Abstract
Abstract. A deep learning rainfall–runoff model can take multiple meteorological forcing products as input and learn to combine them in spatially and temporally dynamic ways. This is demonstrated with Long Short-Term Memory networks (LSTMs) trained over basins in the continental US, using the Catchment Attributes and Meteorological data set for Large Sample Studies (CAMELS). Using meteorological input from different data products (North American Land Data Assimilation System, NLDAS, Maurer, and Daymet) in a single LSTM significantly improved simulation accuracy relative to using only individual meteorological products. A sensitivity analysis showed that the LSTM combines precipitation products in different ways, depending on location, and also in different ways for the simulation of different parts of the hydrograph.
Highlights
All meteorological forcing data available for hydrological modeling are subject to errors and uncertainty
There are possibilities to potentially improve on this benchmark (e.g., Duan et al, 2007; Madadgar and Moradkhani, 2014); as will be shown in Sect. 3, the difference between ensemble averaging and the multi-input Long Short-Term Memory networks (LSTMs) is large, and we would be surprised if any ensembling strategy could account for this difference
We trained n = 10 LSTMs using (1) all of the three forcing products together, (2) for each pairwise combination of forcing products (Daymet and Maurer, Daymet and North American Land Data Assimilation System (NLDAS), and Maurer and NLDAS), and (3) separately for all three forcing products individually. For each of these seven input configurations, we trained an ensemble of n = 10 different LSTMs with different randomly initialized weights
Summary
All meteorological forcing data available for hydrological modeling are subject to errors and uncertainty. While temperature estimates between different data products are frequently similar, precipitation estimates are often subject to large disagreements (e.g., Behnke et al, 2016; Timmermans et al, 2019). Large-scale hydrological models require spatial data (usually gridded), which are necessarily modelbased products resulting from a combination of spatial interpolation and/or satellite retrieval algorithms, and, sometimes, process-based modeling. Every precipitation data product is based on different sets of assumptions that each potentially introduce different types of error and information loss. It is difficult to predict a priori how methodological choices in precipitation modeling or interpolation algorithms might lead to different types of disagreements in the resulting data products (e.g., Beck et al, 2017; Newman et al, 2019). As an example of the consequences of this difficulty, Behnke et al (2016) showed that no existing gridded meteorological product is uniformly better than all others over the continental United States (CONUS)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
