A hybrid deep neural network with statistical learning algorithms for flood susceptibility modeling

Abstract

Flood, with its environmental impact, is a naturally destructive process usually causes severe damage. Therefore, the determination of the areas susceptible to flood by the latest tools, which can render precise estimations, is essential to mitigate this damage. In this study, it was attempted to evaluate flood susceptibility in Lorestan, Iran using a novel hybrid approach including Deep Neural Network (DNN), Frequency Ratio (FR), and Stepwise Weight Assessment Ratio Analysis (SWARA). For this purpose, a geospatial database of floods, including 142 flood locations and 10 flood influencing variables, was used to predict the flood susceptibility areas. FR and SWARA were applied to weigh and score the flood influencing variables, while DNN, which is an excellent tool of machine learning and artificial intelligence, was used to prepare the inference flood pattern. The performance of the models was checked by the area under the curve (AUC), receiver operating characteristic (ROC) curve, and various statistical tests. The outputs indicated that both of the proposed algorithms, DNN-FR and DNN-SWARA, were able to estimate the future flood zones with a precision of over 90%. The outputs also confirmed that although the two algorithms had a high goodness-of-fit and prediction accuracy, the DNN-FR (AUC = 0.953) outperformed the DNN-SWARA (AUC = 0.941). Thus, the DNN-FR algorithm was proposed to be applied as a more reliable and accurate tool for spatial estimation of flood zones.