavy reliance on textual priors from extensive conversational pre-training and neglect of the hierarchical nature of RS scenes limit their performance. To address these limitations, this paper proposes a visual-grounded hierarchical perception (VGHP) framework that activates a pre-trained LLMs open-world knowledge under the joint guidance of visual evidence and linguistic semantics for reliable answer generation. Specifically, a visual-grounded language encoder (VGLE) dynamically binds question semantics to the visual context via cross-modal attention at every encoding layer, effectively mitigating the language prior bias by enforcing vision-conditioned reasoning. Concurrently, a hierarchical vision perceiver (HVP) is designed to capture multi-granular visual information, routing local details and global semantics across feature hierarchies in a vision tower, thereby addressing multi-scale perception challenges in RS images. Extensive experiments on three RSVQA benchmarks showcase the superiority of the proposed VGHP, achieving state-of-the-art accuracy while maintaining computational efficiency.