Abstract: Accurate identification of bare patches caused by Ochotona curzoniae disturbance is fundamental for scientifically assessing the damage level. Traditional methods for recognizing and calculating the area of bare patches are often computationally complex and inefficient. Here, we proposed a wavelet-enhanced U-shaped convolutional neural network (W-UNet) segmentation method based on deep learning for unmanned aerial vehicle (UAV) imagery segmentation, which was based on the U-shaped convolutional neural network (UNet) architecture and used the 16-layer Visual Geometry Group network (VGG16) as the backbone. We introduced the coordinate attention mecha-nism (CA) in the skip connection section to enhance the spatial localization of target regions, and wavelet transform convolution (WTConv) during the encoding stage to improve high-frequency information extraction and the recovery of fine-grained features. Additionally, we employed a composite loss function combining Focal Loss and Dice Loss to effectively address the class imbalance issues. The results showed that the proposed method achieved a mean intersection over union (MIoU) of 81.2%, mean pixel accuracy (MPA) of 89.4%, and overall accuracy (ACC) of 95.8%, significantly outperforming the conventional UNet-Vgg model. This study would provide a robust technical framework for the efficient and accurate monitoring of bare patches induced by O. curzoniae infestation.

Key words: pika-infested land, bare patch, unmanned aerial vehicle, deep learning, W-UNet