A rapid flood inundation modelling framework using deep learning with spatial reduction and reconstruction
Introduction
Flood inundation models are commonly used for assessing flood hazard in planning, design and forecasting. Research in recent years has demonstrated an increasing interest in probabilistic flood inundation estimation using ensembles of inputs to account for uncertainty (Cooten et al., 2011; Georgas et al., 2016; Gomez et al., 2019; Wu et al., 2020). This is a challenging task due to the high computational requirements of existing flood inundation models. As a result, fast flood inundation models are needed to simulate accurate flood inundation extent and depth at high spatial resolutions with small timesteps (Gomez et al., 2019; Teng et al., 2017).
Commonly used flood inundation models can be divided into three major categories, including empirical models, hydrodynamic models and simplified conceptual models (Teng et al., 2017). Among these models, 2-dimensional (2D) hydrodynamic flood inundation models are generally considered accurate and therefore commonly used for engineering applications (Néelz and Pender, 2010). However, they are computationally expensive and therefore not suitable for applications where a large number of model runs are required (Kalyanapu et al., 2011). Though simplified conceptual models are generally more computationally efficient, they are not well suited to capturing the temporal dynamics of floods (Lhomme et al., 2008; Teng et al., 2015; Zheng, 2018).
To reduce the computational cost while maintaining desired accuracy, data-driven emulation models, mostly artificial neural networks (ANNs), have also been developed for flood inundation modelling (Chang et al., 2018; Chu et al., 2019; Xie et al., 2020). One drawback of traditional ANNs, such as multilayer perceptron models, is that they cannot take advantage of serial correlation that is present in the input time series data used to specify flow and water level boundary conditions (Chu et al., 2019; Kratzert et al., 2018). This adds unnecessary computational complexity and uncertainty to the simulations, which could be avoided if explicit account is given to the serial correlation contained in time series input data (Hsu et al., 1997; Kratzert et al., 2018; Kumar et al., 2004).
In addition, current implementations of ANNs for simulating flood inundation typically involve developing models to simulate flood depth (or water level) at each grid cell within the modelling domain. This often results in either a large number of ANN models being developed or a very complex ANN model, which increases the difficulty of model development. For example, Chu et al. (2019) and Xie et al. (2020) developed as many as 19,448 ANN models to model flood depth for an area of 7.8 km2; Kabir et al. (2020) modelled flood inundation depth for 581,061 grid cells using a single convolutional ANN, which has over 297 million parameters. Such approaches also present difficulties when applied to data-sparse regions where there are very limited flood inundation data for model development (Xie et al., 2020). Since flood inundation models simulate continuous water surfaces, water levels at one grid cell are spatially correlated to those in its neighbouring cells. Current ANN-based modelling approaches ignore the redundancy in flood inundation information associated with this spatial correlation, leaving room for potential improvement of modelling efficiency.
In this study, a new flood inundation modelling framework is developed to address the two key issues listed above.
First, deep learning (DL) models are used to accommodate the serial correlation embedded in the multivariate time series inputs for flood inundation modelling. Among different DL models, recurrent neural networks (RNNs) were designed to handle sequential inputs and have been found to be effective for time series modelling (Connor et al., 1994; Hsu et al., 1997; Kumar et al., 2004). Among different RNNs, the long short-term memory (LSTM) architecture has been increasingly applied to discharge modelling in recent years, as LSTM has been found to out-perform the traditional RNNs for time series modelling in a number of studies (Gers et al., 1999; Kratzert et al., 2018; Sahoo et al., 2019). Therefore, the LSTM architecture is adopted for this study. In addition to the traditional LSTM layer, a fully connected input selection layer is included in the LSTM model architecture. Consequently, input selection is incorporated into the model calibration process rather than being a separate step of the model development process, and this significantly simplifies the model development process.
Second, in order to reduce the information redundancy in the current one-model-per-grid approach used in traditional ANN-based flood inundation modelling approaches, a spatial reduction and reconstruction (SRR) method is developed to characterise the flood inundation surface in a more parsimonious manner. This SRR method allows the whole inundation surface across the model domain to be represented by water levels at selected representative locations. This significantly reduces the number of models that need to be developed to represent the flood inundation surface, thus reducing the computational cost (and the number of parameters) involved in the simulation. The representative locations are selected to make the best use of flood information in data-rich regions as this has been found to improve the accuracy of the model simulations (Xie et al., 2020). Finally, the accuracy and efficiency of this framework (which hereafter is referred to as “SRR-DL”) are demonstrated via a real-world case study in Queensland, Australia.
