Performance Evaluation of Three Deep Learning Models for River ICE Segmentation From Aerial Images

River ice monitoring is of great importance due to its impact on various environmental issues. In this study, the performance of three deep learning models, U-Net, MA-FCN, and SegNet, in segmenting UAV images of river ice was compared using the Alberta river ice dataset. The models were evaluated using mean Intersection over Union (mIoU) and Pixel Accuracy (PA). Results showed U-Net achieved the best results with an mIoU of 65.5% and a PA of 88.45%, outperforming MA-FCN and SegNet. U-Net's lightweight design and skip connections preserved spatial information, making it more effective with limited training data. MA-FCN performed slightly weaker than U-Net and SegNet models due to it having more parameters, and SegNet's lack of skip connections led to smoother edges in the final segmentation map.