Use of machine learning in Moroccan soil fertility prediction as an alternative to laborious analyses

Soil fertility assessment and management are challenging for several African countries, as it requires continuous monitoring of important indicators such as organic matter, potassium and phosphorus, which implies time and chemical inputs consuming techniques. The present study aims to test the feasibility of using three machine learning algorithms, multiple linear regression (MLR), support vector machine (SVM) and random forest (RF), to predict three essential soil fertility elements (OM, K2O and P2O5). A sample of 400 soils randomly collected in Doukala, central Morocco, was involved in three machine learning models to evaluate their respective throughput predictions. Data were split in two subsets as follow: 70% of data (n = 280) used for training while 30% (n = 120) for validation. The highest throughput prediction was determined through root-mean-square error (RMSE) and coefficient of determination (R-2). All examined models displayed satisfactory results in predicting organic matter for training and validation with an average R-2 greater than 0.6 coupled with an overall low RMSE value. Regarding K2O and P2O5, the models mentioned above exhibited poor performance levels reflected through the low coefficient of determination for both training (R-2 < 0.5) and validation (R-2 < 0.35) along with small RMSE values. Results also showed that cation exchange capacity, carbonates and texture were the main variables that significantly contributed to the prediction of OM, P2O5 and K2O. As many areas of how these variables can impact soil fertility prediction using machine learning are still needed to be investigated further, this study is of great interest for digital soil fertility assessment and mapping with the lowest cost compared to conventional methods.