The paper is structured as follows. A detailed description of the SRR-DL framework is provided in Section 2, including the SRR method and the design of the DL model architecture. Then, the application of the framework is presented in Section 3, including the introduction of the case study and detailed modelling process applied. Section 4 includes results obtained and discussion of main findings, and this is followed by conclusions in Section 5.
Section snippets
Methodology
The SRR-DL framework consists of two major components:
- 1.
The SRR method is developed and used to select representative locations (RLs) within the modelling domain that can be used to estimate flood inundation behaviour over the whole simulation period based on water levels simulated at these locations.
- 2.
DL models are used to simulate water levels at each of the selected representative locations based on the time series of model inputs (e.g. mainstream and tributary inflows, and ocean water levels
Study area and hydrodynamic model available
The application of the SRR-DL framework to flood inundation modelling is demonstrated using the Burnett River system downstream of Paradise Dam in Queensland, Australia. Fig. 3 shows the location and model domain of the system. The river reach in the study area is approximately 145 km long and ends at the Coral Sea at Burnett Heads. A 2D hydrodynamic model was developed for the system using the TUFLOW modelling suite (Huxley and Syme, 2016). The modelling area extents from the Paradise Dam to
Input selection evaluation
The efficacy of selected inputs at various RLs is illustrated in Fig. 5. Fig. 5a shows the grouping of the 500 LSTM models developed for the 125 RLs based on their distance to the downstream boundary. Each group includes two to seven LSTM models. The 13 concurrent inflows (CFs) are distributed inputs and the middle point of each concurrent inflow location is indicated in the figure. Fig. 5b shows the IFR indices for the 15 input variables, including the main inflow (Q), the lower boundary water
Conclusion
The SRR-DL (Spatial Reduction and Reconstruction – Deep Learning) framework proposed here is found to provide fast time series estimates of the flood inundation without appreciable loss of accuracy. The framework comprises two main components: one component (SRR) reduces the computational burden by taking advantage of spatial correlations across the model domain to reduce the number of locations where models are required, and the other component (Long Short-Term Memory, LSTM, model) is a deep
Software and data availability
The spatial reduction and reconstruction method, as presented in the associated MethodsX paper, is programmed using the Python programming language (version 3.7) and is freely available from https://github.com/yuerongz/SRR-method.
The development of DL models relies on the Python programming language (version 3.7) and open source libraries including PyTorch, Numpy, Scipy, Fiona, GDAL, Geopandas and NetCDF4. Data and program used for DL model development are available from //doi.org/10.26188/13122623.v1
Editorial conflict of interest statement
Given her role as Associate Editor of Environmental Modelling & Software, Wenyan Wu was not involved in the peer-review of this article and had no access to information regarding its peer-review. Full responsibility for the editorial process for this article was delegated to Editor-in-Chief, Daniel P. Ames.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
We thank Hydrology and Risk Consulting (HARC) for providing the TUFLOW 2D hydrodynamic model configuration and SunWater for their permission to use the Burnett River as a case study. We also thank BMT for providing the TUFLOW license to conduct the TUFLOW simulations. This research was undertaken using the LIEF HPC-GPGPU Facility hosted at The University of Melbourne. This Facility was established with the assistance of LIEF Grant [grant number LE170100200]. At last, for a range of open source
References (35)
- et al.
Survey and critique of techniques for extracting rules from trained artificial neural networks
Knowl. Base Syst.
(1995) - et al.
Review and comparison of methods to study the contribution of variables in artificial neural network models
Ecol. Model.
(2003) - et al.
Skill of ensemble flood inundation forecasts at short-to medium-range timescales
J. Hydrol.
(2019) - et al.
A deep convolutional neural network model for rapid prediction of fluvial flood inundation
J. Hydrol.
(2020) - et al.
Assessment of GPU computational enhancement to a 2D flood model
Environ. Model. Software
(2011) - et al.
Calibration and validation of neural networks to ensure physically plausible hydrological modeling
J. Hydrol.
(2005) - et al.
A probabilistic method for assisting knowledge extraction from artificial neural networks used for hydrological prediction
Math. Comput. Model.
(2006) - et al.
A survey of deep neural network architectures and their applications
Neurocomputing
(2017) Deep learning in neural networks: an overview
Neural Network.
(2015)- et al.
The regional oceanic modeling system (ROMS): a split-explicit, free-surface, topography-following-coordinate oceanic model
Ocean Model.
(2005)