Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling

Efficient and accurate flood inundation mapping is essential for flood risk assessment, emergency response, and community safety. The deep learning-enabled rapid flood simulation demonstrates superior computational efficiency compared to traditional hydrodynamic models. However, most deep learning-b...

Full description

Saved in:
Bibliographic Details
Main Authors: Wenke Song, Mingfu Guan, Kaihua Guo, Dapeng Yu
Format: Article
Language:English
Published: Taylor & Francis Group 2025-12-01
Series:Engineering Applications of Computational Fluid Mechanics
Subjects:
Online Access:https://www.tandfonline.com/doi/10.1080/19942060.2025.2481115
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849393384381218816
author Wenke Song
Mingfu Guan
Kaihua Guo
Dapeng Yu
author_facet Wenke Song
Mingfu Guan
Kaihua Guo
Dapeng Yu
author_sort Wenke Song
collection DOAJ
description Efficient and accurate flood inundation mapping is essential for flood risk assessment, emergency response, and community safety. The deep learning-enabled rapid flood simulation demonstrates superior computational efficiency compared to traditional hydrodynamic models. However, most deep learning-based flood mapping models currently focus on predicting the maximum water depth and face challenges in generalizing rainfall events of different durations. This paper proposes a fast flood simulation method based on image super-resolution, utilizing a novel DenseUNet architecture to predict the maximum water depth and velocity of temporal rainfall events. The proposed method integrates physical catchment characteristics to enhance the resolution of flood maps generated by the coarse-grid hydrodynamic model using the deep-learning model. The method is applied to a rural-urban catchment of the Shenzhen River in southern China. The proposed method effectively reproduces maximum water depth and velocity for test rainfall events against the fine-grid hydrodynamic model, achieving root mean square errors below 0.06 and 0.07 m/s, respectively, with a percentage bias within [Formula: see text]5%. For maximum water depth prediction, the method exhibits Nash-Sutcliffe efficiency and Pearson correlation coefficient exceeding 0.99. Similarly, for maximum velocity prediction, both metrics exceed 0.94. The computational efficiency of the proposed method outperforms the fine-grid hydrodynamic model by over 2800 times. The DenseUNet architecture developed in this study exhibits superior regression and classification performance compared to the commonly used ResUNet and UNet architectures. The proposed method is robust for a wide range of super-resolution scale factors. This paper presents an accurate and efficient surrogate model for rapid flood inundation mapping, providing valuable insights for applying deep learning-based image super-resolution methods in flood simulation.
format Article
id doaj-art-84ef819f6c214b33a180a9d6a970930d
institution Kabale University
issn 1994-2060
1997-003X
language English
publishDate 2025-12-01
publisher Taylor & Francis Group
record_format Article
series Engineering Applications of Computational Fluid Mechanics
spelling doaj-art-84ef819f6c214b33a180a9d6a970930d2025-08-20T03:40:25ZengTaylor & Francis GroupEngineering Applications of Computational Fluid Mechanics1994-20601997-003X2025-12-0119110.1080/19942060.2025.2481115Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modelingWenke Song0Mingfu Guan1Kaihua Guo2Dapeng Yu3Department of Civil Engineering, The University of Hong Kong, Hong Kong, People’s Republic of ChinaDepartment of Civil Engineering, The University of Hong Kong, Hong Kong, People’s Republic of ChinaDepartment of Civil Engineering, The University of Hong Kong, Hong Kong, People’s Republic of ChinaGeography and Environment, Loughborough University, Loughborough, UKEfficient and accurate flood inundation mapping is essential for flood risk assessment, emergency response, and community safety. The deep learning-enabled rapid flood simulation demonstrates superior computational efficiency compared to traditional hydrodynamic models. However, most deep learning-based flood mapping models currently focus on predicting the maximum water depth and face challenges in generalizing rainfall events of different durations. This paper proposes a fast flood simulation method based on image super-resolution, utilizing a novel DenseUNet architecture to predict the maximum water depth and velocity of temporal rainfall events. The proposed method integrates physical catchment characteristics to enhance the resolution of flood maps generated by the coarse-grid hydrodynamic model using the deep-learning model. The method is applied to a rural-urban catchment of the Shenzhen River in southern China. The proposed method effectively reproduces maximum water depth and velocity for test rainfall events against the fine-grid hydrodynamic model, achieving root mean square errors below 0.06 and 0.07 m/s, respectively, with a percentage bias within [Formula: see text]5%. For maximum water depth prediction, the method exhibits Nash-Sutcliffe efficiency and Pearson correlation coefficient exceeding 0.99. Similarly, for maximum velocity prediction, both metrics exceed 0.94. The computational efficiency of the proposed method outperforms the fine-grid hydrodynamic model by over 2800 times. The DenseUNet architecture developed in this study exhibits superior regression and classification performance compared to the commonly used ResUNet and UNet architectures. The proposed method is robust for a wide range of super-resolution scale factors. This paper presents an accurate and efficient surrogate model for rapid flood inundation mapping, providing valuable insights for applying deep learning-based image super-resolution methods in flood simulation.https://www.tandfonline.com/doi/10.1080/19942060.2025.2481115Flood inundation predictiondeep learningimage super-resolutionDenseUNet
spellingShingle Wenke Song
Mingfu Guan
Kaihua Guo
Dapeng Yu
Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling
Engineering Applications of Computational Fluid Mechanics
Flood inundation prediction
deep learning
image super-resolution
DenseUNet
title Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling
title_full Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling
title_fullStr Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling
title_full_unstemmed Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling
title_short Rapid flood inundation mapping by integrating deep learning-based image super-resolution with coarse-grid hydrodynamic modeling
title_sort rapid flood inundation mapping by integrating deep learning based image super resolution with coarse grid hydrodynamic modeling
topic Flood inundation prediction
deep learning
image super-resolution
DenseUNet
url https://www.tandfonline.com/doi/10.1080/19942060.2025.2481115
work_keys_str_mv AT wenkesong rapidfloodinundationmappingbyintegratingdeeplearningbasedimagesuperresolutionwithcoarsegridhydrodynamicmodeling
AT mingfuguan rapidfloodinundationmappingbyintegratingdeeplearningbasedimagesuperresolutionwithcoarsegridhydrodynamicmodeling
AT kaihuaguo rapidfloodinundationmappingbyintegratingdeeplearningbasedimagesuperresolutionwithcoarsegridhydrodynamicmodeling
AT dapengyu rapidfloodinundationmappingbyintegratingdeeplearningbasedimagesuperresolutionwithcoarsegridhydrodynamicmodeling