ML framework for global river flood predictions based on the Caravan dataset

,

💡 Research Summary

The paper presents a comprehensive machine‑learning framework designed to predict river flood events on a global scale, leveraging the newly released Caravan dataset—a massive, multi‑source collection that integrates daily precipitation, soil moisture, topographic, and hydrological observations from over 5,000 gauging stations worldwide, together with satellite‑derived products spanning the period 2000‑2022. The authors begin by describing an extensive data‑preprocessing pipeline that addresses missing values through a hybrid of multivariate regression and temporal interpolation, followed by log‑transformation and z‑score normalization to harmonize variable scales. Feature engineering is carried out in four thematic groups: multi‑scale accumulated rainfall (1 h, 3 h, 6 h, 12 h, 24 h windows) processed by a multi‑head attention layer, soil‑moisture indices, basin‑specific flow‑resistance coefficients derived from land‑cover and slope, and climate‑variability descriptors such as the Standardized Precipitation Index.

The core predictive architecture consists of two tightly coupled modules. The first is a Graph Neural Network (GNN) that treats each watershed as a node and encodes upstream‑downstream hydraulic connectivity via an adjacency matrix weighted by basin area ratios and gradient. Message‑passing within the GNN captures the spatial propagation of water volumes across the river network. The second module is a Long‑Short‑Term Memory (LSTM) network that ingests the time‑series of engineered features for each basin, learning temporal dynamics such as lagged runoff response. A cross‑attention mechanism fuses the spatial embeddings from the GNN with the temporal hidden states of the LSTM, producing a unified spatio‑temporal representation. The final output layer simultaneously predicts quantitative discharge values for 24‑, 48‑, and 72‑hour horizons and a probabilistic flood‑risk score (0–1) that can be directly used for early‑warning systems.

Training employs a composite loss function to mitigate the severe class imbalance inherent in flood datasets: a focal loss component emphasizes correctly classifying rare flood events, while a weighted Mean Squared Error term penalizes discharge prediction errors proportionally to basin‑specific variance. To improve generalization across heterogeneous regions, the authors introduce a domain‑adaptation batch normalization scheme that aligns feature distributions between high‑data‑quality basins (e.g., Europe, North America) and data‑scarce regions (e.g., parts of Africa and Southeast Asia). Model evaluation follows a rigorous 5‑fold cross‑validation protocol and an independent hold‑out test set covering 2021‑2022. Performance metrics include RMSE, MAE, R², and the F1‑score for binary flood detection. Compared with the conventional HEC‑RAS hydrological model, the proposed framework reduces RMSE by 18 % and raises the F1‑score from 0.78 to 0.87, indicating a substantial gain in both accuracy and reliability. Inference speed averages 0.15 seconds per basin, making the system suitable for real‑time operational deployment.

A detailed sensitivity analysis reveals that precipitation intensity and soil saturation are the dominant drivers of prediction skill, while basin area and slope contribute moderately. The model’s performance degrades by roughly 7 % in regions where observational coverage is sparse, highlighting the need for improved local monitoring networks. Moreover, under extreme climate scenarios—such as 100‑year return‑period storms—the predictive uncertainty widens, prompting the authors to suggest future integration of Bayesian deep‑learning techniques for calibrated uncertainty quantification.

In conclusion, the study demonstrates that a hybrid GNN‑LSTM architecture, when fed with richly engineered features from the Caravan dataset, can surpass traditional physically‑based flood models in both precision and computational efficiency on a planetary scale. The authors outline three primary avenues for future work: (1) tighter coupling of physics‑based simulators with data‑driven components to enforce hydrological consistency, (2) streaming integration of real‑time IoT sensor data and satellite observations for continuous model updating, and (3) scenario‑based forecasting that incorporates projected climate change impacts. By addressing these directions, the framework has the potential to become a cornerstone of global flood risk management and disaster‑response strategies.