Flood hazard mapping in western Iran: assessment of deep learning vis-à-vis machine learning models

On March 25, 2019, widespread flood events occurred across Iran’s provinces and set a new record for socioeconomic losses and casualties. In hindsight, it opened an engrossed area of research for flood hazard mapping, something that has previously been pursued yet not embraced as must they should. In pursuit of reconciling the decision-makers and authorities with the successful application of machine/deep learning models in pattern extraction and detection of hotspots, we employed a deep learning model named Deep Belief Network (DBN), and then it was hybridized using particle swarm optimization (DBN-PSO) and genetic algorithm (DBN-GA). Their results were compared to Random Forest (RF) and Support Vector Machine (SVM) that have been known as benchmark models in the field of flood susceptibility. The area under the receiver operatic characteristic curve (AUC) and True Skill Statistic (TSS) were used for performance assessment through three sample partitioning replicates, which further indicated models’ predictive performance and robustness. Results revealed that DBN-PSO was the best model in terms of prediction performance (AUCmean = 0.957, TSSmean = 0.748) and robustness (RAUC = 0.6%, RTSS = 0.8%). Furthermore, all the standalone DBN and its hybrid models outperformed benchmark models including RF (AUCmean = 0.911, TSSmean = 0.714, RAUC = 1.5%, RTSS = 2.1%) and SVM (AUCmean = 0.899, TSSmean = 0.692, RAUC = 2.1%, RTSS = 2.3%). In addition, drainage density was the most important factor in flood susceptibility modeling. Accordingly, about 13.5% of the region was addressed as the high hazard zone of flood occurrence based on the DBN-PSO model, which should be considered for further pragmatic actions.