Assessment of the soil fertility status in Benin (West Africa) – Digital soil mapping using machine learning

• Published in: Geoderma Regional Vol. 28; p. e00444
• Main Authors: Hounkpatin, Kpade O.L., Bossa, Aymar Y., Yira, Yacouba, Igue, Mouïnou A., Sinsin, Brice A.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2022
• Subjects: altitude; base saturation; Benin; cation exchange capacity; decision making; Digital soil mapping; exchangeable potassium; issues and policy; Markvetenskap; Multiple soil classes; nitrogen; phosphorus; prediction; Random forest; soil fertility; Soil fertility index; soil organic matter; Soil Science; soil types; spectral analysis; streams; sustainable land management; uncertainty; Benin; Soil fertility index; Random forest; Multiple soil classes; Digital soil mapping
• ISSN: 2352-0094; 2352-0094
• DOI: 10.1016/j.geodrs.2021.e00444

A soil fertility index map (SFIm) can provide key information to decision-makers in regard to spatial planning in the context of sustainable land management. The establishment of such SFIm requires basic soil properties that can be modelled for spatial mapping. The objective of this study was to take advantage of Benin soil legacy data to produce a digital SFIm at a national level based on 8 soil properties (soil organic matter, nitrogen, pH (water), exchangeable potassium, assimilable phosphorus, sum of bases, cation exchange capacity and base saturation). Specific research aims were (1) to model and develop digital soil maps, (2) to identify the key covariates influencing soil nutrients, and (3) to build an SFIm using digital maps of the soil properties. For each soil property, modelling procedures involved the use of different covariates, including soil type, topographic, bioclimatic and spectral data, along with the comparative assessment of the cubist (CB) and quantile random forest (QRF) models. Models were evaluated not only on the basis of classical error metrics (RMSE, R2) but also on the ability to predict local uncertainty based on the prediction interval coverage probability (PICP). The results revealed that CB performed marginally better than the QRF based on classical error metrics (R2, RMSE) but produced the worst uncertainty with an overestimation of the local uncertainty. This suggested that the use of accuracy plots such as PICP to evaluate models can identify accuracy problems not evident with classical error metrics. The analysis revealed that the distance to the nearest stream, which was part of topographic covariates, had strong predictive ability for all the soil properties along with the bioclimatic variables. The spatial distribution of the different classes of SFIm showed a preponderance of low fertility levels with severe limitations for crop development. A limited number of high and average fertility level soils were found in the low elevation areas of southern Benin, and policy could advocate for their sole use for agricultural purposes and promote sustainable management practices. •Creating country level soil fertility index map from different soil properties using a digital soil mapping approach•Prediction uncertainties are better captured by prediction interval coverage probability•Variations of soil properties were related to distance to the nearest stream and bioclimatic variables