Transferring hydrologic data across the continent--how do data from US catchments benefit flood prediction in other countries?

Abstract:

There is a drastic geographic imbalance in available global streamflow gauges and subsurface property data, leading to inequalities in the performance of data-driven hydrologic models. It was typically difficult to migrate knowledge learned from one region to another. The properties of the forcing and attributes also vary substantially so that models calibrated in one region cannot be directly migrated to other regions. Here we show that transfer learning (TL) allows long short-term memory (LSTM) streamflow models that were trained in the Continental United States (CONUS) to be transferred to catchments on other continents, without the need for extensive catchment attributes. We demonstrate this possibility for basins where data are relatively dense and where data are scarce and only globally-available attributes are available. In both China and Chile, while directly applying the CONUS-trained model gave catastrophic results, the TL models significantly elevated model performance compared to locally-trained models, achieving state-of-the-art metrics. The benefits of TL, more prominent for target regions with scarce data, also increased with the amount of available data in the source dataset, suggesting the model leveraged the knowledge extracted from CONUS. The benefits were also greater than pre-training LSTM using the outputs from an uncalibrated hydrologic model and were more prominent when the target datasets have limited lengths. We compared multiple TL options and found that different options were optimal for Chile and China. It suggests hydrologic data around the world have a commonality which could be leveraged by big data models, and we should modify our modeling workflow in data-scarce regions. Finally, this work diversified existing global streamflow benchmarks.