Physics-guided deep learning improves UAV multispectral estimation of strawberry canopy LAI and SPAD

Accurate estimation of strawberry canopy traits from UAV multispectral imagery is important for agricultural productivity but remains difficult under greenhouse conditions because illumination variation, canopy overlap, shadowing, background interference, and coupled structural and biochemical responses weaken purely empirical relationships. We propose DeepRT, a physics-guided multi-task framework for the joint estimation of leaf area index (LAI) and canopy chlorophyll status represented by SPAD. DeepRT combines PROSAIL-generated spectral priors with modality-specific learning from calibrated reflectance, vegetation indices, and texture descriptors. A reliability-aware fusion module adapts the contribution of each modality, while synthetic supervision, latent-domain alignment, and a physical-feasibility regularizer constrain learning under limited field labels. The study comprised 34 temporally ordered UAV campaigns and 310 quality-controlled sampling-unit observations. A chronological train-validation-test split, five repeated runs, and time-series cross-validation were used to limit temporal leakage. DeepRT achieved \(R^2\) values of \(0.92\pm 0.015\) for LAI and \(0.84\pm 0.018\) for SPAD, with RMSE values of \(0.32\pm 0.018\) m \(^2\) /m \(^2\) and \(2.8\pm 0.18\) SPAD units, respectively. Predicted uncertainty was positively associated with absolute error, with Spearman correlations of 0.612 and 0.587. These results support physics-guided heterogeneous fusion as a reproducible approach to greenhouse crop-trait retrieval while also identifying illumination-dependent and high-uncertainty conditions that require cautious interpretation.