Flood susceptibility mapping using ANNs: a case study in model generalization and accuracy from Ontario, Canada

Abstract

AbstractAccurate flood susceptibility mapping (FSM) is critical for mitigating the environmental, social and economic consequences of floods. The influence of model generalizability onto new watersheds, and the impact of arbitrarily selecting a small subset of flooded and nonflooded locations are current major knowledge gaps in FSM research impacting predictive accuracy. As such, this study conducts an assessment of machine learning models – (i) an Artificial Neural Network - Synthetic Minority Oversampling Technique (ANN-SMOTE) hybrid ensemble with (ii) knowledge-based Analytical Hierarchy Process (AHP) and (iii) diversity-based Shannon Entropy approaches. The ANN-SMOTE, AHP and Entropy models were trained and tested on the Don River watershed in Ontario, Canada, with Overall Accuracy (OA) results of 0.549, 0.404 and 0.452, respectively. ANN-SMOTE’s predictive accuracy remained high when it was tested on four independent watersheds from southern Ontario, indicating strong generalization ability. To simulate the commonly used flood point inventory approach, the number of training samples was reduced by a factor of a 1000, which resulted in a 28% decrease in accuracy. The high performance and generalization potential of the ANN-SMOTE model demonstrate its utility and versatility for future FSM studies, and as a support tool in flood risk management decision making.

Keywords

• SMOTE
• AHP
• entropy
• floodplain
• GIS