Frederik Kratzert, Daniel Klotz, Mathew Herrnegger, Alden K Sampson, Sepp Hochreiter, and Grey S Nearing

Traditional hydrological models show best performance when calibrated for an individual basin.  This is because these process-driven models explicitly include conditions and features of a specific basin. In a recent paper we showed that data-driven models trained on hundreds of basins significantly outperform traditional models. Here, we extend this approach to ungauged basins, where prediction is made for a set of basins not used in training the model. Prediction in ungauged basins is an important challenge because the majority of streams in the world are either ungauged or poorly gauged, and obtaining detailed information for an individual basin is expensive.

Our model is based on long short‐term memory (LSTM) networks.  It is trained on 531 basins from the Catchment Attributes and Meteorology for Large-Sample Studies (CAMELS) dataset. The dataset includes about 30 years of daily rainfall‐runoff data from catchments in the United States ranging in size from  4 to 2,000 km2 with varying aridity and vegetation. It also contains several static catchment attributes related to soils, climate, vegetation, topography, and geology.

The model is compared against the Sacramento Soil Moisture Accounting (SAC‐SMA) model and against the US National Water Model (NWM) on 15 years of daily data. SAC‐SMA was calibrated on individual gauged basins; the NWM is a distributed process-based model. We find that our LSTM-based model for ungauged basins outperforms on average both the SAC‐SMA and the NWM models. Our results indicate that there is sufficient information in catchment attribute data to distinguish between different hydrological regimes, and that LSTM is able to extrapolate catchment attributes to new basins. These findings showcase the power of machine learning in synthesizing information from multiple sites and conditions into a single model.

The code developed for this project is available on GitHub.

Water Resources Research. 55, 12, 11344-11354. 2019-12-23.

View paper
IARAI Authors
Dr Sepp Hochreiter
Weather and Climate
Deep Learning, Hydrology, Long Short-Term Memory, Rainfall-Runoff


Imprint | Privacy Policy

Stay in the know with developments at IARAI

We can let you know if there’s any

updates from the Institute.
You can later also tailor your news feed to specific research areas or keywords (Privacy)

Log in with your credentials

Forgot your details?

Create Account