INFLOW-AI v2.1

We present INFLOW AI, a machine learning framework for predicting out of sample extreme seasonal flood extents. The framework employs a two-stage neural network architecture that combines (1) temporal dynamic thresholds with (2) local spatial interpolation to enhance predictive accuracy for out-of-sample extreme events. The first stage employs transformer-based models with multi-headed attention mechanisms to capture long- and short-term hydrometeorological patterns over the past 36 dekads. This stage predicts a seasonally differenced anomaly target, enabling more effective detection of extremes. The second stage uses a ConvLSTM framework to model local spatial flooding probabilities at 1 km² resolution, using the temporal prediction from the first stage as a threshold for spatial flood probability filling. The model generates forecasts up to six dekads (two months) in advance and is deployed on the Joint Analysis System Meeting Infrastructure Needs (JASMIN), a supercomputing cluster managed by the Science & Technology Facilities Council (STFC) for the Natural Environment Research Council (NERC).

Full Changelog: https://github.com/rapsoj/INFLOW-AI/commits/v2.1