Advanced analysis and modelling tools for spatial environmental data. Case study: indoor radon data in Switzerland

The present work deals with development and adaptation of advanced geostatistical models and machine learning algorithms (statistical learning theory – Support Vector Machines) for comprehensive analysis and decision-oriented modelling of environmental spatial data. The real case study is based on indoor radon data. The inherent high variability at different spatial scales of noisy indoor radon measurements coupled with the heavy clustering effect of houses locations make this dataset an excellent candidate to assess the feasibility of traditional and advanced models, trend and risk mapping at local and regional scales. General methodology of spatial data analysis and modelling includes comprehensive exploratory data analysis, qualitative and quantitative description of monitoring networks, analysis and modelling of spatial anisotropic correlation structures, spatial predictions and simulations using geostatistical models and machine learning algorithms. In the present study this methodology is applied for the real case study of indoor radon data. The main attention in the present study is paid to the quantification of monitoring network clustering, modelling of spatial correlations, conditional stochastic simulations and risk mapping and application of statistical learning theory for regional classification. Regularised variography helps to improve the visibility of spatial structures on variograms of both raw and transformed data. Once modelled, the spatial correlation structures of data (variograms) are used to produce traditional maps with ordinary kriging. Due to the spatial variability of data, kriging induce a strong smoothing effect that make this approach only suitable for