[Forum SIS] Seminario Dr Patrick Vetter

[Francesco Finazzi]

Martedì 7 maggio alle ore 10 presso il Dipartimento di Ingegneria
dell'Università di Bergamo, sede di Dalmine,

il Dr Patrick Vetter, dell'European University Viadrina Frankfurt (Oder)

terrà un seminario dal titolo:

"Efficient Approximation of the Spatial Covariance Function - Analysis
of Atmospheric Carbon Dioxide Concentrations".

Tutti gli interessati sono cordialmente invitati a partecipare ed a
estendere l'invito a chiunque possa essere interessato. Segue abstract
dell'intervento.

Abstract: With the emergence of remote sensing data on environmental
quantities obtained from satellites and the increasing spatial and
temporal resolution associated with them, traditional statistical
spatiotemporal models are reaching their computational limits.
Gaussian linear mixed effects models have been widely used in the
spatial statistical analysis of environmental processes. However,
parameter estimation and spatial predictions involve the inversion and
determinant of the spatial covariance matrix of the data process, that
is the size of the dataset. Typically remotely sensed datasets contain
observations of the order of tens or hundreds of thousand on a single
day, which quickly leads to bottlenecks in terms of computation speed
and requirements in working memory.
Therefore techniques for reducing the dimension of the problem are
required. Recent research has been focused on developing approaches to
efficiently approximate the spatial covariance function, whereas one
possibility is to use reduced rank methods (e.g. Fixed Rank Kriging)
and another is to introduce sparseness to the covariance matrix
(Covariance Tapering) by accounting for the fact, that the range of
spatial dependence is (conditionally) finite. The present work will
analyze the efficiency of these approaches in approximating the
spatial covariance function in a real dataset of remotely sensed
carbon dioxide concentrations, obtained from the Atmospheric Infrared
Sounder of NASA’s "Aqua" satellite on the 1st of May 2009, which
consists of 12,846 observations on the globe.
Through a series of cross-validation experiments, with varying choices
for the parameters specifying the degree of the approximation, the
predictive performance is evaluated and the computation time and
storage requirements are monitored for kriging predictions on a
regular grid with 250,000 prediction locations. Furthermore possible
temporal extensions to these approaches are examined.

--
Francesco Finazzi
Assistant Professor

University of Bergamo
Dept. of Management, Economics and Quantitative Methods

http://www.unibg.it/pers/?francesco.finazzi