Md Ishtiaque Alam, Tawfiqur Rahman Sikder, Mohammad Abdus Sami, Md Lutfor Rahman, Md Abu Kawsar Prodhan Hemal, Ahmed Ali Linkon, Mohammad Muzahidur Rahman Bhuiyan, Md Munna Aziz, Md Rashedul Islam, Md Mizanur Rahaman
Crop recommendation systems play a critical role in supporting sustainable agricultural decision making under increasing climate variability. Modern machine learning approaches offer high predictive accuracy, yet their adoption in real-world agri-tech systems depends equally on robustness to environmental change and transparency of decision logic. Using a balanced multi-crop agronomic dataset, this study evaluates classical machine learning models, ensemble methods, and inherently interpretable rule-based learners under two evaluation settings: standard k-fold cross-validation and a rainfall-quartile protocol that simulates shifts in precipitation regimes. The results show that high accuracy under random data splits can substantially overestimate real-world performance when rainfall patterns change. To address this gap, we analyse the accuracy–explainability trade-off by comparing black-box ensembles with interpretable rule-based models. Feature attribution analysis based on SHAP further confirms that rainfall, humidity, and soil potassium are the most influential drivers of crop suitability. The findings provide a data-driven and explainable framework for developing climate-resilient crop recommendation systems that support environmental sustainability, resource-efficient farming, and informed decision making in precision agriculture.