航空学报 > 2025, Vol. 46 Issue (23): 631955-631955   doi: 10.7527/S1000-6893.2025.31955

干扰环境下无人机多源感知专栏

基于强化学习数据合成的典型遥感目标检测

范天麒1, 邹征夏2(), 史振威2   

  1. 1.北京航空航天大学 国家卓越工程师学院,北京 100191
    2.北京航空航天大学 宇航学院,北京 100191
  • 收稿日期:2025-03-10 修回日期:2025-03-16 接受日期:2025-05-15 出版日期:2025-06-09 发布日期:2025-05-30
  • 通讯作者: 邹征夏 E-mail:zhengxiazou@buaa.edu.cn
  • 基金资助:
    国家自然科学基金(62125102)

Typical remote sensing target detection with data synthesis based on reinforcement learning

Tianqi FAN1, Zhengxia ZOU2(), Zhenwei SHI2   

  1. 1.National Graduate College for Elite Engineers,Beihang University,Beijing 100191,China
    2.School of Astronautics,Beihang University,Beijing 100191,China
  • Received:2025-03-10 Revised:2025-03-16 Accepted:2025-05-15 Online:2025-06-09 Published:2025-05-30
  • Contact: Zhengxia ZOU E-mail:zhengxiazou@buaa.edu.cn
  • Supported by:
    National Natural Science Foundation of China(62125102)

摘要:

遥感目标检测,特别是对车辆、飞机等遥感目标的检测,在交通监测、军事侦察等方面具有重要的应用价值,对相关技术的进步也起到了积极的推动作用。然而,遥感目标检测仍面临诸多挑战,包括遥感图像标注成本高,以及复杂天气条件会对目标检测性能产生干扰。为解决上述问题,提出了一种融合强化学习与可控图像渲染的遥感图像合成方法,将虚实混合的遥感图像生成任务转化为基于强化学习的场景参数搜索问题。同时,引入天气干扰因素,通过模拟云层、沙尘和雾气等复杂天气条件对训练数据进行增强,提升数据在复杂环境下的适应性和真实性。实验结果表明,相较于传统方法,所提方法在多项评价指标上均取得显著提升,最高性能提升达4.2%。这验证了所提方法的有效性,显示了合成遥感图像及其复杂天气效果在提升目标检测模型精度方面的潜力,为遥感目标检测技术的发展提供了有力支持。

关键词: 目标检测, 遥感图像, 图像合成, 强化学习, 深度学习

Abstract:

Remote sensing detection of targets such as vehicles and aircraft is essential for traffic monitoring and military reconnaissance, and it continues to drive advances in related technologies. However, remote sensing target detection still faces numerous challenges, including the high cost of image annotation and the interference of complex weather conditions on detection performance. To address these issues, a novel image synthesis method that integrates reinforcement learning with controllable image rendering is proposed, transforming the generation of mixed real and synthetic remote sensing data into a reinforcement learning-based scene parameter search problem. Additionally, weather disturbance factors are introduced by simulating clouds, dust, and fog to augment the training data, enhancing its adaptability and realism in complex environments. Experimental results show that, compared with traditional methods, the proposed method achieves significant improvements across multiple evaluation metrics, with a maximum performance gain of 4.2%. This verifies the effectiveness of the proposed method and highlights the potential of synthetic remote sensing images and complex weather effects in improving target detection model accuracy, thereby providing strong support for the advancement of remote sensing target detection technology.

Key words: target detection, remote sensing image, image synthesis, reinforcement learning, deep learning

中图分类号: