Deep learning for the assimilation of process-based hydrological models

In hydrology, neural networks (NN) are often used to replace hydrological models. While they have been proven to perform well for forecasting and simulating streamflow, they may not always be adequate when transparency and process understanding are required. However, NN can also be used as complements to process-based hydrological models (e.g., physics-based or conceptual) as part of a forecasting chain. For instance, in Boucher et al. (2020), an ensemble of multilayer perceptrons was used to perform data assimilation of streamflow in the GR4J conceptual model, which yielded promising results.

Building on the approach introduced in Boucher et al. (2020), the research presented here aims to improve the NN-based data assimilation method and to alleviate its limitations. To achieve this, multilayer perceptrons are replaced by long short-term memory (LSTM) networks, with an additional attention component. Both streamflow and snow are assimilated, the main focus being on the latter. For this reason, this new methodology has been tested on watersheds located in Canada, Norway and Sweden, including the Mistassibi watershed, which was also used in Boucher et al. (2020). Each watershed has been modelled using two hydrological model structures (GR4J and HMETS) within the Raven modelling framework.

Results show a successful assimilation of both streamflow and snow, which translates into improved daily streamflow simulations compared to the open-loop (according to the CRPS and reliability diagrams), for all catchments and for both models. In particular, results for Mistassibi show an improvement of the post-assimilation simulations compared to Boucher et al. (2020). This presentation will explain those results in detail and also describe the next steps to further expand and generalize the proposed data assimilation method.