DAWindNet: A Deep CNN for Domain Adaptively Extracting Wind Turbine From Satellite HRI
Zhe Zeng, Qing Wei, Wenxia Tan, Yongtian Shen, Zhongheng Wu · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2025
Wind turbines are vital equipment for converting wind energy into electrical energy. Precise detection of wind turbines in high-resolution images is of great importance for building accurate wind turbine databases. However, their distinctive white appearance poses challenges for detection in snowy domain images. To address this problem, a domain adaptive deep convolutional neural network model named DAWindNet, is proposed, aiming at cross-domain extraction of wind turbines in snow background images. First, in the image-level module, the model learns discrepancies between images from different domains, alters the image style, and achieves data alignment for complete images. Second, in the instance-level module, the model focus shifts to the structural information of wind turbine targets, further refining cross-domain data alignment. Perception loss and domain difference loss are designed to preserve semantic consistency across domains and mitigate domain offset phenomena of the target. Finally, bidirectional feature pyramid network and an attention mechanism are incorporated to enhance the network's ability to extract wind power features and achieve higher recognition rates. Experimental results on datasets representing ordinary, bare soil, and snowy domains validate that DAWindNet achieves the satisfactory performance, with a recall rate of 63.8% and an average precision (AP) of 67.3% when transitioning from ordinary to snow background domains, and a recall rate of 65.1% and an AP of 66.1% for the transition between bare soil and snow background domains. These results demonstrate the effectiveness of the proposed modifications for the cross-domain extraction of wind turbines.