Reliability-Aware Federated Ordinal Learning for Weekly Irrigation Management Zone Prediction on Edge Nodes

The vision of smart agriculture depends on systems that can autonomously integrate distributed sensor data to support on-farm decisions such as precision irrigation and nutrient management. In variable-rate irrigation (VRI), this requires timely, high-resolution information to generate management zones that adapt to within-field variability throughout the season. However, current practice often relies on static zone maps and centralized processing, both of which are poorly suited to dynamic field conditions where sensors may fail, or drift and internet connectivity is often limited in resource-constrained environments. This study presents a reliability-aware edge-learning pipeline for weekly VRI zone prediction using federated learning (FL). The proposed framework integrates in-field soil-moisture sensing, weather aggregates, and crop normalized difference vegetation index (NDVI) indicators within an FL setting, enabling field nodes to collaboratively train a shared model without exchanging raw data. This reduces communication demands while preserving data locality. To address sensor unreliability, we develop a reliability model that quantifies weekly data completeness, quality, and stability. These reliability estimates are incorporated through a dual-weighting mechanism: unreliable samples are down-weighted during local training, while more trustworthy clients are given greater influence during global aggregation. The pipeline further employs a rank-consistent ordinal regression (CORAL)-based ordinal classifier to align predictions with ordered irrigation categories: Low, Medium, Medium-High, and High. Experimental evaluation under deployment-oriented operation achieved strong predictive performance, with an accuracy of 0.94, a macro-F1 score of 0.90, and a quadratic weighted kappa (QWK) of 0.98 at the best checkpoint. Multi-week spatial comparisons further showed strong agreement between predicted and geographic information system (GIS)-derived reference zones, while forward-chaining evaluation confirmed useful temporal robustness, with mean accuracy, macro-F1, and QWK values of 0.80, 0.58, and 0.78, respectively. Overall, the proposed framework provides a robust and communication-efficient approach to weekly irrigation zoning that is suitable for farm-scale deployment under unreliable sensing and constrained connectivity.