Application of machine learning techniques for regional bias correction of snow water equivalent estimates in Ontario, Canada

• Published in: Hydrology and earth system sciences Vol. 24; no. 10; pp. 4887 - 4902
• Main Authors: King, Fraser, Erler, Andre R, Frey, Steven K, Fletcher, Christopher G
• Format: Journal Article
• Language: English
• Published: Katlenburg-Lindau Copernicus GmbH 14-10-2020; Copernicus Publications
• Subjects: Algorithms; Bias; Comparative analysis; Data assimilation; Data collection; Datasets; Energy budget; Equivalence; Estimates; Flood forecasting; Flood management; Floods; Ground stations; Interannual variability; Lakes; Learning algorithms; Machine learning; Methods; Moisture content; Nonlinear systems; Nonlinearity; Resource management; River discharge; Snow; Snow-water equivalent; Snowmelt; Snowpack; Spring; Statistical analysis; Statistical methods; Subtraction; Surveying; Water content; Water resource management; Water resources; Water resources management; Weather forecasting; Ontario; Canada; Ontario Canada; United States > US; North America; Quebec Canada
• ISSN: 1607-7938; 1027-5606; 1607-7938
• DOI: 10.5194/hess-24-4887-2020

Snow is a critical contributor to Ontario's water-energy budget, with impacts on water resource management and flood forecasting. Snow water equivalent (SWE) describes the amount of water stored in a snowpack and is important in deriving estimates of snowmelt. However, only a limited number of sparsely distributed snow survey sites (n=383) exist throughout Ontario. The SNOw Data Assimilation System (SNODAS) is a daily, 1 km gridded SWE product that provides uniform spatial coverage across this region; however, we show here that SWE estimates from SNODAS display a strong positive mean bias of 50 % (16 mm SWE) when compared to in situ observations from 2011 to 2018. This study evaluates multiple statistical techniques of varying complexity, including simple subtraction, linear regression and machine learning methods to bias-correct SNODAS SWE estimates using absolute mean bias and RMSE as evaluation criteria. Results show that the random forest (RF) algorithm is most effective at reducing bias in SNODAS SWE, with an absolute mean bias of 0.2 mm and RMSE of 3.64 mm when compared with in situ observations. Other methods, such as mean bias subtraction and linear regression, are somewhat effective at bias reduction; however, only the RF method captures the nonlinearity in the bias and its interannual variability. Applying the RF model to the full spatio-temporal domain shows that the SWE bias is largest before 2015, during the spring melt period, north of 44.5∘ N and east (downwind) of the Great Lakes. As an independent validation, we also compare estimated snowmelt volumes with observed hydrographs and demonstrate that uncorrected SNODAS SWE is associated with unrealistically large volumes at the time of the spring freshet, while bias-corrected SWE values are highly consistent with observed discharge volumes.