A Machine Learning Approach for Snow Depth Estimation From Temperature Sensors

Abstract

Snow is an effective, natural insulator and the differences in its internal temperature dynamics compared to soil and atmosphere allow for estimation of snow depth from snow temperature measurements. We use temperature sensor profiles to estimate snow depth for monitoring multiple winter seasons in a remote 1.3 km² (130 ha) forested watershed in the Adirondack Mountains, New York, United States. Vertical temperature sensor profiles were installed in a grid pattern in 2019 to monitor snow energy state and soil microclimate. Each profile consists of iButton temperature sensors enclosed in PVC pipe at 20 cm vertical spacing, of which eight profiles were paired with trail cameras and snow stakes for daily snow depth estimation starting in November 2021. An additional four temperature profiles with sensors exposed directly to the snow at 10 cm vertical sensor spacing were added in November 2022. We use photographs paired with temperature profiles to train random forest (RF) machine learning models to estimate snow depth from snow temperature profiles and landscape properties. Comparison of our RF model predictions versus camera-derived snow depths shows that we can accurately infer snow depth with a root mean squared error (RMSE) between 1.8 and 6.5 cm, which is lower than or comparable to existing methods. Our random forest method demonstrated effectiveness in an area with a shallow snowpack and frequent midwinter melt events, and showed little sensitivity to sensor mounting method, vertical sensor spacing, or time of day.