Aircraft detection in remote-sensing images is a fundamental task in civil and military applications. Deep learning techniques to achieve end-to-end object detection have attracted the attention of the Earth observation community. One of the primary factors behind the success of deep learning techniques is the utilized data. Several previous studies focused on designing the network infrastructure. Instead, this study pays more attention to the data. With the increasing number of available public datasets, whether directly employing a large number of instances with great variation will lead to a good performance has become a research topic. The ways in which these object instances are collected differ greatly. For example, the image sizes, object sizes in the training images, and geospatial resolution are varied. Therefore, herein, the factors influencing the detection performance, such as the object size, ground sampling distance, and Google zoom view, are investigated. A you-only-look-once-v3-based detection process is proposed for automatic aircraft detection. A nonmaximum suppression algorithm strategy is applied to filter unreliable and redundant bounding boxes detected in the overlapping image blocks. The model generalization ability under different training data combinations is evaluated in several challenging cases. The results prove that more variety of training instances from a greater variety of zoom levels will result in more false alarms. Instead, more variety in the object sizes under a constant zoom level is welcome. A large range of aircraft sizes (i.e., 7 to 77 m in length in this study) can be detected, with a promising F1 score of 0.98.

中文翻译：

---

使用基于YOLOv3的过程在超高分辨率卫星图像中自动检测飞机

在民用和军事应用中，遥感图像中的飞机检测是一项基本任务。用于实现端到端对象检测的深度学习技术吸引了地球观测界的关注。深度学习技术成功的主要因素之一是所利用的数据。先前的一些研究集中在设计网络基础结构上。相反，该研究更加关注数据。随着可用公共数据集数量的增加，是否直接采用大量变化较大的实例将导致良好的性能已成为研究的课题。这些对象实例的收集方式差异很大。例如，图像大小，训练图像中的对象大小以及地理空间分辨率都不同。因此，在这里，研究了影响检测性能的因素，例如物体尺寸，地面采样距离和Google缩放视图。针对自动飞机检测，提出了基于一次仅查看v3的检测过程。应用非最大抑制算法策略来过滤在重叠图像块中检测到的不可靠和多余的边界框。在几种具有挑战性的情况下，评估了不同训练数据组合下的模型泛化能力。结果证明，来自更大缩放级别的更多训练实例将导致更多错误警报。取而代之的是，在恒定的缩放级别下，欢迎使用更多的对象大小。可以检测到各种尺寸的飞机（即，在本研究中，飞机的长度为7至77 m），F1分数有望达到0.98。