Retrieving Crop Traits from Canopy Hyperspectral Reflectance: A Comparative Assessment of Physical, Data-Driven, and Hybrid Models

Hyperspectral remote sensing provides opportunities for accurate and non-destructive retrieval of crop biophysical and biochemical traits based on physical radiative principles; yet robust and transferable retrieval approaches remain challenging. In this study, we systematically compared physically based, data-driven, and hybrid retrieval strategies for estimating leaf chlorophyll content (Cab) and leaf area index (LAI) from 400–2400 nm canopy hyperspectral reflectance from a field spectrometer. Using multi-temporal field observations of potato as a model crop collected across two experimental sites in the Netherlands under contrasting nitrogen and irrigation regimes, we evaluated (i) radiative transfer model inversion using Soil Canopy Observation, Photochemistry, and Energy fluxes (SCOPE) model, (ii) pure data-driven approaches including bidirectional long short-term memory networks (Bi-LSTM) and Gaussian Process Regression (GPR), and (iii) two hybrid methods integrating radiative transfer simulations with machine learning, including GPR hybrid learning and a radiative transfer process-guided machine learning (PGML) framework. Results show that among the data-driven methods, GPR has better performance than Bi-LSTM for Cab retrieval, and slightly lower performance in LAI retrieval. PGML outperformed purely physical and data-driven methods, achieving the highest accuracy for Cab (R² = 0.81, RMSE = 5.41 μg cm⁻²) and LAI (R² = 0.53, RMSE = 0.64 m² m⁻²) in 10-fold cross-validation while requiring limited field measurements. Feature importance analysis revealed that PGML emphasized spectrally and biophysically meaningful regions, including the near-infrared plateau for LAI and the red-edge for Cab. Furthermore, hybrid-derived traits exhibited strong correlations with end-of-season potato yield across key growth stages, comparable to or exceeding those obtained from field measurements. These findings demonstrate the value of hybrid learning for improving the robustness and interpretability of hyperspectral trait retrieval, supporting scalable crop monitoring and precision agriculture applications.