Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TinyRS-R1: Compact Multimodal Language Model for Remote Sensing

Aybora Koksal, A. Aydin Alatan

TL;DR

The paper tackles edge-deployed remote sensing by introducing TinyRS, a 2B RS-focused visual-language model, and its reasoning-augmented variant TinyRS-R1.Built on Qwen2-VL-2B, the models undergo a four-stage pipeline: RS pretraining, visual instruction tuning, CoT-based fine-tuning, and GRPO-based alignment, with a dedicated VHM-Instruct-Think reasoning dataset and a GRPO reward framework.TinyRS-R1 matches or surpasses recent 7B RS models across classification, grounding, and reasoning benchmarks while using a fraction of the memory and latency; TinyRS excels in low-latency VQA, illustrating a trade-off between reasoning depth and speed.Ablation studies highlight the importance of SFT and CoT, the grounding gains from GRPO, and the benefits of dataset balancing, while also showing that 2B models are better aligned to available RS data than 7B variants; future work explores conditional routing via Mixture-of-Experts.

Abstract

Remote-sensing applications often run on edge hardware that cannot host today's 7B-parameter multimodal language models. This paper introduces TinyRS, the first 2B-parameter multimodal small language model (MSLM) optimized for remote sensing tasks, and TinyRS-R1, its reasoning-augmented variant. Built upon Qwen2-VL-2B, TinyRS is trained through a four-stage pipeline: pre-training on million satellite images, instruction tuning on visual instruction examples, fine-tuning with Chain-of-Thought (CoT) annotations from the proposed reasoning dataset, and alignment via Group Relative Policy Optimization (GRPO). TinyRS-R1 achieves or surpasses the performance of recent 7B-parameter remote sensing models across classification, VQA, visual grounding, and open-ended question answering-while requiring just one-third of the memory and latency. Our analysis shows that CoT reasoning substantially benefits spatial grounding and scene understanding, while the non-reasoning TinyRS excels in concise, latency-sensitive VQA tasks. TinyRS-R1 represents the first domain-specialized MSLM with GRPO-aligned CoT reasoning for general-purpose remote sensing.

TinyRS-R1: Compact Multimodal Language Model for Remote Sensing

TL;DR

The paper tackles edge-deployed remote sensing by introducing TinyRS, a 2B RS-focused visual-language model, and its reasoning-augmented variant TinyRS-R1.Built on Qwen2-VL-2B, the models undergo a four-stage pipeline: RS pretraining, visual instruction tuning, CoT-based fine-tuning, and GRPO-based alignment, with a dedicated VHM-Instruct-Think reasoning dataset and a GRPO reward framework.TinyRS-R1 matches or surpasses recent 7B RS models across classification, grounding, and reasoning benchmarks while using a fraction of the memory and latency; TinyRS excels in low-latency VQA, illustrating a trade-off between reasoning depth and speed.Ablation studies highlight the importance of SFT and CoT, the grounding gains from GRPO, and the benefits of dataset balancing, while also showing that 2B models are better aligned to available RS data than 7B variants; future work explores conditional routing via Mixture-of-Experts.

Abstract

Remote-sensing applications often run on edge hardware that cannot host today's 7B-parameter multimodal language models. This paper introduces TinyRS, the first 2B-parameter multimodal small language model (MSLM) optimized for remote sensing tasks, and TinyRS-R1, its reasoning-augmented variant. Built upon Qwen2-VL-2B, TinyRS is trained through a four-stage pipeline: pre-training on million satellite images, instruction tuning on visual instruction examples, fine-tuning with Chain-of-Thought (CoT) annotations from the proposed reasoning dataset, and alignment via Group Relative Policy Optimization (GRPO). TinyRS-R1 achieves or surpasses the performance of recent 7B-parameter remote sensing models across classification, VQA, visual grounding, and open-ended question answering-while requiring just one-third of the memory and latency. Our analysis shows that CoT reasoning substantially benefits spatial grounding and scene understanding, while the non-reasoning TinyRS excels in concise, latency-sensitive VQA tasks. TinyRS-R1 represents the first domain-specialized MSLM with GRPO-aligned CoT reasoning for general-purpose remote sensing.
Paper Structure (8 sections, 2 figures, 6 tables)

This paper contains 8 sections, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Dataset
  3. Method
  4. Experiments
  5. Settings
  6. Results
  7. Ablation Study
  8. Conclusion

Figures (2)

  • Figure 1: Training pipeline of TinyRS and TinyRS-R1 involves four stages: VHM pretraining, instruction tuning with VHM-Instruct, CoT fine-tuning via VHM-Instruct-Think, and GRPO-based reward alignment. System prompts for CoT generation and answer grading are shown. TinyRS-R1 includes reasoning and reward feedback; TinyRS is optimized for concise tasks.
  • Figure 2: Examples from the VHM-Instruct-Think dataset, showing satellite images with associated questions, model-generated Chain-of-Thought reasoning, and concise answers. These samples demonstrate the format used for reasoning-augmented supervision in TinyRS-R1.