Novel Techniques for Mapping of Soil Carbon

Abstract

This thesis presents novel techniques for spatial prediction of soil carbon. Chapter 1 introduces a method to incorporate the local scale spatial variability of soil organic carbon into regional scale mapping. Different to the conventional approach of using globally calibrated single model for the entire region, this method uses a combination of locally and globally calibrated models to predict soil organic carbon at regional scale, using a moving widow approach. Chapter 2 studies how diverse spatial modelling techniques perform under varying training sample sizes, in terms of soil carbon predictions. The study explores the behaviour of various algorithms ranging from simple linear models to complex machine learning techniques trained under numerous sample sizes. Chapter 3 investigates how to optimally use infrared spectroscopy inferred soil carbon data for mapping. Infrared soil spectroscopic data is considered as a timely, low-cost input for spatial modelling of soil carbon. However these data are associated higher measurement errors compared to the standard dry combustion technique. This study establishes a methodology in the model-based geostatistics to filter out the measurement error variability through the inclusion of the error information in the covariance structure of the spatial model. In disaggregating soil information, uncertainty of the disaggregation process is not often discussed. Underestimation of inferential or predictive uncertainty in statistical modelling leads to inaccurate statistical summaries and overconfident decisions. The use of Bayesian inference allows for quantifying the uncertainty associated with disaggregation process. Chapter 4 introduces Bayesian area-to-point regression kriging with a case study of downscaling regional scale soil organic carbon map to farm scale information.