Abstract
Modeling and forecasting spatiotemporal patterns of precipitation is crucial for managing water resources and mitigating water-related hazards. Globally valid spatiotemporal models of precipitation are not available. This is due to the intermittent nature, non-Gaussian distribution, and complex geographical dependence of precipitation processes. Herein we propose a data-driven model of precipitation amount which employs a novel, data-driven (non-parametric) implementation of warped Gaussian processes. We investigate the proposed warped Gaussian process regression (wGPR) using (i) a synthetic test function contaminated with non-Gaussian noise and (ii) a reanalysis dataset of monthly precipitation from the Mediterranean island of Crete. Cross-validation analysis is used to establish the advantages of non-parametric warping for the interpolation of incomplete data. We conclude that wGPR equipped with the proposed data-driven warping provides enhanced flexibility and—at least for the cases studied– improved predictive accuracy for non-Gaussian data.
Highlights
Climate change combined with changes in land use is causing increased frequencies of drought and flooding events in many parts of the world
There is a strong interest in the application of Gaussian processes to model spatial and spatiotemporal data [61,62,63]
In the case of data exhibiting non-Gaussian distribution, nonlinear transforms are applied to the observations in order to allow the application of Gaussian assumptions and methods
Summary
Climate change combined with changes in land use is causing increased frequencies of drought and flooding events in many parts of the world. Certain areas, including the Mediterranean basin, have been characterized as “climate change hot spots” [6] In such regions the interplay of climate change and changes in land use is crucial for water resource availability [4]. The total amount of precipitation received by an area over a specific time window is often modeled by means of parametric, non-Gaussian, probability distributions. We show that non-parametric warped Gaussian process regression (wGPR) can be used to model the spatial distribution of non-Gaussian variables such as precipitation. The remainder of this paper is structured as follows: Section 2 presents the proposed wGPR methodology which involves Gaussian anamorphosis using the kernel-based CDF, spatial interpolation (prediction) of the normalized process employing standard GPR, and generation of the predictive distribution of precipitation by inverting the warping transformation.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have