Canopy Chlorophyll Retrieval in Structurally Complex Pedunculate Oak Forests: Empirical Machine Learning Versus Hybrid PROSAIL–GPR

Hybrid radiative transfer model–machine learning (RTM–ML) approaches can incorporate physical constraints into vegetation-trait retrieval, but their advantage over empirical learning remains uncertain for very-high-resolution, limited-band observations of structurally heterogeneous forest canopies. We compared six empirical machine-learning regression algorithms with a hybrid PROSAIL–Gaussian Process Regression (GPR) approach for retrieving derived canopy chlorophyll content (CCC) from four-band UAV multispectral imagery of pedunculate oak (Quercus robur L.) stands under contrasting oak lace bug (Corythucha arcuata) pressure. CCC was derived from field-derived leaf chlorophyll content and UAV-estimated leaf area index. Data were collected across three sampling periods from 2021 to 2023, and models were evaluated using tree-grouped three-fold cross-validation to maintain independence between training and held-out trees. Empirical performance varied among algorithms, sites, and sampling periods. At the infested site, GPR provided the strongest performance in May (R2 = 0.61, RMSE = 27.04 µg cm−2, NRMSE = 10.87%), whereas Random Forest performed best in September. At the reference site, Random Forest, least-squares linear regression, and kernel ridge regression provided the strongest results depending on sampling period and performance metric, with R2 reaching 0.60. According to the primary cross-validated predictive metrics (R2, RMSE, and NRMSE), the tested PROSAIL–GPR approach generally showed lower predictive performance than the empirical models under the same tree-grouped evaluation framework. Empirical GPR maps resolved substantial within-stand spatial variation in derived CCC. Overall, incorporating PROSAIL-simulated information did not improve predictive performance under the tested four-band, crown-scale configuration. Rather than identifying a single cause, the results indicate that the benefit of physically constrained retrieval depends on compatibility among model representation and parameterization, canopy structure, spectral information, and available training data.