Non-stationary geostatistical modeling of daily rainfall over complex topography

In mountains, topography-atmosphere interactions generate orographic effects which make windward slopes usually wetter than leeward ones, and highlands wetter than lowlands. The transitions between wet and dry areas can occur within few kilometers, which creates strong horizontal gradients of rainfall statistics such as frequency of occurrence, daily mean accumulation, or extreme intensities. This spatial variability of rainfall statistics breaks the hypothesis of stationarity on which rely most geostatistical models that are used for the spatial analysis of rainfall data. Using stationary models to process non-stationary data can lead to a degraded performance in spatial prediction (e.g., mapping rainfall by interpolation of sparse rain gauge observations) and to unrealistic rainfall features in simulations (e.g., emulation of synthetic rain fields using a stochastic rainfall generator). 
 
To overcome these limitations, we present in this work a fully non-stationary trans-Gaussian geostatistical model dedicated to the spatial analysis of daily rainfall over complex topography. This model allows not only for a non-stationary marginal distribution of daily rainfall accounting for rainfall intermittency and non-Gaussian intensity, but also for a non-stationary covariance structure of Matérn type that models the spatial dependencies.

The model is tested for the Island of Hawai‘i (State of Hawaii, USA) where rainfall gradients are amongst the strongest on Earth and can reach 1000 mm.year^-1/km. To make our model operable in practice, we designed a procedure to infer model parameters from rain gauge observations that are freely available in near-real-time on the Hawai‘i Climate Data Portal. Model assessment demonstrates good skills at reproducing the spatial variability of daily rainfall occurrence, intensity distribution and spatial dependencies.