|
Project properties |
|
| Title | Mapping multiple soil properties using state-of-the-art machine learning |
| Group | Soil Geography and Landscape |
| Project type | thesis |
| Credits | 36 |
| Supervisor(s) | Gerard Heuvelink |
| Examiner(s) | Gerard Heuvelink |
| Contact info | Gerard.Heuvelink@wur.nl |
| Begin date | 2022/01/03 |
| End date | |
| Description | A sub-discipline of soil science, Digital Soil Mapping (DSM), consists in producing computer-assisted soil maps by the use of statistical methods. This relies on the large availability of spatial data, computer power and GIS tools. Recently, there has been a growing interest in using machine-learning methods for the purpose of generating spatial soil information. In particular, random forest has good predictive potential, as well as tools to quantify the uncertainty associated with the spatial prediction of a soil property. However, little has been done to investigate how several soil properties can jointly be predicted. This is relevant, for example, when predicting correlated variables such as sand, silt and clay content of the soil.
This thesis research will investigate whether it is possible to use random forest to predict jointly several soil properties. A second objective is to derive uncertainty measures associated with these predictions. Several tools, such R packages, are available but have not yet been used for the purpose of generating soil maps. The methodology can be applied to our large database of several soil properties from the whole Europe. |
| Used skills | R programming |
| Requirements | Basic knowledge on spatial statistics |