Snow accumulation in mountainous watersheds plays a paramount role in the hydrological cycle and environmental stability. In the current research, three different snow modeling approaches including conceptual models, Random Forest (RF) and their nested (Nested-RF) model have been implemented to project climate change. Also, three distinct calibration scenarios employing Snow Cover Area (SCA) and Snow Water Equivalent (SWE) have been used. These scenarios involve using SCA and SWE data individually or combined to calibrate both conceptual snowpack models and machine learning methods. Considering comparative analysis in the mountainous and complex topographic area of Tashk-Bakhtegan watershed as a case study, the Guo model (calibrated using both SCA and SWE data) achieves the highest prediction accuracy in predicting SCA/SWE using conceptual method. Considering various statistics, RF using the same calibration scenario demonstrates high accuracy in estimating SWE and shows promising results in detecting SCA. Considering the feature importance analysis and statistics of the RF and Nested-RF models indicate that the feeding conceptual outputs as extra information into the Nested-RF had minimal influence and even slightly reduced accuracy. Future projections using RF indicate significant reductions in mean annually SCA/SWE, especially under SSP8.5, with a 28.4% decrease in SCA and a reduction in SWE to 13.62 mm from a historical baseline of 19.4 mm. Lower emission scenarios (SSP2.6 and SSP4.5) show less drastic declines. While minor differences were observed between the best conceptual model and the RF approach in capturing spatiotemporal and annual snow patterns, these discrepancies did not reach statistical significance.
Read full abstract