Geostatistical modeling of geochemical variables in3D

by Milutin M Pejovi

Institution: Univerzitet u Beogradu
Year: 2017
Keywords: 3D modeliranje zemljita; 3D regresioni kriging; lasso; ugnjedena unakrsna validacija; procena zagadenosti; topografskaizloenost
Posted: 02/01/2018
Record ID: 2153620
Full text PDF: https://fedorabg.bg.ac.rs/fedora/get/o:15303/bdef:Content/get


Geodesy - Modeling and Management in Geodesy /Geodezija - Modeliranje i menadzment u geodeziji Geostatistical mapping of soil properties in 3Drefers to the application of geostatistical methods to the soildata in order to produce maps of soil properties at differentdepths. Through two separate studies, this thesis elaborates on twodifferent approaches for 3D soil mapping. At first, the wellestablished Spline-Than-Krige approach for the mapping of soilpollutants atmospherically deposited from the copper smeltingplant, was used. In the absence of the monitoring data, which canbe used for a detailed characterization of the plume spreadingprocess, this study was confined to the consideration of terrainexposure to explain spatial trend in arsenic distribution atdifferent depths. This study aims to explore the extent to whichthe commonly available information, such as the prevailing winddirection, or the location of the source of pollution, incombination with the digital terrain model, can be used to quantifythe terrain exposure, and hence to improve the spatial predictionof the arsenic concentration at several soil depths. Next, theinnovative geostatistical approach to 3D mapping of soilproperties, based on soil profile data, was proposed. It providesthe semi-automatic way for 3D modeling of soil variables,prediction over the regular grids (rasters) and also the evaluationof prediction accuracy. Methodologically, this approach operateswithin the 3D regression kriging framework. 3D trend model isconceptualized as hierarchical or non-hierarchical linearinteraction model. This means that the model includes theinteractions between the spatial covariates and depth in thehiearchial or non-hierarchial manner. The trend modeling is basedon the application of the penalized regression technique, lasso.The lasso uses a specific regularization penalty in a fittingprocedure to enable the efficient parameter estimation and variableselection (including interaction terms) at the sametime...Advisors/Committee Members: Bajat, Branislav, 1963-.