Evaluation of AI Predictability on Crop Yield using XAI - An Experimental Approach

Abstract

Accurate crop yield prediction is essential for improving agricultural planning and sustainability. This study explores the integration of explainable artificial intelligence (XAI) with machine learning models to enhance transparency in agricultural decision-support systems. A comparative evaluation of multiple regression algorithms - Linear, Ridge, Lasso, Elastic Net, Decision Tree, Random Forest, AdaBoost, XGBoost, CatBoost, LightGBM, Gradient Boosting, and K-Nearest Neighbour Regression (KNR) is performed for crop yield prediction. Categorical attributes are transformed using One-Hot Encoding, while model reliability is ensured through 5-Fold Cross-Validation and hyperparameter optimization using Grid Search. Among the evaluated approaches, KNR demonstrated the best predictive capability with an R² value of 0.9827 and reduced errors (MAE:9.03, RMSE:117.58, MSE:13824.73). To improve interpretability, SHAP, LIME, and ELI5 techniques are employed to explain feature contributions and model decision. The proposed framework provides transparent and reliable insights, supporting data-driven agricultural management and sustainable crop production strategies.