Spatial prediction of flood-prone areas using geographically weighted regression

Abstract An important non-structural solution in flood management is susceptibility mapping, which identifies the likelihood of flood occurrence in an area. Although various models have been applied in flood susceptibility mapping with different successes, Geographically Weighted Regression [GWR] has not been sufficiently tested despite its effective advantages in interpreting spatially heterogeneous relationships. To test GWR's effectiveness in flood susceptibility modelling, this study included 16 morphometric parameters as the explanatory variables, and historical flood occurrence as the dependent. Multicollinearity was eliminated based on Variance Inflation Factor [VIF], which resulted in six screened parameters: stream order, drainage texture, relief ratio, bifurcation ratio, topographic wetness index, and topographic position index. Five tests were carried out, with the first involving direct inputs of the VIF-screened variables into the GWR modelling process. The other four tests incorporated morphometric parameter normalisation into 1-to-5 ranking scores according to Ordinary Least Square [OLS] coefficients or literature using all or only the VIF-screened parameters. The best-performing model was the first test, indicating that direct input of the screened parameters was the ideal modelling process. This test had the lowest corrected Akaike Information Criterion (160.01), the highest percentage of deviance explained (46.18%), lowest spatial autocorrelation of residuals (0.1122) and transformed residuals (0.2827), highest success accuracy (91.24%), and second-best prediction accuracy (75.15%). These findings show that accounting for spatial variation improved global flood model performance. The results also show that GWR may have the potential for better flood susceptibility mapping when compared with other traditional models such as non-spatial logistic regression and frequency ratio.