Authors: Juliette Mukangango – Colorado School of Mines
Title: Benchmarking the relative quality of available meteorological input datasets for large domain streamflow and snow (hence, hydrology) modeling
Presentation Type: Lightning Talk
Abstract: Hydrological forecasting is essential for effective water resource management and requires high-quality meteorological forcing inputs. This project, which is a part of the CIROH Testbed effort, evaluates the quality of multiple meteorological forcing datasets using deep learning methods to benchmark the information content of each dataset from the combined perspectives of both rainfall-runoff and snow modeling. Meteorological inputs for precipitation and air temperature are collected from various datasets, including Daymet, Livneh, Maurer, NLDAS2, AORC, ERA5-Land, CONUS404, and an ensemble GPEP forcing dataset. These are evaluated for their ability to support calibrated simulations of streamflow and snow water equivalent (SWE). LSTM models for both streamflow and snow are trained for each dataset using the Neural-Hydrology (NH) package, with analysis conducted across 671 basins in the continental US (CONUS) and approximately 800 SWE measuring (SNOTEL) sites located in high-elevation mountain watersheds in the western US. Model performance is evaluated using common metrics such as the RMSE, NSE, and KGE. We report on the experimental framework and results to date. Overall, this research provides evidenced-based support for decisions on forcing dataset development to support hydrologic forecasting applications.