FUSAR-R1: New Reasoning Model for SAR Image Interpretation

Interpreting Synthetic Aperture Radar (SAR) images presents unique challenges due to their complex features, speckle noise, and target-background coupling, which arise from coherent imaging mechanisms. While large-scale vision-language models have shown promise in remote sensing, existing SAR models lack the step-by-step analysis, logical judgment, and self-correction capabilities of human experts, making reliable interpretation difficult in complex scenarios. To address these limitations, the FUSAR-R1 model has been developed. This large-scale reasoning model for SAR image interpretation first constructs explicit chain-of-thought reasoning data by simulating human expert interpretation processes. This data then guides instruction learning, imbuing the model with foundational reasoning abilities. Subsequently, a reinforcement learning strategy is integrated to optimize the model's outputs based on inference results, enabling self-correction and more dependable reasoning. Experimental results confirm that FUSAR-R1 consistently surpasses current multimodal large-scale models across various SAR interpretation tasks, including target detection, counting, classification, and land-cover recognition.