A new algorithmic framework that can predict flooding could help save lives and reduce the devastation as climate change
drives more intense and unpredictable rainfall.
The model described in the International Journal of Information and Communication Technology uses the Multi-Scale
Adaptive Neuro-Fuzzy Inference System (MS-ANFIS) and combines deep learning with a form of fuzzy logic that quantifies
uncertainty; features that were missing from earlier data-driven flood models.
Flood prediction usually focuses on hydrologic models that simulate how rainfall moves across landscapes and into
rivers. These are grounded in environmental science but depend on detailed land-surface information and can be
computationally expensive, limiting their usefulness for rapid or large-scale forecasting.
Attempts to reduce the computing demands as well as speed up predictions using statistical and early machine-learning
approaches have proved useful but still struggle to cope with diverse data sources or respond to highly localized
Even cutting-edge deep-learning models, which can spot patterns in vast datasets, treat river systems as deterministic
in behavior and do not take into account the inherent variability that arises because of extreme weather.
MS-ANFIS might plug the holes in earlier approaches. It uses a feature pyramid network. This is a deep-learning
architecture that extracts information at multiple scales. In doing so, it can capture detailed runoff patterns and
broader rainfall trends visible in satellite data.
The fuzzy layer then interprets the data and expresses uncertainty in a structured, interpretable way. The result is
flood prediction with a measure of confidence in the prediction built in.
The researchers have tested their system on data from five major river basins, covering markedly different weather
patterns and hydrologic behavior.