Perera et al. (2026) Hybrid methods in flood inundation modeling: a systematic review

Identification

• Journal: Natural Hazards
• Year: 2026
• Date: 2026-04-01
• Authors: Uditha Perera, Athanasios Angeloudis, Adil Siripatana, Lindsay Beevers
• DOI: 10.1007/s11069-026-08078-w

Research Groups

Institute for Infrastructure and Environment (IIE), School of Engineering, University of Edinburgh, Edinburgh, EH8 9YL, UK

Short Summary

This systematic review defines and classifies hybrid flood inundation models, evaluates their advantages and limitations over standalone models, and proposes a standardized benchmarking framework to guide their development and application, highlighting Physics-Informed Neural Networks (PINNs) as a promising future direction.

Objective

• To explore the limitations of standalone hydrodynamic and common machine learning (ML) models that drive research into hybrid models.
• To define 'hybridization' and analyze the techniques applied in creating hybrid flood models.
• To examine common performance metrics for evaluating hybrid flood models and propose a benchmarking framework for model comparison.
• To critically analyze the success of current 'hybrid flood models' and suggest a way forward, particularly concerning physics-informed machine learning.

Study Configuration

• Spatial Scale: The review covers literature on flood inundation models applicable across various catchment scales (small, medium, large) and resolutions (fine, medium, coarse), including urban and large-scale river systems.
• Temporal Scale: The review covers literature published from 2010 onward, discussing flood prediction models with both short-term (hourly, daily, weekly) and long-term (longer than a week) prediction horizons.

Methodology and Data

• Models used: This paper is a systematic review of existing hybrid flood inundation models. It analyzes combinations of process-based (hydrodynamic) models and data-driven (machine learning, deep learning) models, including specific architectures like Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forests (RF), Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and Physics-Informed Neural Networks (PINNs).
• Data sources: The systematic review itself was conducted using keyword searches in Scopus and Web of Science databases. The models reviewed in the literature utilize diverse data sources including satellite aperture radar (SAR) imagery, numerical model outputs, observed gauge data, social media/crowd-sourced data, topographical data (e.g., Digital Elevation Models), meteorological data, geological data, geographical data, and anthropogenic data.

Main Results

• Research on hybrid flood inundation models shows an increasing trend, indicating growing interest in combining process-based and data-driven approaches.
• Hybridization is defined as a method of combining the strengths of standalone process-based (hydrodynamic) and machine learning models to enhance the input, structure, or processing of the resulting hybrid model, aiming to improve performance in terms of physics incorporation, representability, accuracy, speed, and generalizability.
• Hybridization techniques are classified into three main categories: enhancing inputs (e.g., incorporating SAR data or hydrodynamic model-generated data), enhancing structure (e.g., optimization, feature-informed models, Physics-Informed Neural Networks, combining ML structures), and enhancing processing (e.g., ensemble models, parallel processing).
• Evaluation metrics for flood models are categorized into accuracy-based metrics (e.g., RMSE, MAE, R², Nash-Sutcliffe Efficiency) and speed-based metrics (e.g., computational time, CPU time).
• Hybrid models demonstrate advantages over standalone models in rapid prediction, real-time applicability, accuracy, utility in data-scarce areas, and lower computational cost, but still face limitations in data dependency, generalizability, and the need for specialized expertise.
• A standardized benchmarking framework is proposed, which involves classifying models by hybridization method, catchment scale, and resolution, and then comparing them against hydrodynamic and standalone ML models using standardized datasets, tests, and a hierarchy of primary and secondary evaluation metrics.
• Physics-Informed Neural Networks (PINNs) are identified as a promising way forward, as they embed physical laws into ML models, reducing their black-box nature, increasing data efficiency, and improving generalizability, despite potentially higher computational costs for training.

Contributions

• Provides a comprehensive systematic review of state-of-the-art hybrid methods in flood inundation modeling.
• Proposes a standardized definition and a novel classification framework for hybridization techniques in flood modeling.
• Categorizes and discusses the diversity of assessment metrics used for hybrid flood models.
• Develops and proposes an extensive benchmarking framework for the objective comparison and selection of hybrid flood models.
• Identifies and critically analyzes the advantages and limitations of hybrid models compared to standalone approaches.
• Suggests a clear way forward for future research, particularly emphasizing the potential of Physics-Informed Machine Learning (PINNs) in flood inundation modeling.

Funding

• UKRI grant EPSRC EP/X041093/1
• Funding from the School of Engineering of the University Of Edinburgh

Citation

@article{Perera2026Hybrid,
  author = {Perera, Uditha and Angeloudis, Athanasios and Siripatana, Adil and Beevers, Lindsay},
  title = {Hybrid methods in flood inundation modeling: a systematic review},
  journal = {Natural Hazards},
  year = {2026},
  doi = {10.1007/s11069-026-08078-w},
  url = {https://doi.org/10.1007/s11069-026-08078-w}
}