Abstract
<title>Abstract</title> <p>The rapid digitization of the agricultural sector has created a critical reliance on centralized cloud architectures, often resulting in high latency and significant data privacy vulnerabilities. This study presents the Soil Intelligence System (SIS), a novel, hardware-agnostic, and hierarchical framework designed to perform localized, real-time agronomic assessment at the edge. The proposed SIS architecture utilizes a four-layer intelligence engine, featuring an asynchronous Extended Kalman Filter (EKF) for multimodal sensor fusion—integrating sparse chemical telemetry with high-frequency temporal data—and an asymmetric autoencoder to resolve dimensionality imbalances across disparate sensing platforms. By implementing deterministic L1/L2 rule-based filtering and a non-linear L3 One-vs-Rest (OvR) gradient boosting ensemble, the system achieves robust crop suitability matching while maintaining complete offline operational capability. Furthermore, an integrated L4 online learning protocol enables continuous adaptation to localized micro-ecological concept drift without the need for periodic cloud retraining. Experimental validation across 90 managed plots and 14,250 multi-sensor arrays demonstrates that the SIS architecture provides high-fidelity, interpretable agricultural decision support while operating entirely within local farm-edge computational boundaries.</p>