Main text
Flooding is one of the most devastating natural disasters that affects communities globally. In early May 2023, catastrophic floods swept through Rwanda and the Democratic Republic of the Congo, leading to over 600 deaths and displacing thousands. As climate change is supercharging extreme rainfall worldwide, similar floods could further intensify. For developing regions with limited resources to build flood defenses, accurate flood forecasting with sufficient lead time is a cost-effective adaptation strategy to mitigate flood impacts. However, the lack of ground measurements in these regions hinders our understanding of flood-generating mechanisms, thereby limiting our ability to build better models for improved flood forecasts.
Two decades have passed since the International Association of Hydrological Sciences launched the Predictions in Ungauged Basins initiative in 2003. Despite advances in catchment hydrological processes, explosions of Earth observation data, and surges in computational power, the progress primarily benefits regions with dense gauge networks. Developing regions, which lack sufficient streamflow observations, continue to struggle with inaccurate and unreliable flood forecasts. There is potential to leverage the improved forecast performance from data-rich watersheds to enhance forecasts in data-scarce areas. However, this is challenging for traditional hydrological forecasts because the simplified conceptual representations of hydrological processes limit their transferability across watersheds with different features. Recent strides in deep learning offer new opportunities to address this. In The Innovation, Binlan Zhang and colleagues1 design a watershed-aware streamflow forecast model based on artificial intelligence (AI) to bridge the flood forecast gap in data-scarce developing regions, leveraging transfer learning techniques in AI. They show that AI-based models can capture watershed-dependent hydrological responses while learning complex rainfall-streamflow relationships. This capability opens doors to significant improvements in streamflow forecasts in data-scarce regions.
Design watershed-aware AI models
Traditional physics-based streamflow forecasts (Figure 1A) typically require years of gauge data to calibrate hydrological parameters. Forecasting floods requires even longer data to achieve robust statistical analysis, as floods lie in the tail part of the streamflow distribution. To facilitate streamflow forecasts in ungauged regions, hydrological regionalization has been proposed to “transfer” hydrological behaviors from gauged to ungauged catchments. This is typically done by linking hydrological model parameters to key watershed properties using similarity-based interpolations or linear regressions. Although conventional methods acknowledge the dependence of hydrological responses on watershed properties, they often implicitly assume watershed-dependent rainfall-streamflow relationships characterized by a limited set of key watershed features. This assumption may not fully capture the complexities of streamflow processes across geographically and climatically diverse watersheds, thereby limiting the performance of hydrological models when transferred to data-sparse regions.
Figure 1.
Improving flood forecasts over data-scarce regions with watershed-aware AI models
(A) Schematic illustration of the conventional hydrological model, where meteorological inputs are translated to hydrological outputs through updates in the state variables . The basin-specific parameters are optimized through calibration processes to reflect the basin’s properties.
(B) Architecture differences of three types of hydrological models (top: conventional hydrological model; middle: Zhang et al.’s feature-conditioned LSTM1; bottom: Google’s feature-augmented LSTM3).
(C) Taxonomy of existing and recommended model architectures for streamflow forecasts based on data requirements and interpretability. Feature-agnostic AI (i.e., AI models that do not consider basin features) tend to require less gauged data but suffer from lower interpretability compared to traditional hydrological models. By incorporating basin-specific attributes, similar to traditional parameter regionalization approaches, feature-informed AI models progress toward an ideal forecast paradigm. LSTM, long short-term memory; AI, artificial intelligence.
Recent strides in transferable AI offer a powerful tool to develop a more universal streamflow model that can generalize complex streamflow mechanisms across diverse watersheds. The so-called feature extraction techniques, which abstract salient patterns of inputs, enable models to learn and transfer knowledge across different contexts.2 Zhang et al.1 exemplify this approach by encoding basin spatial properties into long short-term memory (LSTM) networks. This facilitates the transfer of hydrological information from densely gauged regions to ungauged basins, even when they are located on different continents or hemispheres. Their AI model, trained on 671 watersheds in the data-rich United States, demonstrates its effectiveness by providing decent streamflow forecasts for 76.9% of 160 watersheds in data-scarce Chile.
Coincidentally, another recent study3 by Google also utilized AI for flood forecasts but at a global scale using data from 5,680 streamflow gauges. Google’s method also relies on an LSTM architecture but encodes watershed features differently. This approach demonstrates improved performance in forecasting fierce floods with return periods of up to 10 years. It outperforms the benchmark physics-based flood forecast system operated by the European Center for Medium Range Weather Forecasts (ECMWF). These two studies highlight the power of AI for flood forecasts in ungauged basins, paving the way for wider adoption of flood early warning systems.
A tale of two transferable LSTM models for streamflow forecasts
Conventional streamflow forecasts rely on process-based hydrological models. These models use human-interpretable equations to simulate complex watershed hydrological responses from meteorological inputs to streamflow outputs (Figure 1A). To maintain interpretability, key hydrological processes—such as runoff generation, infiltration, groundwater flow, and routing—are represented based on simplified physical laws (illustrated as water buckets in Figure 1A). Central to these physics-based models are dynamic intermediate state variables (e.g., snow water, soil moisture, and groundwater) and parameters (either empirical or physical) capturing relatively static basin-specific hydrological characteristics (such as soil properties and surface roughness). Both components are challenging to directly observe and thus often rely on model calibration. Calibration of these hydrological parameters involves two stages: a spin-up stage to adjust the states of the model toward equilibriums reflective of actual conditions (e.g., soil moisture levels) and a parameter-tuning stage to optimize the model’s ability to replicate hydrological responses influenced by watershed-specific factors such as climate, topography, soil types, and land use (top of Figure 1B). However, these simplified representations may not be sufficient to capture the full complexity, nonlinearity, and heterogeneity of hydrological processes at the watershed scale.
