Urban flooding is a complex natural hazard event that incorporates climate change impacts with urban planning and developing challenges, requiring comprehensive strategies for mitigation and adaptation. Flood susceptibility mapping is one of the first steps in an appropriate strategy to reduce economic disruption and damage to urban environments due to flooding. This paper proposes a family of new deep neural networks, namely “deep abstract networks” (DANet) algorithm, which has not been conducted earlier on the susceptibility assessment worldwide, to be trained for producing reliable urban flood susceptibility maps, using Sanandaj City, Iran, as an example. In this procedure, 174 urban and 174 non-urban flood locations are considered in tandem with 19 flood factors prioritized using the reliefF attribute evaluation (RAE) feature selection technique. We determine the goodness-of-fit and prediction accuracy of our models using sensitivity, specificity, accuracy, kappa, root mean square error (RMSE), mean absolute error (MAE), and area under the curve (AUC). Furthermore, the new proposed deep learning algorithm is compared to the five state-of-the-art benchmark learning algorithms, i.e., Convolutional Neural Network (CNN), Support Vector Machine with Linear (SVM-Linear) and with radial basis function (SVM-RBF), Artificial Neural Network-Multi-Layer Perceptron (ANN-MLP), and Logistic Regression (LR). Here, land use, building density, distances to buildings, rainfall, and distances to passages are the five most influential factors in urban flood occurrence in the study area. The DANet algorithm achieves RMSE = 0.535, AUCmodel = 0.811, and AUCmap = 0.840, and thus outperforms the ANN-MLP, SVM-RBF, SVM-Linear, LR and CNN algorithms as an excellent alternative algorithm for managing areas prone to urban flooding with caution.
