A Gaussian Markov Random Field Model for Total Yearly Precipitation over the African Sahel

Johan Lindström and Finn Lindgren


Centre for Mathematical Sciences
Mathematical Statistics
Lund Institute of Technology,
Lund University,
2008

ISSN 1403-9338
Abstract:
A spatio-temporal model is constructed to interpolate yearly precipitation data from 1982 to 1996 over the African Sahel. The precipitation data used in the analysis comes from the Global Historical Climatology Network.
The spatio-temporal model is based on a Gaussian Markov random field approach with AR(1)-dependence in time and a spatial component modeled using an approximation of a field with Matérn covariance. The model is defined on an irregular grid on a segment of the sphere, both avoiding the issue of matching observations to a regularly spaced grid, and handling the curvature of the Earth.
The model is estimated using a Markov chain Monte Carlo approach. The formulation as a Markov field allows for relatively efficient computations, even though the field has more than
     3 x 104 nodes.