Author(s): |
E. Meerschman, Ghent University (BE)
G. Pirot, University of Neuchâtel (CH)
G. Mariethoz, The University of New South Wales (AU)
J. Straubhaar, University of Neuchâtel (CH)
M. Van Meirvenne, Ghent University (BE)
P. Renard, University of Neuchâtel (CH)
|
Abstract: |
The Direct Sampling (DS) algorithm is a recently developed multiple-points geostatistical simulation technique. Its particularity is that it does not store a catalogue of patterns, but instead directly scans the training image (TI) for a given data event, which provides a number of computational and practical advantages. By using distances between data events, the method was extended from categorical variables to continuous variables and multivariate cases. However, in order to benefit from the wide spectrum of potential applications of the DS method, one needs to understand the significance and sensitivity of the user defined input parameters. We performed a systematic study of the most important parameters and assessed their impact on the simulations. Real case TIs were used, including a categorical image of ice wedge polygons and a continuous image of braided river topography. A quantitative sensitivity analysis was conducted on three important parameters balancing simulation quality and computational cost: the acceptance threshold t, the fraction of TI to scan f and the number of neighbors n. We compared the realizations to the TIs using various statistical measures, such as histograms, variograms, connectivity functions and multiple-points histograms. Our findings show that adjusting f offers substantial computational gains without important degradation of the statistical properties of the simulations. Similarly, t and n allow decreasing the CPU time, but it comes at the expense of degraded spatial features. Furthermore, we illustrate the improvement resulting from post-processing and the potential of DS to honor conditioning data and to simulate bivariate fields, hereby giving attention to the weight given to the conditioning data and the relative weights of each variable. We provide a comprehensive guide to performing multiple-points geostatistical simulations with the DS algorithm and recommendations on how to set the input parameters appropriately. |
