Data-driven models using an artificial neural network (ANN), deep learning (DL) and numerical models are applied in flood analysis of the urban watershed, which has a complex drainage system. In particular, data-driven models using neural networks can quickly present the results and be used for flood forecasting. However, not a lot of data with actual flood history and heavy rainfalls are available, it is difficult to conduct a preliminary analysis of flood in urban areas. In this study, a deep neural network (DNN) was used to predict the total accumulative overflow, and because of the insufficiency of observed rainfall data, 6 h of rainfall were surveyed nationwide in Korea. Statistical characteristics of each rainfall event were used as input data for the DNN. The target value of the DNN was the total accumulative overflow calculated from Storm Water Management Model (SWMM) simulations, and the methodology of data augmentation was applied to increase the input data. The SWMM is one-dimensional model for rainfall-runoff analysis. The data augmentation allowed enrichment of the training data for DNN. The data augmentation was applied ten times for each input combination, and the practicality of the data augmentation was determined by predicting the total accumulative overflow over the testing data and the observed rainfall. The prediction result of DNN was compared with the simulated result obtained using the SWMM model, and it was confirmed that the predictive performance was improved on applying data augmentation.

As rainfall intensity varies irregularly, urban floods can cause extreme damage. Furthermore, they are extremely nonlinear phenomena that are complex to analyze. Therefore, a classification-based real-time flood prediction model for urban areas is constructed in this study, by combining a numerical analysis model based on hydraulic theory with a machine learning model. Flood databases are constructed in advance for different rainfall scenarios using the Environmental Protection Agency-Storm Water Management Model (EPA-SWMM) and a two-dimensional inundation model. The flood depth data for each map grid are divided into five categories based on the average flood depth using the Latin hypercube sampling (LHS) and probabilistic neural network (PNN) classification techniques for higher-precision flood range prediction. A model is constructed to predict the representative cumulative volume if the observed rainfall is entered. For spatial expansion of the flood depth with the predicted representative cumulative volume, a system capable of generating a real-time flood map is constructed by linking the cumulative volume of each grid with the representative cumulative volume using linear and nonlinear regression. When compared with the results of a verified two-dimensional (2D) flood model, the developed-model goodness-of-fit is 85%, with a required run time of 1 min 12 s. Using the developed system, rainfall-induced flooding can potentially be predicted, facilitating disaster risk management and minimizing damage to property and health.

A flood hazard rating prediction model was developed that is based on a long short-term memory (LSTM) neural network and random forest. The target area was Samseong District in Seoul, which has a history of severe flooding. The Storm Water Management Model was used to generate training data for the LSTM model to predict the total overflow as the rainfall input data. Two-dimensional numerical analysis was performed to calculate inundation and flow velocity maps for training the random forest, which was used to generate a map of the predicted flood hazard rating of grid units given the total accumulative overflow of the target area. To confirm the goodness of fit, the proposed model was used to predict a flood hazard rating map for a rainfall event observed on July 27, 2011. The prediction accuracy for the flood hazard rating of each grid was 99.86% when the debris factor was considered and 99.99% when the debris factor was not considered.

Damage caused by flash floods is increasing due to urbanization and climate change, thus it is important to recognize floods in advance. The current physical hydraulic runoff model has been used to predict inundation in urban areas. Even though the physical calculation process is astute and elaborate, it has several shortcomings in regard to real-time flood prediction. The physical model requires various data, such as rainfall, hydrological parameters, and one-/two-dimensional (1D/2D) urban flood simulations. In addition, it is difficult to secure lead time because of the considerable simulation time required. This study presents an immediate solution to these problems by combining hydraulic and probabilistic methods. The accumulative overflows from manholes and an inundation map were predicted within the study area. That is, the method for predicting manhole overflows and an inundation map from rainfall in an urban area is proposed based on results from hydraulic simulations and uncertainty analysis. The Second Verification Algorithm of Nonlinear Auto-Regressive with eXogenous inputs (SVNARX) model is used to learn the relationship between rainfall and overflow, which is calculated from the U.S. Environmental Protection Agency’s Storm Water Management Model (SWMM). In addition, a Self-Organizing Feature Map (SOFM) is used to suggest the proper inundation area by clustering inundation maps from a 2D flood simulation model based on manhole overflow from SWMM. The results from two artificial neural networks (SVNARX and SOFM) were estimated in parallel and interpolated to provide prediction in a short period of time. Real-time flood prediction with the hydraulic and probabilistic models suggested in this study improves the accuracy of the predicted flood inundation map and secures lead time. Through the presented method, the goodness of fit of the inundation area reached 80.4% compared with the verified 2D inundation model.

For flood risk assessment, it is necessary to quantify the uncertainty of spatiotemporal changes in floods by analyzing space and time simultaneously. This study designed and tested a methodology for the designation of evacuation routes that takes into account spatial and temporal inundation and tested the methodology by applying it to a flood-prone area of Seoul, Korea. For flood prediction, the non-linear auto-regressive with exogenous inputs neural network was utilized, and the geographic information system was utilized to classify evacuations by walking hazard level as well as to designate evacuation routes. The results of this study show that the artificial neural network can be used to shorten the flood prediction process. The results demonstrate that adaptability and safety have to be ensured in a flood by planning the evacuation route in a flexible manner based on the occurrence of, and change in, evacuation possibilities according to walking hazard regions.

Flooding due to the increase of heavy rainfall caused even larger damage in metropolitan areas. Therefore, numerical simulation and probabilistic models have been used for flood prediction, but the methodologies for real-time flood prediction by drainage district in metropolitan areas are still not sufficient. In this study, a flood scenario database was established by using one- and two-dimensional hydraulic analysis models to propose a realtime urban flood prediction method by drainage districts in metropolitan areas. Flood prediction models were constructed for each drainage district through the Nonlinear Auto-Regressive with eXogenous inputs and Self-Organizing Map (NARX-SOM). Suggested prediction model is a data-driven model because it is based on flood database which composed with diverse flood simulation results. To evaluate the predictive capacity of the models, flood prediction was performed for the actual heavy rainfall in 2010 and 2011 that caused severe flooding in Seoul, Republic of Korea. Flood prediction models for a total of 24 drainage districts were constructed, and it was found that the goodness of fit on the flood area ranged from 68.7 to 89.7%. In terms of the expected inundation map, the predictive power was found to be high when the SOM result with 5 × 5 dimension was mainly used. Through this study, it was possible to identify the predictive capability of the NARX-SOM flood prediction algorithm. The time for inundation map prediction for each area was within two minutes, but the one- and two-dimensional flood simulation usually takes 60–80 minutes. Moreover, when the calculated goodness of fit was examined, the proposed method was found to be a practical methodology that can be helpful in improving flood response capabilities.