AI-DRIVEN MONITORING OF WHEAT DYNAMICS THROUGH REMOTE SENSING AND FIELD DATA IN MULTAN DISTRICT

Abstract

Wheat (Triticum aestivum L.) is Pakistan's most strategically important cereal crop, yet productivity in Multan District remains persistently constrained by co-occurring biotic and abiotic stresses, and absent phenological growth monitoring. These limitations cannot be addressed by conventional manual monitoring methods, which lack the temporal resolution and spatial coverage required for effective precision agriculture in Punjab's wheat belt. This research develops and evaluates an artificial-intelligence-driven integrated monitoring system for wheat dynamics, addressing two core tasks: five-class stress detection and phenological growth stage classification. The system combines a Random Forest and a Temporal Convolutional Network as independent base learners, with the Random Forest capturing spectral-spatial feature patterns from multi-temporal Sentinel-2 vegetation indices and Sentinel-1 radar backscatter, and the Temporal Convolutional Network modelling sequential phenological dynamics across the Rabi growing season, whose outputs are fused through a trained meta-learner that adaptively weights each base learner's contribution based on demonstrated performance, rather than applying a fixed average. Ground truth was compiled across five consecutive Rabi seasons (October 2020 to May 2025) at 248 sample points in the Multan Saddar and Shujabad tehsils, with stress labels derived from the 28 field-verified observation points and growth stage labels assigned across all 248 points, and all configurations were evaluated under Leave-One-Season-Out cross-validation. The stacked ensemble achieved a macro-averaged F1-score of 0.9004 and an area under the receiver operating characteristic curve of 0.9769 for stress detection, and an accuracy of 0.9054 with almost perfect agreement (kappa = 0.8863) for growth stage classification, confirming that the integrated system delivers the strongest stress-detection performance of the three configurations, while matching the best base learner on growth-stage classification. Feature-importance analysis confirms that agronomically relevant indices (normalized difference red edge, land surface water index and enhanced vegetation index) are the most informative features for stress class discrimination, and that the self-attention mechanism assigns higher weights to phenologically critical Rabi months, providing interpretable temporal diagnostics for each monitoring task. The proposed system establishes a scalable and interpretable framework for comprehensive wheat dynamics monitoring, with direct applicability to agricultural advisory services and precision farming in Punjab, Pakistan.