Improving Environmental Prediction by Assimilating Auxiliary Information

Abstract

The concern of this work is the systematic synthesis of site-specific samples and auxiliary information (including continuous and categorical variables) aiming at improving spatial prediction of natural attributes (soil properties, contaminant processes etc.). Bayesian Maximum Entropy (BME) is the theoretical support of the proposed synthesis. The significance of the synthesis is that it can increase the prediction accuracy of natural attributes in a physically meaningful and technically efficient manner. The spatial prediction approach is applied in a real world case study that combines soil organic matter (SOM) content samples with auxiliary information (terrain indices, soil types, and soil texture) to generate predictive maps. Prediction was affected by soil type and soil texture (prediction accuracy increased when categorical variables were included). In the same case study, the BME-based approach was compared with mainstream spatial statistics techniques, like Regression Kriging (RK) with auxiliary information, and hard data-driven Ordinary Kriging (OK). The numerical results demonstrated the superiority of the BME-based approach over the Kriging-based techniques, whereas it was found that some key BME parameters (counts of soft data, predicted variables categories, and continuous auxiliary variable categories) can have different effect on SOM prediction accuracy. The success of BME-based prediction relied heavily on finding adequate auxiliary information about the study situation.

Keywords: spatial statistics, prediction, BME, Kriging, auxiliary information, soil, nitrogen