ABSTRACT
This study aimed to assess the effective spatial characteristics of input features by using physics-informed, machine learning (ML)-based inundation forecasting models. To achieve this aim, inundation depth data were simulated using a numerical hydrodynamic model to obtain training and testing data for these ML-based models. Effective spatial information was identified using a back-propagation neural network, an adaptive neuro-fuzzy inference system, support vector machine, and a hybrid model combining support vector machine and a multi-objective genetic algorithm. The conventional average rainfall determined using the Thiessen polygon method, raingauge observations, radar-based rainfall data, and typhoon characteristics were used as the inputs of the aforementioned ML models. These models were applied in inundation forecasting for Yilan County, Taiwan, and the hybrid model had the best forecasting performance. The results show that the hybrid model with crucial features and appropriate lag lengths gave the best performance.
Editor A. Castellarin Associate Editor O. Kisi
Editor A. Castellarin Associate Editor O. Kisi
Acknowledgements
We thank the editor and anonymous reviewers for their comments and suggestions to improve the quality of this manuscript. We extend special thanks to Professor Yasuto Tachikawa from the Department of Civil and Earth Resources Engineering, Graduate School of Engineering, Kyoto University, who discussed the process of model development and provided positive comments on this study. We also thank the personnel and students at the Hydrology and Water Resources Research Laboratory, Kyoto University.
Disclosure statement
No potential conflict of interest was reported by the authors.
Supplementary material
Supplemental data for this article can be accessed online at https://doi.org/10.1080/02626667.2022.2092406