Machine learning to generate soil information

Abstract

This thesis is concerned with the novel use of machine learning (ML) methods in soil science research. ML adoption in soil science has increased considerably, especially in pedometrics (the use of quantitative methods to study the variation of soils). In parallel, the size of the soil datasets has also increased thanks to projects of global impact that aim to rescue legacy data or new large extent surveys to collect new information. While we have big datasets and global projects, currently, modelling is mostly based on "traditional" ML approaches which do not take full advantage of these large data compilations. This compilation of these global datasets is severely limited by privacy concerns and, currently, no solution has been implemented to facilitate the process. If we consider the performance differences derived from the generality of global models versus the specificity of local models, there is still a debate on which approach is better. Either in global or local DSM, most applications are static. Even with the large soil datasets available to date, there is not enough soil data to perform a fully-empirical, space-time modelling. Considering these knowledge gaps, this thesis aims to introduce advanced ML algorithms and training techniques, specifically deep neural networks, for modelling large datasets at a global scale and provide new soil information. The research presented here has been successful at applying the latest advances in ML to improve upon some of the current approaches for soil modelling with large datasets. It has also created opportunities to utilise information, such as descriptive data, that has been generally disregarded. ML methods have been embraced by the soil community and their adoption is increasing. In the particular case of neural networks, their flexibility in terms of structure and training makes them a good candidate to improve on current soil modelling approaches.