Comparison of machine learning algorithms for emulation of a gridded hydrological model given spatially explicit inputs. (February 2022)

• Record Type: Journal Article
• Authors: Lim, Theodore; Wang, Kaidi
• Is Part Of: Computers & geosciences. Volume 159(2022)
• Publication Date: 2022-02
• Subjects: Emulation modeling -- Surrogate modeling -- ParFlow.CLM -- Machine learning; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2021.105025
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: This study compares the performance of several machine learning algorithms in reproducing the spatial and temporal outputs of the process-based, hydrological model, ParFlow.CLM. Emulators or surrogate models are often used to reduce complexity and simulation times of complex models, and have typically been applied to evaluate parameter sensitivity or for model parameter tuning, without explicit treatment of variation resulting from spatially explicit inputs to the model. Here we present a case study in which we evaluate candidate machine learning algorithms for suitability emulating model outputs given spatially explicit inputs. We find that among random forest, gaussian process, k-nearest neighbors, and deep neural networks, the random forest algorithm performs the best on small training sets, is not as sensitive to hyperparameters chosen for the machine learning model, and can be trained quickly. Although deep neural networks were hypothesized to be able to better capture the potential nonlinear interactions in ParFlow.CLM, they also required more training data and much more refined tuning of hyperparameters to achieve the potential benefits of the algorithm. Highlights: Machine learning algorithms trained to emulate spatial-temporal output of gridded hydrological model. Among tested algorithms, random forest regressor exhibited best fidelity, stability, and training time. Hyperparameter tuning and data science expertise present additional barriers to emulation modeling using machine learning. Need to consider costs of training sample generation when selecting machine learning algorithm for emulator.