Information about soil nutrient contents is key for explaining measured crop responses to soil fertility management practices and for updating and upscaling of soil fertility management recommendations. This project, implemented primarily upon request of the Netherlands Environmental Assessment Agency,(PBL), produced soil nutrient maps of Sub-Saharan Africa with the aim to contribute to increasing agricultural sustainable productivity.
Results include spatial predictions of soil nutriënt contents (macro-, meso- and micronutriënts) across Sub-Saharan Africa at a spatial resolution of 250 m for 0–30 cm. Model training was run using soil samples from over 59,000 point locations and an extensive stack of covariates including remote sensing covariates, landform, lithologic and land cover layers (maps). Spatial predictions were made using an ensemble model from the machine learning algorithms random forest and gradient boosting, implemented in R packages ranger and xgboost. The results of cross-validation showed that apart from extractable S and P, significant models can be produced for the targeted nutrients (R-square between 40–85 %). A limiting factor for mapping nutrients using the point data available for Africa is the high spatial clustering of the sample locations, following the AfSIS design, causing many countries and land cover / land use groups being unrepresented.
This work presents a systematic follow up to maps published in Hengl et al. (2015); R code used to generate the revised predictions is available via the github repository.
The work should be cited as follows:
