One of the backbones of the Indian economy is agriculture, which is conditioned by the poor soil fertility.
In this study we use chemical soil measurements to classify many relevant soil parameters: village-wise
fertility indices of organic carbon (OC), phosphorus pentoxide (P2 O5 ), manganese (Mn) and iron (Fe); soil
pH and type; soil nutrients nitrous oxide (N2 O), P2O5 and potassium oxide (K2 O), in order to recommend
suitable amounts of fertilizers; and preferable crop. To classify these soil parameters allows to save time
of specialized technicians developing expensive chemical analysis. These ten classification problems are
solved using a collection of twenty very diverse classifiers, selected by their high performances, of fam-
ilies bagging, boosting, decision trees, nearest neighbors, neural networks, random forests (RF), rule based
and support vector machines (SVM). The RF achieves the best performance for six of ten problems, over-
coming 90% of the maximum performance in all the cases, followed by adaboost, SVM and Gaussian
extreme learning machine. Although for some problems (pH;N2 O;P2 O5 and K2O) the performance is moderate, some classifiers (e.g. for fertility indices of P2O5 ; Mn and Fe) trained in one region revealed valid for other Indian regions.
Keywords: Soil type, Machine learning, Random forest, Soil fertility, Fertilizer recommendation