Snowfall has important significance in water resources management and disaster prevention worldwide. Accurate prediction of both mean and extreme snowfall is challenging because of multiple controlling mechanisms at different spatial and temporal scales. By using a 65 years long in-situ snowfall observation, we evaluated seven different machine learning algorithms for predicting monthly snowfall in the Lower Peninsula of Michigan (LPM). The Bayesian Additive Regression Trees (BART) demonstrates the best fitting (R2 = 0.88) and out-of-sample prediction skills (R2 = 0.58) for the monthly mean snowfall followed by the Random Forest model. The BART also demonstrate strong predictive skills for seasonal and the extreme monthly snowfall. Both machine learning models also demonstrate signals of key physical processes controlling the snowfall including topography, local/regional environmental factors, and teleconnections. Particularly, models with the non-parametric framework can incorporate signals from multiple scales and nonlinear responses from the snowfall to environmental factors and that substantially improved the model prediction skills. The multiscale machine learning approach provides a reliable and computationally efficient alternative approach to predict/forecast weather and climate and has potential to be applied to other extreme weather prediction scenarios.