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.