Comparison and evaluation of various methods to estimate the variogram in geostatistics
2 Ecological Modelling, BayCEER, University of Bayreuth, Dr.-Hans-Frisch-Straße 1–3, 95448 Bayreuth, Germany, University of Bayreuth
3 Soil Physics Group, BayCEER, University of Bayreuth, 95440 Bayreuth, Germany
O 2.1 in Research in its Prime: First Results of Ongoing Research
10.10.2013, 09:15-09:30, H6, GEO
Several basic statistical methods can be used to describe the structure of spatial similarity in environmental data. The overall goal is to predict unknown values from the known data by fitting an appropriate model to it. In the present study we compare the prediction accuracy of different weighted least-squares methods, the maximum-likelihood estimator and several ways of cross validation. Using the R environment [R Core Team, 2013] and the package RandomFields [Schlather u.a., 2013] a virtual random field is created and sampled. This sampled data are randomly split into a training and a validation data set. Subsequently, one of the methods mentioned above is chosen to estimate the variogram or the parameters of the random field. By testing the estimated model on the independent validation data set we are able to access its prediction accuracy via the validation error. Our results show that the smallest validation error is achieved by the maximum-likelihood estimator. However, the cross validation that is hardly used in geostatistics so far, provides similarly good results. This method avoids the estimation of a variogram and is therefore less error-prone. Irrespective of the applied method, sampling design is crucial: Data from irregular sampling seem to capture the specific characteristics of a random field much better than data obtained from a regular grid.
Export as iCal:
