Are you an EPFL student looking for a semester project?
Work with us on data science and visualisation projects, and deploy your project as an app on top of Graph Search.
In a world of increasing exposure of populations to natural hazards, the mapping of extreme rainfall remains a key subject of study. Such maps are required for both flood risk management and civil engineering structure design, the challenge being to take into account the local information provided by point rainfall series as well as the necessity of some regional coherency. Two approaches based on the extreme value theory are compared here, with an application to extreme rainfall mapping in West Africa. The first approach is a local fit and interpolation (LFI) consisting of a spatial interpolation of the generalized extreme value (GEV) distribution parameters estimated independently at each station. The second approach is a spatial maximum likelihood estimation (SMLE); it directly estimates the GEV distribution over the entire region by a single maximum likelihood fit using jointly all measurements combined with spatial covariates. Five LFI and three SMLE methods are considered, using the information provided by 126 daily rainfall series covering the period 1950-1990. The methods are first evaluated in calibration. Then the predictive skills and the robustness are assessed through a cross validation and an independent network validation process. The SMLE approach, especially when using the mean annual rainfall as covariate, appears to perform better for most of the scores computed. Using the Niamey 104 year time series, it is also shown that the SMLE approach has the capacity to deal more efficiently with the effect of local outliers by using the spatial information provided by nearby stations.
Mario Paolone, Vladimir Sovljanski
Sabine Süsstrunk, Majed El Helou, Kaan Okumus
Volkan Cevher, Grigorios Chrysos, Fanghui Liu