RS_DeepReason: LLM-Driven Deep Reasoning for Multigranularity Remote Sensing Scene Interpretation

Vision-Language Models (VLMs) have achieved remarkable results in remote sensing scene interpretation. However, existing models primarily rely on a single-step reasoning paradigm, which suffers fromIncomplete PerceptionandGranularity Limitationwhen confronting complex tasks requiring comprehensive, multi-granularity visual contexts. To overcome these bottlenecks, we propose DeepRS, a training-free, LLM-driven Deep Reasoning Framework. This framework leverages the logical planning capabilities of Large Language Models (LLMs) to restructure the reasoning process into a hierarchical, iterative Reason-Observe-Re-reason cycle. Specifically, we construct a dynamic reasoning tree where the LLM recursively decomposes the problem into fine-grained queries, actively guiding the perception module to mine visual clues. Subsequently, a multistage reflection mechanism consolidates these multi-level contexts to derive the final conclusion. Extensive experiments demonstrate that DeepRS significantly outperforms both single-VLM baselines and collaborative baselines in accuracy and robustness. Furthermore, our framework generates clear, traceable reasoning paths, substantially enhancing interpretability in complex remote sensing scenarios.