How To Calibrate A Process-Based Model Using Deep-Learning: Applying Simulation-Based Inference To A Hydrologic Model Of The Upper Colorado River Basin
Presenter: Quinn Hull1
Co-Author(s): Andrew Bennett
Luis De La Fuente
Elena Leonarduzzi
Peter Melchior
Reed Maxwell
Hoang Viet Tran
Advisor(s): Dr. Laura Condon
1Department of Hydrology and Atmospheric Sciences, University of Arizona
Panapto Presentation Video
Poster PDF
Poster Session 2
High-fidelity, process-based (PB) hydrologic models are needed to make smart decisions about water in a future without historical analog. However, calibrating them to observations is hampered in part by their complexity leading to equifinality and high computational demands. Recent advances and availability of deep learning technology provide new opportunities to (1) conduct Bayesian-style inference of probable parameter sets relaxing assumptions of a single ‘best’ model configuration, and (2) reduce the computational cost of inference by emulating behaviors-of-interest learned from the PB. This study confronts the challenges of inferring two common parameters in hydrologic PBs, Manning’s coefficient and hydraulic conductivity, in a snowmelt-dominated catchment in the Colorado River. Here we apply a simulation-based inference (SBI) approach. First, a Long Short-Term Memory machine learning (LSTM) emulator is trained on an ensemble of ParFlow (PB model) outputs with systematically varied surface and subsurface parameters. Second, a neural density estimator (NDE) is applied in concert with the emulator to learn a full probabilistic mapping between simulation parameters and outputs to generate posterior distributions of parameters. The performance of SBI is tested via experiments that vary the degree (and sources) of uncertainty associated with the processes involved. We find that the NDE powerfully constrains parameter uncertainty in controlled cases but is only as robust as the diversity of system representations used to train it. While other Bayesian inference approaches have been applied in watershed studies, SBI has not been widely used in the field of hydrology. Our study provides a proof-of-concept example of how deep learning techniques may make inroads in resolving the 'intractable inverse problem' related to many PB models. Further work is needed to identify ways to implement this method in operational settings.