Performance benchmarking on several regression models applied in urban flash flood risk assessment.

To evaluate the performances of regression models applied in the urban flash flood risk assessment, the historical urban flash flood occurrences points were used to build the Voronoi polygon networks for calculating Ripley's K values which can be adopted to be the risk value and the predictands in regression. The first level risk indicators of hazard, vulnerability, sensitivity and exposure risk factors in the risk assessment, as well as the sensitivity subordinate indicators of imperviousness and terrain factor, were listed to be the predictors in the regression model. Subsequently, methods of the linear regression equation (LRE), nonlinear regression power-form function (PF) and a simplified power-form function (SPF), as well as support vector machine (SVM) model and random forests (RF) model, were all nominated for the performance evaluation and comparison of the fitness of their regression relationships between the predictors and the predictands. With the support of samples, the benchmarking firstly demonstrated the SPF is the best of the regression equation; but the full PF equation cannot be figured out on account of the sample data deficiency. The SVM model behaves better than the regression equations of SPE and LRE, while the SVM of nonlinear polynomial kernel function is slightly better than that of the nonlinear Gaussian kernel function. Above all, the RF model performed perfectly in the regression fitting, which the relative bias index is − 0.009 and the relative mean squared error is 0.0773. Meanwhile, it mostly resolves the problems of overfitting, outliers and noise in regression. The variable importance (VI) evaluated by the RF model indicated that the top four important risk factors are the imperviousness, terrain factor, vulnerability, and exposure factor, which the VI index value is 0.38, 0.16, 0.11 and 0.1, respectively. Unexpectedly, the hazard factor appears to be the least important factor with a VI value of 0.04. The homogeneity of invariable hazard being preserved in regional climate background makes the hazard a minor role in risk contribution. The model performance evaluation demonstrated the artificial intelligence RF model should be recommended to be the common-use model for aftermath meteorology-related risk assessment. On the other hand, the VI analysis tools of RF were also recognized to be a welcome toolbox items for the risk analysis. [ABSTRACT FROM AUTHOR]