Machine Learning Applications in Urban Flood Prediction

Urban flood prediction has become an increasingly data-driven discipline, driven by the need for rapid, accurate forecasts in the face of climate change and expanding impervious surfaces. Traditional physics-based hydraulic models deliver reliable representations of flow dynamics but are often computationally prohibitive for large-scale or real-time applications. Machine learning approaches, including ensemble methods, deep neural networks and hybrid frameworks that couple physical simulators with data-driven algorithms, have emerged to bridge this gap. These techniques excel at processing heterogeneous inputs—such as rainfall time series, terrain elevation, sewer network layouts and land-use data—to generate fine-grained inundation maps, water-depth forecasts and early-warning indicators. Advances in spatiotemporal modelling permit explicit characterisation of interactions across manholes, pipes and surface runoff, while probabilistic strategies quantify prediction uncertainty. Interpretability tools further elucidate how key drivers—such as antecedent rainfall or underground drainage constraints—influence model outputs. Collectively, these innovations enable rapid scenario analysis for emergency management, inform infrastructure design and underpin adaptive flood-risk planning in cities worldwide.

Research from Nature Portfolio

Recent work has demonstrated a scalable framework that marries two-dimensional hydraulic simulations with machine learning classifiers and regressors. A flood-mapping classifier first discriminates water-covered from dry nodes, after which a multilayer perceptron estimates water depth at wetted locations. This two-stage hybrid approach achieves over 98 % accuracy in wet/dry detection and a strong regression fit for depth, while reducing computational costs by orders of magnitude compared with full hydraulic runs. The framework’s adaptability to arbitrary catchment scales and its retention of physical consistency make it a promising tool for continental and global flood forecasting systems.

Machine Learning Applications in Urban Flood Prediction publication trend

The graph below shows the total number of articles in machine learning applications in urban flood prediction across all publications each year (not limited to Nature Index journals).

Technical terms

Hydraulic model: A computational tool that solves conservation equations of mass and momentum to simulate water flow and inundation dynamics.

Bayesian neural network: A neural network framework that treats model weights as probability distributions, enabling explicit quantification of predictive uncertainty.

Shapley Additive Explanations (SHAP): An interpretability technique based on cooperative game theory, attributing the impact of each input feature on a model’s output.

Gaussian process emulator: A statistical surrogate model that approximates complex simulation outputs and provides uncertainty estimates without rerunning the original solver.

Spatiotemporal effects: The combined spatial and temporal interactions within a system, such as evolving water levels across a sewer network over time.

References

1. A Framework for Modeling Flood Depth Using a Hybrid of Hydraulics and Machine Learning. Scientific Reports (2020).
2. SHAP-powered insights into spatiotemporal effects: Unlocking explainable Bayesian-neural-network urban flood forecasting. International Journal of Applied Earth Observation and Geoinformation (2024).
3. Gaussian process emulation of spatio-temporal outputs of a 2D inland flood model. Water Research (2022).
4. U-FLOOD – Topographic deep learning for predicting urban pluvial flood water depth. Journal of Hydrology (2021).