Towards Deep Learning River Network Models

Deep Learning models for streamflow prediction are now more than five years old (Kratzert et al., 2018, 2019), and lumped LSTMs, trained on as many basins and forcing products as we can get our hands on, continue to pose the state of the art. Or do they?

While traditional hydrologic modeling has long moved beyond lumped modeling, Deep Learning methods are only now starting to leverage the inherent graphical topology of rivers through graph neural networks (GNNs). Such models come with their own set of challenges, both from an engineering standpoint (e.g., dealing with the sheer amount of data from many small sub-basins) and from a modeling standpoint (e.g., ensuring generalization to ungauged basins along the river graph). Yet, GNNs promise more accurate predictions, the ability to assimilate real-time up- and downstream data, make predictions at arbitrary points along a river, or integrate knowledge about human intervention. 

We present a Deep Learning semi-distributed hydrologic model that combines the time-series capabilities of LSTMs with a learned GNN routing mechanism. The model is trained on streamflow data from all around the world, providing predictions that are strong competitors to their lumped counterparts—especially on large, ungauged rivers.

Kratzert, F., Klotz, D., Brenner, C., Schulz, K., and Herrnegger, M.: Rainfall–runoff modelling using Long Short-Term Memory (LSTM) networks, Hydrol. Earth Syst. Sci., 22, 6005–6022, 2018.

Kratzert, F., Klotz, D., Shalev, G., Klambauer, G., Hochreiter, S., and Nearing, G.: Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets, Hydrol. Earth Syst. Sci., 23, 5089–5110, 2019.