Change detection in snow water equivalent using a Siamese UNet model

Snow water equivalent (SWE), an important physical property of snow, has been used to demonstrate the potential impacts of climate variables on snow regime during winter. This information is crucial for informed decision-making on critical infrastructure, such as hydropower, transportation, and irrigation systems in the cold regions. However, detecting daily changes in SWE is fraught with challenges due to its internal variability, which may not be associated with underlying climate variables. In this study, changes in SWE across Canada’s cold regions are presented at high temporal resolution using GlobSnow’s version 3.0 SWE data. A Siamese UNet (Si-UNet) model is trained to detect daily changes. The model is guided using the structural similarity (SSIM) index. The architecture of the model is modified to introduce the SSIM index in the last layer. The scalar output from structural similarity computation is transferred to the contrastive loss function, enabling the model to learn the structural differences between a pair of SWE maps. The model detects daily changes in SWE, producing an overall accuracy of 99% on independent validation data. The model predicts snow daily melt-off events to increase significantly in April after 2008.