A physics-informed neural network for continuous rice canopy thermal monitoring and forecasting from sparse UAV observations

Continuous monitoring of canopy temperature (Tc), a key indicator of crop water-heat stress and physiological dynamics, using unmanned aerial vehicle (UAV) imagery is inherently limited by temporal discontinuity and the limited physical realism of purely data-driven models. This study develops a physics-informed neural network (PINN) framework to transform temporally sparse UAV thermal observations into continuous hourly rice Tc reconstruction and 48 h forecasting products. The model leverages sparse UAV thermal measurements as supervisory signals while integrating them with continuous meteorological forcing and daily UAV-derived crop phenotypic features. Validated through a comprehensive season-long rice field experiment using walk-forward cross-validation, the proposed PINN framework demonstrated superior performance. It achieved R 2 values of 0.92 for reconstruction and 0.90 for forecasting, with RMSE of 0.71 °C and 0.82 °C, respectively. Ablation analysis further showed that crop phenotypic variables contributed more strongly than temporal descriptors, reducing predictive uncertainty by approximately 4.8–14.3 %, while the integration of SEB physical constraints and uncertainty modeling improved R 2 by 8.4–9.5 % and reduced Total STD by 28.4–37.7 %. The model successfully captures diurnal dynamics, spatial variability, and canopy thermal hysteresis while maintaining physical consistency through improved energy closure. This framework bridges sparse aerial observations with continuous physiological monitoring and highlights its potential to support precision irrigation, early stress detection, and high-throughput phenotyping in smart agriculture.