随着遥感影像的可获取性逐渐提高，其应用日益广泛，其中，目标探测技术在国防军事等领域具有重要意义。强化学习作为机器学习的重要分支，因其具有自学习和自适应的特点而得到关注。为实现端到端的遥感影像目标探测，本研究基于深度学习卷积神经网络(CNN)与强化学习策略梯度网络(PGN)模型，提出了遥感影像目标探测的新方法。以RGB遥感影像为研究对象，设计可自动调节大小的探测窗口，实现在大场景遥感影像上快速准确识别并定位目标地物的功能。结果表明，该模型训练得到的智能体能够在大场景下对小目标物体进行准确定位。同时，本研究验证了深度强化学习解决遥感目标探测任务的可行性，为利用深度强化学习解决遥感领域相关研究与应用问题提供了可能。

Currently, remote sensing imagery has been increasingly available and its application fields are becoming wider. In particular, object detection is of great significance in defense and military fields. Reinforcement learning, an important branch of machine learning, is getting more and more attention due to its advantage of self-learning and self-adaptation. In this work, we propose an end-to-end model aiming to detect target objects in remote sensing imagery, combining Convolution Neural Network (CNN) in deep learning and Policy Gradient Network (PGN) in reinforcement learning. The detection window in the model was designed to automatically adjust the size, extract useful deep features of selected areas and finally locate the target objects efficiently and accurately. The result demonstrates the strong ability of the well-trained intelligent agent to pinpoint the target objects in large-scale scenes, and verified the feasibility of deep reinforcement learning to solve the object detection tasks. Our work linked traditional remote sensing imagery with cutting-edge reinforcement learning technology, and the model could be generalized to solve more remote sensing related tasks.