Building on earlier advances in LSTM-based streamflow simulations, Zhang and colleagues develop a feature-conditioned LSTM1 that encodes spatially distributed basin features in the model spin-up phase, which shares similarities with conventional hydrological models (middle of Figure 1B). Their watershed feature encoder incorporates not only static geographic attributes (e.g., elevations, channel shapes) but also key climatic states (e.g., soil moisture, groundwater) in spatially distributed forms. The resulting feature embeddings are then used to initialize the LSTM, similar to the traditional model spin-up practice using long-term climatology. Moreover, the extended weather hindcast period (180 days before the forecast time) acts similarly to a spin-up phase, allowing the watershed states (represented by LSTM’s cell and hidden states) to converge toward equilibriums reflective of actual conditions at the forecast time point.
Different from the feature-conditioned LSTM,1 Google has developed a feature-augmented LSTM3 that directly concatenates encoded basin-specific features with meteorological inputs in a tensor format (bottom of Figure 1B). Similar to Zhang et al.,1 their model uses an extended hindcast period (365 days) as a comprehensive spin-up phase to reduce biases from incorrect initializations. A structural distinction lies in the decoder LSTM. This decoder is continuously fed with dynamic meteorological forecasts from the ECMWF and may enhance the model’s predictive performance and its value for operational use. However, Google’s method only encodes watershed-averaged characteristics and discards the spatial heterogeneity information. A hybrid approach combining the two strategies—concatenating static spatially distributed geographical embeddings with meteorological inputs (feature-augmented LSTM) and initializing LSTMs using climatic state data (feature-conditioned LSTM)—could potentially further improve the forecast performance.
Challenges, cautions, and coming opportunities
Improving physical awareness and interpretability is a pivotal challenge for AI-based flood forecasts. While both LSTM architectures take innovative steps toward automated hydrological parameter regionalization by integrating the feature embedding module with the dynamic simulation module (LSTM blocks), this integration often compromises interpretability due to the “black box” nature of deep neural networks. Although efforts have been made in both studies to evaluate model performance attributable to the original physical watershed attributes, clear interpretability remains elusive. This gap hinders communications between researchers and real-world decision-makers, potentially undermining the usefulness of AI-based flood forecasts.
Nonetheless, we stand on the cusp of advancing flood forecasts that balance enhanced accuracy with interpretability (Figure 1C). In the AI era, explainable AI techniques offer a promising path to resolve the interpretability challenge for AI-based flood forecasts.2 For example, post hoc explanation methods, such as local interpretable model-agnostic explanations and Shapley values, can be used to analyze how streamflow responds to meteorological inputs. Meanwhile, for conventional physics-based hydrological models, the latest advances in differentiable parameter learning4 hold promise as an alternative to traditional regionalization methods. This technique allows efficient model parameter calibration without compromising the interpretability intrinsic to process-based models, potentially leading to further improvement in forecast performance in ungauged basins.
AI-based models can learn complex relationships between basin characteristics and streamflow processes, opening new avenues to investigate how watersheds respond to a changing environment. For example, certain Australian watersheds may not fully recover to pre-Millennium drought levels even after 7 years despite an increase in precipitation. AI models embedded with basin features could help assess how changes in basin characteristics influence watershed resilience to such climate shocks. Furthermore, flood events are manifesting in various forms that compound each other, including pluvial floods and coastal floods5 beyond the fluvial floods discussed earlier. Further research is needed to embed additional features relevant to multiple flood-generating drivers into AI-based models. In addition, growing evidence shows that climate-fueled floods are getting more extreme, with record-shattering events. This poses new challenges for AI-based models, such as forecasting floods that fall outside the historical range. Deep generative models, such as generative adversarial networks and diffusion models, offer a promising approach to simulating these unprecedented events. However, to ensure the simulated unseen floods remain physically plausible, incorporating hydrological knowledge into these models is crucial.
Acknowledgments
This work was supported by the Singapore Ministry of Education (MOE) Academic Research Fund Tier-1 project (A-0009297-01-00 and A-8001177-00-00) and Tier-2 project (A-8001886-00-00). X.H. acknowledges the National University of Singapore’s College of Design and Engineering for providing additional financial support through the Outstanding Early Career Award (A-8001228-00-00, A-8001389-00-00, and A-8001389-01-00).
Declaration of interests
The authors declare no competing interests.
Published Online: May 29, 2024
Contributor Information
Xiao Peng, Email: xp53@nus.edu.sg.
Xiaogang He, Email: hexg@nus.edu.sg.
References
- 1.Zhang B., Ouyang C., Cui P., et al. Deep learning for cross-region streamflow and flood forecasting at a global scale. Innovation. 2024;5(3) doi: 10.1016/j.xinn.2024.100617. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Reichstein M., Camps-Valls G., Stevens B., et al. Deep learning and process understanding for data-driven Earth system science. Nature. 2019;566(7743):195–204. doi: 10.1038/s41586-019-0912-1. [DOI] [PubMed] [Google Scholar]
- 3.Nearing G., Cohen D., Dube V., et al. Global prediction of extreme floods in ungauged watersheds. Nature. 2024;627(8004):559–563. doi: 10.1038/s41586-024-07145-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Tsai W.P., Feng D., Pan M., et al. From calibration to parameter learning: Harnessing the scaling effects of big data in geoscientific modeling. Nat. Commun. 2021;12(1):5988. doi: 10.1038/s41467-021-26107-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Qiu J., Liu B., Yang F., et al. Quantitative stress test of compound coastal-fluvial floods in China's Pearl River Delta. Earth's Future. 2022;10(5) doi: 10.1029/2021ef002638. [DOI] [Google Scholar]

