| Last Updated:
Creating DOI. Please wait...
Prediction of flood quantiles at ungauged sites has been investigated using several nonparametric regression methods including: local regression based on regions of influence, neural networks and generalized additive models (GAM). These methods were used to describe the relationship between run-off variables and catchment descriptors to predict flood quantiles. Previous work reported the presence of spatial correlation in the residuals for these models. To this end, this study proposes and investigates ways of incorporating spatial components. An L-moments regression technique (LRT) is developed to predict L-moments of target sites and flood quantiles are derived by aggregating quantiles from multiple candidate distributions. The predictive power of the proposed methods is evaluated on a large database of Canadian rivers using cross-validation. The results are examined inside different provinces and hydrological regions to assess the behaviour of the methods. The results show that GAM and local regression using respectively thin plate spline and kriging provide the best predictive powers among the considered methods. Additionally, the LRT method is found to improve prediction power over the well-known index-flood model and has similar results to quantile regression techniques (QRT) when using the same nonparametric regression approaches.