Continuous streamflow prediction in ungauged basins: Long Short-Term Memory Neural Networks clearly outperform hydrological models

Abstract

Abstract. This study investigates the ability of Long Short-Term Memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A series of state-of-the-art, hydrological model-dependent regionalization methods is applied to 148 catchments in Northeast North America and compared to a LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model’s structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.

In Defense of Metrics: Metrics Sufficiently Encode Typical Human Preferences Regarding Hydrological Model Performance

“…In recent years, a range of papers (including those from authors of this manuscript) have shown that data-driven LSTM-based models clearly and consistently outperform existing traditional PC-based models for streamflow prediction, across a large variety of regions and scenarios (e.g., gauged (Kratzert, Klotz, Shalev, et al, 2019;Lees et al, 2021;Koch & Schneider, 2022;Gauch et al, 2021;Mai et al, 2022), ungauged (Kratzert, Klotz, Herrnegger, et al, 2019;Feng et al, 2021;Arsenault et al, 2022;Mai et al, 2022), and extreme events (Frame et al, 2022)). Despite these successes, parts of the scientific community have questioned the viability of LSTM-based models as "hydrologic" models in their own right (Razavi, 2021;Nearing et al, 2021)-as models that can not only provide better predictions, but that also advance hydrologic knowledge.…”

“…The large number of available hydrologic models raises the obvious challenge of how one should evaluate and compare models when faced with the need to choose the most suitable model for a given task (e.g., Krause et al, 2005;Gupta et al, 2009;Moriasi et al, 2015;Garcia et al, 2017). Benchmark studies are one approach that can help address this problem, and thus have gained substantial traction (e.g., Best et al, 2015;Kratzert, Klotz, Shalev, et al, 2019;Lees et al, 2021;Mai et al, 2022;Arsenault et al, 2022). Overall, the community usually bases their strategy for model performance evaluation and comparison on sets of quantitative metrics.…”

Probabilistic Forecasts of Flood Inundation Maps Using Surrogate Models

“…Alternatively, surrogate models with different structures can be explored to compose the multi-model ensemble. Other recurrent algorithms commonly applied for river flow forecasting, such as the Gated Recurrent Unit (GRU) [46] and the Long Short-Term Memory (LSTM) [47][48][49][50], are suggested alternatives for the NARX structure adopted in this work, while convolutional neural networks can be used to compose the flood inundation maps [51]. Additionally, Mosavi et al [52] (Section 4.2 of their work) listed a series of hybrid models already explored for short-term forecasting that potentially can be adapted for the prediction of flood inundation maps.…”