Open AccessIdentification and Regionalization of Streamflow Routing Parameters Using Machine Learning for the HLM Hydrological Model in Iowa
Author(s) -
Velásquez Nicolás,
Mantilla Ricardo,
Krajewski Witold,
Quintero Felipe,
Zanchetta André D. L.
Publication year - 2022
Publication title -
journal of advances in modeling earth systems
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 3.03
H-Index - 58
ISSN - 1942-2466
DOI - 10.1029/2021ms002855
Subject(s) - routing (electronic design automation) , streamflow , computer science , benchmark (surveying) , interpolation (computer graphics) , flood control , random forest , flood forecasting , hydrology (agriculture) , environmental science , meteorology , flood myth , machine learning , artificial intelligence , geology , geography , cartography , drainage basin , motion (physics) , computer network , geotechnical engineering , archaeology
We present a novel approach to determine spatially distributed routing parameters for the distributed hydrological Hillslope Link Model (HLM), an ordinary differential equations‐based streamflow forecasting model implemented and tested in Iowa. We being by developing a technique to determine two model parameters that control the channel routing equation in gauged catchments draining less than 1,300 km 2 . Then, we implement a parameter regionalization methodology using machine learning classification techniques and a bootstrap procedure, in which we trained 400 Random Forests (RFs) using physical and geomorphological features for classification. We made a regional interpolation using an ensemble of selected RF realizations that exhibited the best performance. We used as benchmark of our results a more straightforward interpolation technique based on USGS Hydrological Units Codes. We performed simulations of the HLM over the entire state of Iowa between 2012 and 2018 using the two regionalization methods, comparing them to the operational model used by the Iowa Flood Center, which applies a single set of parameter values to the entire domain. After evaluating the results at 148 USGS stations, the Random‐Forest approach captures the value of observed peak flows more precisely without losing performance in terms of the Kling Gupta Efficiency index. The improvements obtained using our proposed strategy that uses data, hydrological modeling, and a machine learning technique to identify and regionalize routing parameters are modest, indicating that the parameters that control the rainfall‐runoff transformation dominate uncertainty in our flood forecast model.