Statistical modelling of analytical and spectral soil measurement errors

Soil is vital for plants, animals and humans. However, human activities have led to rapid soil degradation, prompting global initiatives like the UN Sustainable Development Goals to improve soil quality and promote sustainable land management. Achieving these goals requires comparable and quality-assessed soil data. Soil data from different sources are typically managed in soil information systems. To reduce errors from varying measurement units or methods, data standardisation and harmonisation are crucial. Despite these efforts, laboratory soil data contain measurement errors, due to factors like analyst error or laboratory conditions.

This thesis aimed to quantify measurement errors in analytical soil data and mid-infrared (MIR) spectra, and examine their impact on pedotransfer functions and spectral models. The research also aimed to provide tools for data providers to incorporate repeated measurements in their experimental measurement design: an important step towards routinely supplying soil data with quantified uncertainties.

Measurement errors in pHH2O and total organic carbon (TOC) data were quantified through a linear mixed-effect model approach. Laboratory soil data from the Wageningen Evaluating Programmes for Analytical Laboratories were used, which included replicated measurements over different rounds, batches and laboratories. The results showed the importance of replicates for accurate estimation of measurement error variance, especially in small datasets.
The effect of additive and multiplicative measurement errors on pedotransfer (PTF) calibration and validation was studied. PTFs were calibrated to predict cation-exchange capacity (CEC) in the A- and B-horizons of four soil orders in the contiguous United States. Multiple linear regression (MLR) and random forest (RF) PTFs were calibrated with both ‘true’ and error-contaminated data. The results showed that the mean regression coefficients (MLR) and the variable importance scores (RF) were largely unaffected by measurement errors in the calibration data. However, model performance deteriorated with large measurement errors and small calibration datasets.

The impact of modeller’s choices, and measurement error in calibration data on spectral model performance was examined. Spectral models were developed to predict CEC, pHH2O, total carbon and total nitrogen. Four modeller’s choices were considered: 1) applying a calibration transfer function, 2) defining subsets from a global soil spectral library, 3) applying spectral pre-processing methods, and 4) modelling method. The effect of these modeller’s choices on model performance varied, and no single optimal selection was identified. Measurement errors in analytical soil data notably reduced model performance for most soil properties, while MIR spectral errors had a smaller impact.

Measurement error quantification should become routine practice when deriving new soil data, achievable by including replicate measurements in the experimental measurement design. Developing standard operating procedures is a crucial first step toward this goal. Accurate, standardised measurement information will enable data users to make well-informed decisions about error-contaminated data. The thesis provides insights into the sensitivity of common soil science models to measurement errors, helping data providers balance additional measurements and resources. By offering tools to quantify errors in soil data and MIR spectra, this thesis equips soil scientists to routinely account for measurement errors in prediction models and digital soil maps.