An explainable and auditable IoT–AI engineering framework for soil–environment regime discovery in arid agriculture

Sustainable crop planning in arid and semi-arid regions is constrained by soil degradation, water scarcity, and high environmental variability, while the adoption of data-driven decision methods remains limited by poor transparency and trust. This study investigates whether unsupervised soil–environment regimes can be discovered and interpreted in a scientifically transparent manner to support crop planning and soil management in data-scarce arid regions. To address this need, an integrated IoT–AI–XAI–blockchain framework is presented for interpretable and auditable soil–environment regime discovery. Using soil and climatic variables including nitrogen, phosphorus, potassium, pH, temperature, humidity, and rainfall, we evaluate multiple unsupervised clustering approaches K-Means, DBSCAN, and Agglomerative clustering to identify homogeneous agro-environmental regimes. Cluster quality and stability are assessed using repeated internal validation. K-Means produced four stable and interpretable regimes, achieving a mean Silhouette Score of 0.62 ± 0.03 and a Davies–Bouldin Index of 0.71 ± 0.05 showing stronger internal validation performance than the alternative methods considered in this study. Instance-level and global explanations are generated using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) through a surrogate-based explainability strategy for better interpretability of soil monitoring parameters. Cryptographic hashes of analytical outputs are recorded for auditability. The results suggest that the integration of explainable unsupervised learning and lightweight provenance logging can provide a structured, interpretable, and auditable analytical basis for data-driven agronomic analysis.