Abstract
Plateau zokor mounds, created by the burrowing activity of Plateau zokor, cause significant damage to crops, grasslands, and infrastructure, particularly in the alpine meadows of the Tibetan Plateau. Traditional field surveys are inefficient and labor-intensive, limiting the ability to conduct large-scale monitoring. Accurate detection of zokor mounds is essential for effective rodent control and sustainable grassland management. This study introduces VGG-Dice-PSA UNet(VDP_UNet), an enhanced deep learning model designed to segment zokor mounds from UAV imagery captured at 30 m. Based on the UNet architecture, VGG16 is used to replace the original UNet backbone, enabling the model to capture global contextual information and enhance feature extraction in complex backgrounds. Additionally, a Polarized Self-Attention (PSA) module is integrated into the feature fusion stage following the encoder-decoder skip connections to better capture fine-grained semantic features related to zokor mounds. To reduce overfitting and address class imbalance, Dice Loss is introduced during training. VDP_UNet was trained and evaluated on a custom high-resolution zokor mound dataset. It achieved an IoU of 51.99%, MIoU of 75.63%, mean Pixel Accuracy of 82.66%, Precision of 71.44%, FPS of 42.13 f/s, Accuracy of 99.27%, and an F1-score of 68.41%, outperforming recent deep learning models. Experimental results indicate that the proposed VDP_UNet model efficiently segments zokor mounds in alpine meadows, markedly improving the extraction of mound features from UAV images. Furthermore, this study establishes a practical foundation for estimating mound areas in real sample plots and provides solid technical support for rodent control and the sustainable development of alpine ecosystems.