Machine Learning-Based Estimation of High-Resolution Snow Depth in Alaska Using Passive Microwave Remote Sensing Data

Abstract

Snow depth (SD) knowledge is significant in many applications related to hydrology, climate, and disaster management. Many SD models are developed using multifrequency spaceborne passive microwave (PMW) brightness temperature (Tb) observations because of their sensitivity to SD. The sensitivity of Tb to SD is affected by snow metamorphism, which constrains the utility of several empirical and conceptual models for estimating SD. For the first time, extremely randomized trees (ERT), a machine learning algorithm, which is less susceptible to data noise, is used in this study for estimating SD at high resolution (1 km × 1 km) for Alaska. Different ERT SD models (i.e., Alaska wide model, zonal model) are developed using Advanced Microwave Scanning Radiometer-2 data and auxiliary datasets for various Alaska regions during 2012–2021. These models are evaluated using three different cross-validations (i.e., sample, spatial, and temporal). Further, ERT models' predictive power assessment is performed using independent spatial, temporal datasets. The results indicate that: 1) inclusion of auxiliary parameters improves the accuracy of ERT SD estimates; 2) there is no substantial difference between the zonal and Alaska wide ERT model estimates; 3) when SD $ > $30 cm, the ERT models have outperformed the AMSR-2 product, the GlobSnow product, and the Chang model; 4) the mean absolute error in SD estimates increases with a decrease in latitude, an increase in elevation, and from early winter to late winter across Alaska. Overall, this study shows that the ERT SD model has good potential for improving moderate to deep SD estimates.

Keywords

• Machine learning
• passive microwave remote sensing
• snow depth (SD)