RSGround-R1: Rethinking Remote Sensing Visual Grounding through Spatial Reasoning

TL;DR

RSGround-R1 tackles remote sensing visual grounding by reframing it as a spatial reasoning problem. It introduces a two-stage post-training pipeline: first, Chain-of-Thought SFT (CoT-SFT) to seed explicit position-aware reasoning, then Reinforcement Fine-Tuning (RFT) using GRPO with a continuous positional reward and a spatial consistency optimization to stabilize rollouts. The approach achieves state-of-the-art results on multiple RS grounding benchmarks and demonstrates strong out-of-domain generalization, while requiring only a fraction of training data. This work advances geospatial understanding by enabling more robust, interpretable spatial localization in large-scale aerial imagery. The combination of explicit reasoning priors and dense spatial feedback offers practical impact for urban planning, disaster response, and environmental monitoring.

Abstract

Remote Sensing Visual Grounding (RSVG) aims to localize target objects in large-scale aerial imagery based on natural language descriptions. Owing to the vast spatial scale and high semantic ambiguity of remote sensing scenes, these descriptions often rely heavily on positional cues, posing unique challenges for Multimodal Large Language Models (MLLMs) in spatial reasoning. To leverage this unique feature, we propose a reasoning-guided, position-aware post-training framework, dubbed \textbf{RSGround-R1}, to progressively enhance spatial understanding. Specifically, we first introduce Chain-of-Thought Supervised Fine-Tuning (CoT-SFT) using synthetically generated RSVG reasoning data to establish explicit position awareness. Reinforcement Fine-Tuning (RFT) is then applied, augmented by our newly designed positional reward that provides continuous and distance-aware guidance toward accurate localization. Moreover, to mitigate incoherent localization behaviors across rollouts, we introduce a spatial consistency guided optimization scheme that dynamically adjusts policy updates based on their spatial coherence, ensuring stable and robust convergence. Extensive experiments on RSVG benchmarks demonstrate superior performance and generalization of our model.

Figures (6)

• Figure 1: (a) Word cloud comparison between natural VG datasets RefCOCO+ yu2016modeling, which relies exclusively on semantic attributes, and RSVG datasets zhan2023rsvgliu2024rotatedli2024vrsbench. Phrase size corresponds to its frequency. (b) Examples from DIOR-RSVG dataset zhan2023rsvg with positional and relational phrases.
• Figure 2: (i) CoT-SFT for position awareness. With CoT-SFT, the model can provide thinking process with clearer, position-aware spatial reasoning. (ii) Positional reward for spatial guidance. Positional reward increases as prediction approaches the target, whereas IoU reward remains zero. (iii) Illustration of spatial inconsistency. Samples representing spatially inconsistent/consistent rollouts.
• Figure 3: Overview of RSGround-R1 pipeline. Our framework has two consecutive stages. CoT-SFT (Stage 1): we curate high-quality, position-aware CoT data to finetune the base model and instill explicit positional understanding. RFT (Stage 2): after initializing the policy and reference model with the CoT-SFT finetuned based model, we further incentivize spatial reasoning using RFT by GRPO. Along with format and IoU rewards, we introduce a positional reward for progressive localization. To mitigate spatial inconsistent rollouts, we apply a spatial consistency guided optimization that reweights the per token GRPO loss $\mathcal{L}_{\text{GRPO}}$ to yield $\mathcal{L}_{\text{SC}}$, promoting more stable and contextually aligned grounding behavior.
• Figure 4: Qualitative comparison. RSGround-R1 produces more plausible reasoning trajectories and more accurate bounding box predictions compared to the baseline Qwen2.5-VL-3B w/ GRPO. More visualizations can be found in the Appendices.
• Figure 5: Spatial consistency guided optimization: reward std reduction vs. baseline on DIOR-RSVG dataset.