Reinforcement learning-driven object detection method for degraded remote sensing images

Satellite remote sensing image object detection constitutes a pivotal technique for Earth observation and intelligent interpretation. However, most existing research has concentrated on ideal imaging conditions, and the resulting detection performance remains notably insufficient under complex degradations, such as adverse weather, atmospheric turbulence, and noise interference. To address this limitation, a reinforcement learning-based adaptive object detection methodology is proposed for degraded remote sensing images. This methodology achieves robust detection in complex scenarios by dynamically orchestrating image preprocessing operators. The core principle is to optimize object detection performance by leveraging reinforcement learning's decision-making capability to adaptively and iteratively select and compose preprocessing operations, including denoising, deblurring, and contrast enhancement, thereby improving both remote sensing imagery quality and detection precision. Experiments on degraded scenarios constructed from the DIOR and DOTA satellite remote sensing datasets with the YOLO11m-OBB detector demonstrate that the proposed method achieves superior performance in all cases. On DIOR, the proposed method achieves mAP₅₀ improvements of 11.1% and 2.5% over Raw-Syn (trained on pristine data, validated on degraded data) and Syn-Syn (trained and validated on degraded data) baselines, respectively, achieving a final mAP₅₀ of 80.8%. On DOTA, mAP₅₀ is improved by 7.2% and 2.8% over the same baselines, reaching 76.6%. Furthermore, the quality of processed remote sensing imagery is significantly enhanced (PSNR > 25 dB), substantiating the efficacy and applicability of the proposed approach in challenging environments.