Spatio-temporal mapping of Leaf Area Index (LAI) in open-canopy forests remains challenging due to pronounced structural heterogeneity, sparse vegetation cover, and strong effects of soil background and understory vegetation on canopy reflectance. This study assesses empirical and hybrid approaches for spatio-temporal LAI retrieval in the Zagros open-canopy forests using Sentinel-2 imagery, with the aim of developing a robust and transferable framework specifically adapted to heterogeneous forest structures. Empirical models, including Stepwise Multiple Linear Regression, Artificial Neural Networks, Random Forest, Support Vector Regression (SVR), k-Nearest Neighbors, and Gaussian Process Regression, were developed using Sentinel-2 spectral bands, vegetation indices, and a canopy height model (CHM) as predictors. For the hybrid approach, PROSAIL radiative transfer simulations were used to generate a look-up table, which was then inverted with SVR to retrieve LAI from Sentinel-2 reflectance. Models were trained using plots from the main study site (SS1) and evaluated based on an independent site (SS2) to assess predictive accuracy and spatio-temporal transferability. Results indicated that empirical models achieved high accuracy when integrating with the canopy height model. Notably, Stepwise Multiple Linear Regression performing best (R2 = 0.83, RMSE = 0.18 m2.m-2). However, hybrid approaches demonstrated superior spatio-temporal transferability, maintaining robustness when applied to an independent site sampled across different spatial and temporal conditions (R2 = 0.52, RMSE = 0.32 m2.m-2), outperforming empirical models (R2 = 0.39, RMSE = 0.45 m2.m-2). Spatial analysis further indicated that hybrid models better captured fine-scale heterogeneity, particularly in complex mountainous terrain, whereas empirical models tended to overestimate low LAI values. Overall, while empirical methods offer computational efficiency, their limited transferability reduces suitability for dynamic open-canopy systems. Hybrid approaches, despite higher computational demands, provide enhanced ecological realism and stability across space and time. These findings highlight the value of hybrid frameworks for robust spatio-temporal LAI mapping and support their application in sustainable management of fragile and heterogeneous forest ecosystems.
