Regional flood inundation nowcast using hybrid SOM and dynamic neural networks Li-Chiu Chang a,⇑ , Hung-Yu Shen a , Fi-John Chang b,⇑ a Department of Water Resources and Environmental Engineering, Tamkang University, New Taipei City 25137, Taiwan, ROC b Department of Bioenvironmental Systems Engineering, National Taiwan University, Taipei 10617, Taiwan, ROC article info Article history: Received 1 April 2014 Received in revised form 1 July 2014 Accepted 18 July 2014 Available online 30 July 2014 This manuscript was handled by Geoff Syme, Editor-in-Chief Keywords: Artificial neural network (ANN) Self-organizing map (SOM) Recurrent configuration of nonlinear autoregressive with exogenous inputs (R-NARX) Flood inundation map Regional flood forecasting model summary This study proposes a hybrid SOM–R-NARX methodology for nowcasting multi-step-ahead regional flood inundation maps during typhoon events. The core idea is to form a meaningful topology of inundation maps and then real-time update the selected inundation map according to a forecasted total inundated volume. The methodology includes three major schemes: (1) configuring the self-organizing map (SOM) to categorize a large number of regional inundation maps into a meaningful topology; (2) building a recurrent configuration of nonlinear autoregressive with exogenous inputs (R-NARX) to forecast the total inundated volume; and (3) adjusting the weights of the selected neuron in the constructed SOM based on the forecasted total inundated volume to obtain a real-time adapted regional inundation map. The pro- posed models are trained and tested based on a large number of inundation data sets collected in an inun- dation-prone region (270 km 2 ) in the Yilan County, Taiwan. The results show that (1) the SOM–R-NARX model can suitably forecast multi-step-ahead regional inundation maps; and (2) the SOM–R-NARX model consistently outperforms the comparative model in providing regional inundation maps with smaller forecast errors and higher correlation (RMSE < 0.1 m and R 2 > 0.9 in most cases). The proposed modelling approach offers an insightful and promising methodology for real-time forecasting 2-dimensional visible inundation maps during storm events. Ó 2014 Elsevier B.V. All rights reserved. 1. Introduction Floods have posed an increasing threat worldwide over recent decades and the corresponding disasters have also noticeably increased due to urbanization and extensively subsequent devel- opment in the vicinities of rivers. Obviously, flood control deserves adequate attention and demands to develop appropriate technolo- gies for early warning and disaster prevention. Effective real-time flood forecasting models could help mitigate flood disasters through the rapid dissemination of inundation information regard- ing threatened areas. Flood forecasting in urban areas, however, is a great challenge because of the complex interactions and disrup- tions associated with non-riverine urban flooding and the lack of high-resolution hydro-geomorphological data. In recent years, dis- tributed numerical models have become an increasingly attractive solution of flood inundation estimation owing to the advances in numerical modelling techniques and computing power (Bates et al., 2010; Dottori and Todini, 2011; Frank et al., 2012; Hunter et al., 2007; Kang, 2009; Mason et al., 2009; Neal et al., 2012; Yamazaki et al., 2011). However, these models commonly require various types of hydro-geomorphological monitoring data and huge computational time of hydrodynamic calculations, which result in lacking the accessibility of implementation and prohibit- ing real-time applications. There is a continuous need to conduct in-depth research in flood inundation disaster management through implementing the latest scientific tools such as soft com- puting and numerical simulation modelling for developing disaster management scenarios. To model real-time flood inundation forecasts, more flexible modelling techniques that can effectively extract the complex input–output relationship of a system in an adaptive manner have to be explored and conducted. Applications of soft computing tech- niques in the field of time series forecasting have been available for decades. Among these modelling techniques, artificial neural net- works (ANNs) have been successfully applied in hydrosystem stud- ies and obtained many encouraging results (Abrahart et al., 2012; Besaw et al., 2010; Chang et al., 2008; Chen and Chang, 2009; http://dx.doi.org/10.1016/j.jhydrol.2014.07.036 0022-1694/Ó 2014 Elsevier B.V. All rights reserved. ⇑ Corresponding authors. Address: No. 151, Yingzhuan Road, Tamsui District, New Taipei City 25137, Taiwan, ROC. Tel.: +886 2 26215656x3269 (L.-C. Chang). Address: No. 1, Sec. 4, Roosevelt Road, Da-An District, Taipei 10617, Taiwan, ROC. Tel.: +886 2 33663452 (F.-J. Chang). E-mail addresses: changlc@mail.tku.edu.tw (L.-C. Chang), changfj@ntu.edu.tw (F.-J. Chang). Journal of Hydrology 519 (2014) 476–489 Contents lists available at ScienceDirect Journal of Hydrology journal homepage: www.elsevier.com/locate/jhydrol