Interpretable crop-environment regime analysis using robust unsupervised clustering in arid agro-ecosystems

Interpreting crop-environment patterns in arid agro-ecosystems is challenging since soil properties and weather conditions interact in complex ways, while reliable crop labels, yield records, or field-level suitability classes may not always be available. This study presents an interpretable unsupervised clustering analysis for identifying and explaining crop–environment regimes using seven variables: nitrogen, phosphorus, potassium, pH, temperature, humidity, and rainfall. A controlled twelve-crop analytical subset was organized into six post-hoc crop–environment regime groups. The comparative evaluation included K-Means, Gaussian mixture modelling, DBSCAN, and hierarchical agglomerative clustering. The crop and regime labels were retained only for interpretation and were not used during unsupervised model fitting. Cluster validity was assessed through internal validation metrics, k-sensitivity analysis, repeated-run stability testing, pairwise adjusted Rand index analysis, and DBSCAN parameter sensitivity. The selected six-regime K-Means solution achieved a Silhouette Score of 0.4836, Davies-Bouldin Index of 0.8181, and Calinski–Harabasz Score of 4948.41. Repeated K-Means analysis across 30 independent initializations showed stable performance, with a mean Silhouette Score of 0.4825 ± 0.0060 and mean Davies-Bouldin Index of 0.8179 ± 0.0010. DBSCAN produced a valid seven-cluster density-based structure with near-zero noise, indicating possible finer substructure within the crop–environment feature space. To improve interpretability, a high-fidelity Random Forest surrogate was trained to reproduce the final K-Means cluster assignments. The surrogate achieved an accuracy and macro-F1 score of 1.0000. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) identified rainfall, nitrogen, potassium, phosphorus, and humidity as the main variables contributing to cluster discrimination, while temperature and pH showed comparatively lower explanatory influence. The findings show that robust unsupervised clustering combined with surrogate-based explainability can support transparent crop–environment regime interpretation without claiming direct crop recommendation, yield prediction, or field-level suitability assessment.