Comparative Study of Parametric and Non-Parametric Models in Crop Recommendation

Abstract

Agriculture is facing mounting challenges from population growth, food insecurity, soil degradation, and climate variability, making traditional practices insufficient to sustain production. AI and machine learning (ML) have recently been applied to agriculture, offering opportunities to improve crop selection and land use through data-driven recommendations. Previous studies have explored crop recommendation systems, but their findings have often been inconsistent due to differences in datasets and methods. This study addresses that gap by benchmarking parametric and non-parametric models on the same dataset to evaluate their performance in a multi-class crop recommendation task. Using an agricultural dataset of 8,000 entries across 11 crop types, data preprocessing included one-hot encoding, label encoding, and median imputation. A neural network implemented in TensorFlow was compared against K-Nearest Neighbors (KNN) and Random Forest models. Results showed that the neural network achieved 80–83% accuracy, while the non-parametric models remained near 9–10%, close to random guessing. These findings suggest that parametric models are better suited for capturing the complex, non-linear patterns in agricultural data. Despite limitations such as overfitting and reliance on a single dataset, the study highlights the potential of AI to provide more accurate and sustainable crop recommendations, ultimately supporting global food security.