尺度和密度约束下基于YOLOv3的风电塔架遥感检测方法

Detection of wind turbine towers in remote sensing based on YOLOv3 model under scale and density constraints

风电场分布是风电投资监测预警、占地监测和清洁能源消纳能力评价的重要依据,卫星遥感技术是大范围提取风电场分布信息的有效方法。风电塔架作为风电场的遥感解译标识,其在高分影像中是一种多尺度目标,且受影像获取时间、光照条件、地表覆盖等影响导致特征差异大,遥感自动检测难度大。针对以上问题,提出一种尺度和密度约束下基于YOLOv3模型的风电塔架自动检测方法。首先,在风电场遥感特征分析基础上,确定样本构建条件,分析风电塔架目标尺度;然后,压缩YOLOv3模型特征提取网络以提高多尺度目标特征表征能力,并将目标尺寸作为先验知识输入模型;最后,基于噪声与风电塔架的密度差异,采用DBSCAN密度聚类算法抑制误检。实验结果表明,该方法在典型试验区取得的风电塔架检测准确率为96%,召回率为94%,F1为95%,效果优于Faster R-CNN和FPN等基准模型,表明本文方法对于遥感影像复杂背景下的小目标具有良好的检测效果。

The distribution of wind farms is an important basis for the monitoring and early warning of wind power investment,the analyses of land occupation,and the assessment of clean energy consumption capacity.Remote sensing technology serves as an effective method for extracting wind farm distribution on a large scale.As the remote sensing interpretation marks of wind farms,wind turbine towers are a kind of multi-scale targets in high-resolution images.However,their characteristics greatly differ due to the effects of image acquisition time,illumination conditions,and surface coverage.Therefore,it's difficult to automatically detect wind turbine towers in remote sensing images.Aiming at the above problems,this paper proposed an automatic detection method of wind turbine towers based on the YOLOv3 model,and the steps are as follows.Firstly,determine the sample construction conditions and the target scale of wind turbine towers according to the analyses of the remote sensing characteristics of a wind farm.Secondly,optimize the depth of the feature extraction network of the YOLOv3 model to improve the characterization capacity of multi-scale targets.Finally,suppress false detections using the DBSCAN density clustering algorithm according to the density difference between noise and wind turbine towers.The experimental results show that the proposed method exhibits superiority over the benchmark models such as Faster R-CNN and FPN.With a detection accuracy rate of 96%,a recall rate of 94%,and F1 of 95%,the proposed method has good effects for the detection of small targets in the remote sensing images with complex background.