Author Identifiers

Hassan Talebi
ORCID: 0000-0002-4038-4200

Date of Award


Degree Type


Degree Name

Doctor of Philosophy


School of Science

First Advisor

Associate Professor Ute Mueller

Second Advisor

Dr Johnny Lo


Spatial uncertainty modelling and prediction of a set of regionalized dependent variables from various sample spaces (e.g. continuous and categorical) is a common challenge for geoscience modellers and many geoscience applications such as evaluation of mineral resources, characterization of oil reservoirs or hydrology of groundwater. To consider the complex statistical and spatial relationships, categorical data such as rock types, soil types, alteration units, and continental crustal blocks should be modelled jointly with other continuous attributes (e.g. porosity, permeability, seismic velocity, mineral and geochemical compositions or pollutant concentration). These multivariate geospatial data normally have complex statistical and spatial relationships which should be honoured in the predicted models.

Continuous variables in the form of percentages, proportions, frequencies, and concentrations are compositional which means they are non-negative values representing some parts of a whole. Such data carry just relative information and the constant sum constraint forces at least one covariance to be negative and induces spurious statistical and spatial correlations. As a result, classical (geo)statistical techniques should not be implemented on the original compositional data. Several geostatistical techniques have been developed recently for the spatial modelling of compositional data. However, few of these consider the joint statistical and/or spatial relationships of regionalized compositional data with the other dependent categorical information.

This PhD thesis explores and introduces approaches to spatial modelling of regionalized compositional and categorical data. The first proposed approach is in the multiple-point geostatistics framework, where the direct sampling algorithm is developed for joint simulation of compositional and categorical data. The second proposed method is based on two-point geostatistics and is useful for the situation where a large and representative training image is not available or difficult to build. Approaches to geostatistical simulation of regionalized compositions consisting of several populations are explored and investigated. The multi-population characteristic is usually related to a dependent categorical variable (e.g. rock type, soil type, and land use). Finally, a hybrid predictive model based on the advanced geostatistical simulation techniques for compositional data and machine learning is introduced. Such a hybrid model has the ability to rank and select features internally, which is useful for geoscience process discovery analysis.

The proposed techniques were evaluated via several case studies and results supported their usefulness and applicability.

Access Note

Access to Chapters 2, 3 and 4 of this thesis is not available.