Blog

Sunday, February 10th

I will present a talk “Can one geostatistical model practice many arts with success?” at the 5 th Spatial
Statistics conference, which will be held in Sitges (near to Barcelona), Spain, from 10 – 13 July 2019,
https://www.elsevier.com/events/conferences/spatial-statistics.
Here is an abstract of the talk: "One man cannot practice many arts with success" - Plato, Republic, 374 A - but can one geostatistical model? We discuss challenges to the statistical data analysis when one or more of the following classic geostatistical model assumptions are violated: data stationarity, data Gaussianity, knowing true covariance function, random data locations, constant measurement error, constant data support, and error-free mutually uncorrelated covariates. Large samples are increasingly common, particularly with spatial data collected using GPS-enabled devices. Ideally, increasing the number of data samples should have a positive effect on the statistical inference. In practice, data quality often gets worse, and the number of violated assumptions increases. Fitting models to big data based on the gold standard for small/medium size data is problematic. A reliable geostatistical model can be fitted using reasonable approximations even when some, or all, of the statistical assumptions above are partially violated . We show examples of such approximations, including the divide and conquer approach, empirical Bayesian methodology for generalized covariance functions estimation, individual measurement errors usage, flexible data transformation, adjustable search neighborhood, and the principal component regression algorithm. Another useful approximation for modeling data over a large portion of Earth is the usage of chordal distances. Although there are many GIS applications where the best distance metric is geodesic, we show that data interpolation using chordal distance metrics is accurate and also very efficient. Additionally, long computation time required to find optimal model parameters is unacceptable to many statistical software users, limiting access to optimal spatial data analysis to theoretical researchers. Therefore, we discuss several examples of useful approximation techniques which reduce the time of computation without a significant compromise of the model output. Come find out whether one geostatistical model can practice many arts with success. Sunday, July 7th A paper Krivoruchko K. and Gribov A. (2019) Evaluation of the empirical Bayesian kriging methodology is published in the Spatial Statistics Journal, Volume 32, August 2019. Link: https://authors.elsevier.com/c/1ZKVt7su79m-18 is valid till August 22, 2019. Here is an abstract: Using Bayesian bootstrap as the informative prior distribution is a key feature of the empirical Bayesian kriging method. We evaluate correctness of the method using two large simulation experiments. We also compare empirical Bayesian kriging performance with several popular geostatistical models implementedin R using two real datasets: the average annual precipitation in South Africa and the daytime land surface temperature satellite data. Results show that empirical Bayesian kriging performs better than other predictors, and increasingly so with data complexity.
Konstantin Krivoruchko