Research on Long-Distance Snow Depth Measurement Method Based on Improved YOLOv8

Real-time monitoring of snow depth is crucial in meteorological forecasting, transportation management, disaster warning, and ecological protection. With the development of deep learning, the YOLO model has become widely used for automated snow depth detection due to its efficient and accurate instance segmentation. However, the standard YOLOv8-seg model still faces challenges, such as insufficient detection accuracy and blurred edges, when dealing with snow detection in complex backgrounds. To address these issues, this paper proposes a snow depth detection method based on an improved YOLOv8-seg model, named YOLOv8-AE. First, an efficient multiscale attention (EMA) module is incorporated into the C2f module to enhance feature extraction capabilities. Second, the introduction of the variable kernel convolution (AKConv) module improves the adaptability of convolutional operations, boosting the model’s performance in snow depth detection. Comparative experimental results show that, compared with the YOLOv8-seg model, the YOLOv8-AE model increases the ratio of snow depths that deviate by no more than 2 cm from the reference value, from 84% to 88% on average. The RMSE decreases by 0.74, and the MAE decreases by 0.93, significantly improving the accuracy of snow depth measurements.