Fertilizer Recommendation Using Deep Soil Inspection

Conventional fertilizer-appointment systems generally provide a one-stop solution to nutrient management which may not be optimal in different agricultural conditions. This paper proposes an explainable XGBoost-based framework, which considers soil properties, climatic parameters, and crop-specific information to provide precise fertilizer recommendations. A novel contribution is the introduction of SHAP (SHapley Additive exPlanations) for model interpretability to help farmers understand the key drivers of the recommendations. The dataset consisted of 1,200 samples of NPK values, pH, temperature, rainfall, soil color, and crop type, which were preprocessed by standardization and label encoding. Grid-search-based hyperparameter tuning helped in model optimization, achieving 92.3% accuracy on test data as compared to 88.5% for Random Forests and 84.7% for SVMs. SHAP analysis highlighted soil Nitrogen (weight=0.35) and Rainfall (weight=0.18) as the most influential features. A user interface to provide recommendations in real time demonstrated the working of this system. The work brings forward a connection between complex machine learning models and agricultural insight.