Abstract—Machine learning models for soil fertility are usually explained with SHAP or LIME. These methods describe what a model has learned, but they do not check whether the learned behaviour agrees with soil science, they do not state how reliable a single prediction is, and they do not tell the farmer which fertiliser to apply. This paper presents SoilCARE, a framework that addresses these three points. First, agronomic knowledge is written as sign constraints in an ordinal gradient-boosting model, so that the predicted fertility never decreases when nitrogen, phosphorus, potassium, organic carbon, sulphur or a micronutrient increases, and never increases with salinity. Second, split conformal prediction gives each prediction a set of classes with a guaranteed coverage level, and soils with an uncertain set are referred to the laboratory. Third, a search procedure finds the lowest-cost and realistic fertiliser plan that moves a Low or Medium soil to the next class. Two measures, the Domain-Consistency Score (DCS) and the paradox rate, are used to check agreement with agronomy. Experiments on a public dataset of 882 Indian soil-test records with 12 nutrients used five repetitions of stratified 5-fold cross-validation. SoilCARE gave a macro-F1 of 0.789, which is not significantly different from unconstrained XGBoost (0.801, p = 0.67). It raised the DCS from 0.56 to 1.00 and reduced the share of soils with contradictory responses from 83% to 0%. The conformal sets reached 90.4% coverage at the 90% target. The recourse search found a valid plan for 99.0% of Low and Medium soils at the lowest median cost among the compared methods (₹3,228 ha-1), and all plans remained valid when the doses were increased by 50%.
 

Keywords—soil fertility, explainable AI, monotonic constraints, conformal prediction, counterfactual explanation, fertiliser recommendation, precision agriculture