During a flood, the best physics-based forecasts can be too slow to guide same-day decisions. This study asks whether a faster learner can preserve the useful part: knowing where water will rise, and why. As flood occurrences become more intense and frequent, decision makers, emergency managers, infrastructure owners and operators, and first responders have struggled to expeditiously assess and respond to flood accidents.

Physics-based models are typically used to compute flood hazards, but they could get computationally expensive and even prohibitive as the area, detail, and complexity grow. The challenge is to minimize computational cost while maintaining accuracy, reliability, and interpretability for near-time flood prediction.

It is like needing a detailed map during a crisis, but having to draw the map before anyone can use it. MaxFloodCast provides peak inundation predictions while taking into account nearby and upstream precipitation information, all while maintaining model interpretability.

Because historical inundation data can be scarce, MaxFloodCast is trained and tested using physics-based model simulations. It was applied in the multi-watershed region of Harris County and validated by reproducing Hurricane Harvey and Tropical Storm Imelda events.

The first setup incorporated peak and cumulative precipitation within each cell. The second setup added watershed heavy cumulative precipitation ratio and heavy peak precipitation. Using only peak and cumulative precipitation within each cell, Experiment 1 demonstrates strong performance in predicting inundation depth in the non-channel cells.

The performance in the channel cells was comparatively weaker because the first setup did not consider precipitation in contributing nearby and upstream drainage areas. Including precipitation information from contributing watersheds and drainage areas improved results in the channel cells.

But those watershed features slightly worsened predictions in non-channel cells, likely because they introduced noise and extra information. Since approximately 80% of the cells are non-channel cells, the overall performance of the second setup was inferior to the first.

The model was tested against two severe storms: Hurricane Harvey and Tropical Storm Imelda. They were chosen because their size and rainfall patterns differed. Peak water depth records collected at channel gages during those storms were used to validate the model’s peak depth predictions, with the physics-based model provided for reference.

Models trained using physics-based simulations can provide comparable prediction performance to hydrodynamic models while improving computational efficiency and model interpretability. For most flood intensities considered, MaxFloodCast’s prediction performance aligned with that of a standalone hydrodynamic model.

However, prediction of extreme events was less successful, especially where flood depths were greater than 15 feet, partly because extreme events were lacking in the training set. In its current form, the model only looks at peak flood depth, which represents a proxy for flood severity and is used to estimate damage to buildings and infrastructure.

But time to peak is essential for timely decisions, and inundation duration affects the length of emergency and recovery phases. Future developments aim to include more information and produce other parts of the flood’s rise and fall in addition to the peak, helping emergency operators understand what happens and when.

A model trained on physics-based flood simulations matched their prediction performance for most conditions while using information people can interpret. It could help emergency teams see likely flood depths sooner, though the most extreme floods remain difficult.