INFLOW-AI v2.1: a machine learning framework for predicting out-of-sample extreme seasonal flood extents

Rapson, J., Stephens, E. ORCID: https://orcid.org/0000-0002-5439-7563, Maidment, R. I. ORCID: https://orcid.org/0000-0003-2054-3259 and Bonifacio, R. (2026) INFLOW-AI v2.1: a machine learning framework for predicting out-of-sample extreme seasonal flood extents. Geoscientific Model Development. ISSN 1991-9603 (In Press)

Abstract/Summary

Forecasting extreme flooding remains a major challenge in hydrological modelling, particularly in data-scarce and highly dynamic floodplain systems characterised by complex surface water, groundwater, and evapotranspiration processes. Machine learning provides an alternative by learning nonlinear relationships directly from data but often struggles to predict events that exceed those represented in the training data. We present INFLOW-AI v2.1, a two-stage machine learning framework for forecasting seasonal flood extent under hydrometeorological regime shifts. The framework addresses the extrapolation problem by predicting the seasonally differenced anomaly in total flood extent rather than the raw flood extent, thereby improving robustness to out-of-sample extreme events. The first stage uses a transformer with multi-head attention to model long- and short-term temporal dependencies from the preceding 36 dekads and predict the transformed total flood extent. The second stage uses a ConvLSTM to predict inundation likelihoods at a 1 km spatial resolution, with the predicted total inundation extent constraining the spatial forecasts. The model was evaluated over the Sudd wetlands in South Sudan, where the seasonally differenced anomaly transformation improved predictions of extreme flooding beyond the range of observations used for training. INFLOW-AI has been deployed operationally since the 2024 flood season, providing real-time flood forecasts with a two-month lead time to support humanitarian operations in South Sudan.

Item Type	Article
URI	https://centaur.reading.ac.uk/id/eprint/131461
Refereed	Yes
Divisions	Science > School of Mathematical, Physical and Computational Sciences > Department of Meteorology
Publisher	European Geosciences Union
Download/View statistics	View download statistics for this item