Editors’ Highlights are summaries of recent papers by AGU’s journal editors.
Source: Water Resources Research
In the realm of hydrologic prediction and forecasting, deep learning (DL) has been demonstrated to be useful and competitive with process-based modeling. The question remains whether DL is applicable in the context of hydrologic projections under climate change conditions due to the extrapolation challenge beyond the value range of past climate, which is intrinsic to any DL method.
In a formalized approach, Wi and Steinschneider [2022] study the applicability of long short-term memory networks (LSTMs) in projecting streamflow in a changing climate. In the approach, they train LSTM networks on historic data over watersheds in California and compare the streamflow projection results to results from simulations using three different hydrologic models. The comparison shows that the LSTM-based projections are unrealistic in some watersheds. The quality of the projections may be improved if LSTM training includes additional input from physics-based model projections.
In addition, by using a previously trained LSTM over 500 watersheds over the continental United States, the authors show that including diverse big data in the training is key in tackling the challenge of hydrologic projections under climate change conditions. The results also suggest that learned hydrologic responses can be transferred between watersheds, which is important for hydrologic prediction and forecasting.
Citation: Wi, S., & Steinschneider, S. (2022). Assessing the physical realism of deep learning hydrologic model projections under climate change. Water Resources Research, 58, e2022WR032123. https://doi.org/10.1029/2022WR032123
—Stefan Kollet, Editor, Water Resources Research