<p>For the wind energy sector, rapid and accurate online detection of surface damage on wind turbines is essential to minimize downtime and prevent catastrophic failures. Using drones to photograph wind turbines and detecting damage to the surface of wind turbines using computer vision techniques is an efficient solution. However, the sections in the drone-captured images that show signs of damage are usually quite small, but they are filled with numerous mountains, skies and other elements unrelated to wind turbines. It is difficult for traditional visual detection methods to accurately detect damage by relying on limited damage features. To address this challenge, we propose a detection model WTD-YOLO for surface damage detection of wind turbines. Based on PPYOLOE, we propose adjacent layer aggregated attention (ALAA) to optimize the extraction of damage features in the low-level feature maps by using the high-level feature maps as a guide to eliminate the negative impact of redundant features to reduce and enable efficient fusion of high- and low-level feature maps. We also integrate a transformer encoder block and a convolutional block attention module (CBAM) to exploit the prediction potential through the self-attention mechanism and control the model to focus on key regions in complex scenes. Experimental results on the wind turbine surface damage dataset show that our model achieves an average precision of 62.8%, outperforming PPYOLOE-L by 2.5% and YOLOv8-L by 11.7%. The inference speed reaches 82.3 FPS. The ablation studies suggest that ALAA, transformer encoder block and CBAM all play key roles. These results demonstrate the potential of our approach to solve the problem of small wind turbine damage detection and could be a viable method for wind turbine damage detection using drones in wind farms.</p>