Snow and Ice Detection Using UAV Remote Sensing Imagery

As extreme weather events intensify under global climate change, the snow and ice has become a critical factor affecting infrastructure, transportation, and public safety. To achieve accurate and timely distribution of snow and ice, a comprehensive machine learning -based approach is proposed in this paper. First, overlapping UAV images are stitched using a high-precision registration framework called Convolutional and Graph Neural Network (CGNet), which integrates convolutional neural networks (CNN) and graph neural networks (GNN). The stitched images are then used to construct a high-quality annotated dataset for deep learning models training. Three semantic segmentation models-Fully Convolutional Network (FCN), High-Resolution Network (HRNet), and Dynamic Graph Convolutional Network (DGCNet)-are evaluated for snow and ice detection. Experimental results show that HRNet performs best in overall segmentation accuracy, while DGCNet achieves superior results in ice region detection. The proposed method demonstrates strong potential for large-scale, high-resolution monitoring of snow and ice.