An approach for crop recommendation with uncertainty quantification based on machine learning for sustainable agricultural decision-making

While machine learning (ML) models for crop recommendation have demonstrated high predictive accuracy, a critical gap persists in their practical reliability: the omission of uncertainty quantification. Existing studies predominantly deliver deterministic recommendations, neglecting inherent uncertainties arising from data noise. This raises concerns about the reliability of the decision support systems for crop recommendation. To address this, we propose an ensemble ML framework incorporating entropy-based uncertainty quantification. Trained on a publicly available Indian agricultural dataset with 2200 samples across seven agro-climatic features (nitrogen, phosphorus, potassium, temperature, humidity, pH, and rainfall) and 22 crop classes, the model achieves a predictive accuracy of 99.54 %. By estimating prediction confidence using entropy, the framework offers probabilistic recommendations that support environmentally informed decision-making under uncertainty. These findings suggest that integrating uncertainty measures into ML-driven crop recommendation systems can enhance reliability and promote sustainable agricultural practices.