Abstract
Subsurface drainage is a key loss pathway for water and nutrients from agricultural land in Eastern Canada. Winter is presently a dominant period of subsurface drainage and nutrient loss in cold climates. Under climate change, however, future winter drainage patterns may change significantly due to reductions of snow cover and soil freezing. This study evaluated the performance of the RZ-SHAW model and four machine-learning (ML) models in simulating winter subsurface drainage for five sites in Eastern Canada. The calibrated/trained RZ-SHAW and ML models were then applied to predicted future climate (high emission scenario: RCP8.5) spanning from 1950 to 2100 to comprehend the potential alteration in winter drainage patterns under global warming. Among ML models, the Cubist and SVM-RBF models emerged as the most accurate, offering competing short-term simulation capabilities compared to the RZ-SHAW modelwith lower computational demand. Simulation by both the RZ-SHAW and ML models predict a significant increase in winter drainage volume and frequency by the end of the 21st century (1950-2005 vs. 2070-2100) (RZ-SHAW: 243 mm to 328 mm (+35%); 75.5 days to 102.9 days (+45%), ML models: 250 mm to 425 mm (+70%); 121.9 days to 129.2 days (+8%)). RZ-SHAW simulated a shift towards a more evenly spread drainage pattern throughout the winter months from baseline to the end of the century. This shift was driven by the simulated shorter snow coverage periods, advancement of snowmelt timing, and fewer days of freezing soil. Thus, the timing of peak and trough winter drainage is expected to reverse, with February becoming the peak month and April the lowest by century's end.