A Robust Crop Recommendation System Leveraging Soil and Climate Parameters

---

Résumé

We present a benchmarking study of classical machine learning (ML) methods for crop recommendation from soil and climate parameters with an emphasis on methodological transparency, interpretability and deployability in low-resource contexts. We evaluate K-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Classifier (SVC), Gaussian Naıve Bayes (NB), Bagging (BG) and a soft-voting ensemble, trained and validated on a curated dataset of N=2,200 instances comprising N–P–K, pH, rainfall, humidity and temperature features [1]. Models are tuned via nested, stratified k-fold cross-validation and assessed using accuracy, precision, recall and F1-score with 95% confidence intervals (bootstrap). Beyond aggregate metrics, we report global permutation importance and partial dependence to enhance interpretability; we further discuss temporal extensions (e.g., LSTM) for seasonal dynamics [2]. While the voting ensemble attains state-of-the-art accuracy on this dataset (best single-run accuracy: 99.77%), we underscore the risk of overfitting given the limited sample size and advocate external validation and real-world trials. We detail hyper- parameter searches, preprocessing and scaling strategies, statistical significance testing (McNemar) and ethical/practical considerations for inclusive deployment. The contribution is a rigorously controlled benchmark and deployment-oriented blueprint rather than a new algorithmic novelty.

Mots-clés

Crop Recommendation, Precision Agriculture, Ensemble Learning, Interpretability