Authors: Andy Wood, Rhys Pulling, Guoqiang Tang, Xinxin Sui, Xia Feng
Title: Assessing a large-sample emulator strategy for calibrating the NextGen CFE+ model in CAMELS US watersheds
Presentation Type: Poster Presentation
Abstract: Accurate prediction of streamflow is a fundamental challenge in hydrology with valuable applications in mitigating flood hazards and managing water resources. Conceptual and process-based hydrology models have played a central role in this endeavor for decades, yet as their deployment moves from local to national scales, long-standing methods for calibrating their parameters often fail to offer a scalable, efficient strategy for achieving fully regionalized parameter sets that yield uniformly skillful model performance. Traditional parameter optimization approaches have often relied on basin-specific tuning using local observations, which does not scale well to large domains and complicates the transfer of parameters to ungauged or unobserved regions. Recently, a new process-model calibration strategy has emerged that uses a ‘Large-Sample Emulator (LSE)’, i.e., a machine learning (ML) model emulator trained on model results from a large sample of watersheds that can be used to search cheaply for optimal parameters in a large-domain process model. We have assessed the performance of the LSE approach for calibrating Conceptual Functional Equivalent (CFE) model, a core conceptual formulation used in the NextGen modeling framework, focusing on a 498-basin subset of watersheds from the Catchment Attributes for Meteorological Studies (CAMELS) collection. We extended an existing open-source LSE codebase designed for conceptual modeling to incorporate the CFE model (modified to include a snow temperature index model, labeled CFE+) and evaluate its calibration and temporal validation performance. This presentation discusses our recent study details and findings, which demonstrate that the LSE-calibrated CFE+ models achieve comparable accuracy to other CAMELs conceptual model results, with low computational cost. The study argues for the viability of the LSE strategy for NextGen models and also yields the first strong CFE model baseline of calibrated performance on CAMELS basins, which can in turn be used to benchmark advances in other models and calibration strategies.