GeoVLM-R1: Reinforcement Fine-Tuning for Improved Remote Sensing Reasoning

TL;DR

GeoVLM-R1 tackles shallow reasoning in Earth Observation vision–language models by applying a two-stage training pipeline: supervised fine-tuning to bootstrap EO-domain knowledge, followed by GRPO-based reinforcement learning with task-aware, dual-format rewards to produce structured, interpretable reasoning traces before final answers. The method introduces a novel dual-objective reward design (format and task-aware accuracy) and a GRPO objective that uses intra-group reward differences to stabilize updates. Across 28 EO benchmarks, GeoVLM-R1 achieves consistent improvements over state-of-the-art VLMs in tasks ranging from classification and detection to region captioning, grounding, VQA, and temporal change analysis, including notable gains on BigEarthNet and xBD datasets. The approach demonstrates that task-specific rewards coupled with group-relative policy optimization yield robust, scalable, and interpretable EO reasoning with practical implications for multi-task remote sensing analytics.

Abstract

Recent advances in reinforcement learning (RL) have delivered strong reasoning capabilities in natural image domains, yet their potential for Earth Observation (EO) remains largely unexplored. EO tasks introduce unique challenges, spanning referred object detection, image or region captioning, change detection, grounding, and temporal analysis, that demand task aware reasoning. We propose a novel post training framework that incorporates task aware rewards to enable effective adaptation of reasoning based RL models to diverse EO tasks. This training strategy enhances reasoning capabilities for remote sensing images, stabilizes optimization, and improves robustness. Extensive experiments across multiple EO benchmarks show consistent performance gains over state of the art generic and specialized vision language models. Code and models will be released publicly at https://mustansarfiaz.github.io/GeoVLM-R1/ .

Figures (8)

• Figure 1: Comparison of recent generic and specialized VLMs over diverse EO tasks. GeoVLM-R1 shows favorable improvements across classification, detection, and captioning tasks.
• Figure 2: Illustration of the overall proposed training paradigm for GeoVLM-R1. The model is first initialized via supervised fine-tuning using diverse earth observation tasks. It is then successively optimized using GRPO-based reinforcement learning (RL) for each task. The GeoVLM-R1 processes queries and outputs a structured format that comprises an interpretable reasoning trace (<think>...</think>) and a final prediction (<answer>...</answer>).
• Figure 3: Overall pipeline of GeoVLM-R1 policy update mechanism (left). During fine-tuning, the GRPO module generates multiple candidate responses. These responses are evaluated, and each is assigned a distinct reward equipped with our reward mechanism. In particular, our reward mechanism comprises (i) a format reward to enforce structural compliance and (ii) a task-aware accuracy reward to ensure accuracy compliance. We present a few examples showcasing GeoVLM-R1 using a unique task-aware accuracy reward function, resulting in better performance (right).
• Figure 4: Ablation over change detection MUDS dataset shows that GeoVLM-R1 with HSLR performs better. Whereas for image captioning task, GeoVLM-R1 with LR reward performs favorably.
• Figure 5: Ablation using various reward functions for the classification task. Our method with the recall reward is more effective than other models.