African soil properties and nutrients mapped at 30 m spatial resolution using two-scale ensemble machine learning

Summary

This continental-scale study produced a 30 m resolution Soil Information System for Africa by combining approximately 150,000 soil samples with Earth observation data (MODIS, Sentinel-2, Landsat) using a two-scale 3D ensemble machine learning framework. The resulting maps provide predictions for 20+ soil properties at three depths, with varying accuracy; soil pH was most reliably predicted (CCC = 0.900) whilst extractable phosphorus and sulphur were more challenging (CCC = 0.654–0.708). The research demonstrates the feasibility of fine-resolution pan-African soil mapping and identifies satellite spectral bands and climatic variables as the most important predictors for continental-scale soil characterisation.

UK applicability

This study is primarily focused on African soil conditions and may have limited direct applicability to UK soil mapping, which already benefits from established national soil surveys and finer-scale legacy data. However, the machine learning framework and methodological approach could inform future refinement of UK soil databases and pan-European soil monitoring initiatives.

Key measures

Concordance correlation coefficient (CCC) for predictions of soil pH, organic carbon, total nitrogen, effective cation exchange capacity, extractable phosphorus, potassium, calcium, magnesium, sulphur, sodium, iron, zinc, silt, clay, sand, stone content, bulk density, and depth to bedrock

Outcomes reported

The study produced predictions for 20+ soil properties including pH, organic carbon, total nitrogen, extractable nutrients (P, K, Ca, Mg, S, Na, Fe, Zn), texture components, bulk density and depth to bedrock at three soil depths (0, 20, 50 cm) across the African continent using machine learning. Spatial cross-validation showed accuracy levels varying from soil pH (CCC = 0.900) to extractable phosphorus (CCC = 0.654).

Topic tags