The application of digital soil mapping to improve the resolution of national soil properties across Great Britain.

Many countries have created soil maps to illustrate the variety of soil properties and support how soils can be used. Traditional soil mapping by field survey and interpretation has been the most recognised form of soil mapping for many years and an effective way to capture a variable soil landscape. Such maps have enabled scientists and stakeholders to improve their understanding of relationships between soils and other landscape factors such as geology and land cover. However, with the amount of soil information growing and technology improving, Digital Soil Mapping (DSM) has been developed as an alternative approach to generate soil property predictions and to produce finer resolution soils data. Currently, DSM produces maps based on training of models with observed soils data and environmental covariates and then releases these to stakeholders to evaluate their utility. This PhD has taken a different approach by addressing stakeholder needs at the beginning of the process. The overall aim of this PhD was to improve the spatial resolution of soil properties across Great Britain (GB), as informed by stakeholders. Three main aims were identified. The first assessed what current soils data and information stakeholders currently use, and what improvements they want to see from future soil-related products. The second aim, using information from the questionnaire survey and a comparison of laboratory and analytical methods, is to develop DSM which could be applied across the whole of GB. This was done by comparing two modelling approaches: Boosted Regression Trees (BRTs) and Multivariate Adaptive Regression Splines (MARS) for mapping soil properties (loss-on-ignition, texture and pH) across two pilot areas. The characteristics of MARS and BRT models at both training and deployment stages are examined. The third outcome investigated how well the soil properties mapped across GB, building on the development of DSM in the pilot areas and whether they reflect expert pedological knowledge. This section also focusses on how suitable an independent validation dataset is at evaluating soil property predictions. This PhD has shown that stakeholders are aware of what soils data and information they are using and could clearly express what is needed to improve current maps. Wider use of soil information by non-soil experts would be improved by increasing data accessibility and user- friendly supporting materials. Fundamentally, most stakeholders require finer resolution than what is currently available which identifies an opportunity for DSM to fill some of this gap. To address these gaps and develop DSM across GB, this PhD focussed on mapping soil properties that were directly comparable across Scotland and England & Wales and also key to many stakeholder information needs. After investigation of laboratory and analytical methods from the two national soil surveys of Scotland and England & Wales, soil loss on ignition, soil texture and soil pH were chosen for developing DSM for GB. From the development of DSM, results showed that MARS models produced better statistical performances than BRTs for predicting soil properties within a training environment. However, when MARS models are deployed to larger areas, they extrapolate beyond their means and BRTs performed better. This is because MARS models perform more consistently when many variables are required. Furthermore, MARS models struggle with overfitting and missing data which subsequently leads to incorrect and unfeasible pedological relationships between soil properties. BRT models, despite not performing as well statistically, produce more consistent relationships between pedology and mapped soil properties. This is because BRT models introduce randomness in the boosting which reduces overfitting and improves the predictive performance. BRTs have shown to be more consistent in the mapping outputs than MARS because regressing to the mean is more favourable when most data matches up with one another. However, this does not necessarily mean that the full range of soils in these areas were being captured by the BRT model. This led to scaling up from the pilot areas to modelling soil properties across GB using a single regional BRT model and evaluating its performance. BRT modelling results for GB at 2D and 3D predict well for pH and LOI but less so for texture. Going forward, more data are required to produce more consistent modelling outputs especially for areas across GB where soil properties are not currently being predicted well. The GB modelling results also highlighted a poor performance of the model against an independent validation dataset. This is because the original data for both GB training and validation datasets were analysed and collected for different purposes. These datasets were taken at different time periods under a different sampling design. Furthermore, the data for both training and validation GB datasets were collected at different scales. At present, these first versions of soil property DSM maps for GB have produced variable results. However, this exercise has shown that the development of reliable DSM maps would benefit from interaction between pedologists, modellers and stakeholders to ensure that mapped outputs are of sufficient quality at adequate finer resolution and can be usable. Such DSM maps, alongside management recommendations, will help to address many global challenges associated to soils. However, DSM is not the panacea for all mapping needs. Until such time that DSM fully develops into DSA and adequately incorporates the breadth of information available in traditional soil maps, mapping from field survey and observation will continue to be necessary for stakeholders.