Hybrid Explainable Machine Learning Framework for Soil-Specific Crop Recommendation

Abstract

Accurate crop recommendation is a critical decision-support task in precision agriculture, requiring the integration of heterogeneous soil and environmental information. Using real-world results from soil laboratories in Pune, India, this study provides an example of a hybrid explainable artificial intelligence-based soil-specific crop recommendation system using machine learning and data from weather/crop and crop suitability by region. The framework uses soil nutrients (N, P, K, organic carbon, pH, EC), weather severity and derived fertility indices, and seasonal climatic variables to determine the most suitable crop category for the field. Data heterogeneity and class imbalance were tackled with preprocessing, feature engineering, and SMOTE-based balancing. A Random Forest, XGBoost, and a Bayesian Neural Network (MC Dropout) were trained on and fused through weighted probability ensembles. The hybrid model achieved an average of 94.1% ± 0.31% over 5-fold stratified cross-validation and 96% on a holdout data set, demonstrating generalization. The implementation of LIME and SHAP for explainable AI was used to obtain (un)certainty estimation to assist with confidence-aware support. The evidence indicates that real-world agricultural advisory systems can use this framework to tackle problems with explainability, reliable crop recommendation systems, and accurate systems, in addition to providing a base for subsequent modules for the recommendation of fertilizers.