ABSTRACT
Flood susceptibility projections relying on standalone models, with one-time train-test data splitting for model calibration, yields biased results. This study proposed novel integrative flood susceptibility prediction models based on multi-time resampling approaches, random subsampling (RS) and bootstrapping (BT) algorithms, integrated with machine learning models: generalized additive model (GAM), boosted regression tree (BTR) and multivariate adaptive regression splines (MARS). RS and BT algorithms provided 10 runs of data resampling for learning and validation of the models. Then the mean of 10 runs of predictions is used to produce the flood susceptibility maps (FSM). This methodology was applied to Ardabil Province on coastal margins of the Caspian Sea which faced destructive floods. The area under curve (AUC) of receiver operating characteristic (ROC) and true skill statistic (TSS) and correlation coefficient (COR) were utilized to evaluate the predictive accuracy of the proposed models. Results demonstrated that resampling algorithms improved the performance of Standalone GAM, MARS and BRT models. Results also revealed that Standalone models had better performance with the BT algorithm compared to the RS algorithm. BT-GAM model attained superior performance in terms of statistical measures (AUC = 0.98, TSS = 0.93, COR = 0.91), followed by BT-MARS (AUC = 0.97, TSS = 0.91, COR = 0.91) and BT-BRT model (AUC = 0.95, TSS = 0.79, COR = 0.79). Results demonstrated that the proposed models outperformed the benchmark models such as Standalone GAM, MARS, BRT, multilayer perceptron (MLP) and support vector machine (SVM). Given the admirable performance of the proposed models in a large scale area, the promising results can be expected from these models for other regions.
ABSTRACT
Bivariate frequency analysis of extreme rainfall and runoff is crucial for water resource planning and management in a river basin. This study is aimed at accounting for uncertainties in bivariate analysis of extreme rainfall-runoff frequency in the Taleghan watershed, one of the major watersheds in northern Iran, using copulas. Two types of paired rainfall and runoff data, including annual maximum series (AMS) and peaks over threshold (POT) are adopted to investigate the uncertainties that arose due to the input data. The Cramer von-Mises goodness-of-fit test and Akaike information criteria (AIC) reveal that the Student's t copula is the best-fit copula for PAMS-QAMS with Gaussian-Pearson III (P3) margins, while the Plackett copula is the best-fit copula for PPOT-QPOT with generalized Pareto (GPAR-GPAR) margins. A nonparametric bootstrapping method for sampling from p-level curves is established to investigate the effects of univariate and bivariate models selection and uncertainties induced by input data. The results indicated that the sampling uncertainty reduces POT data compared to AMS data due to the increased sample size. However, the parameterization uncertainty of the POT data increases because of the weaker dependence structure between rainfall and runoff for the POT data. The results also reveal that the larger sampling uncertainties are associated with higher p-level curves for both AMS and POT data which are induced by lower data density in the upper tail. For the study area, the input-data uncertainty is most significant in bivariate rainfall-runoff frequency analysis and quantile estimation, while the uncertainty induced by probabilistic model selection is least significant.
