From Canopy Phenology to Lithological Signals: Evaluating Biophysical Traits with Machine Learning in the Hațeg Basin

Using vegetation indices for lithological signal detection is a well-known practice; however, these indices are characterized by strong equifinality, as they encapsulate both biochemical and biophysical characteristics, causing subtle lithological differences to be lost. Our study presents a new framework that breaks down the remote-sensed spectrum into physically grounded features. A nine-year (2017–2025) Sentinel-2 time series (310 scenes) was used to estimate biophysical parameters—chlorophyll content (Cab), water content (Cw), and leaf area index (LAI)—using the 1D PROSAIL radiative transfer model in the Hațeg Basin. Random Forest and Multi-Layer Perceptron (MLP) classifiers were used with a rigorous spatial block-based cross-validation framework. The PROSAIL model decomposes the spectrum into independent, physically meaningful variables and substantially reduces the ambiguity problem associated with empirical indices. Using the combined dataset containing vegetation indices and biophysical parameters along with the MLP, we achieved 66.07% accuracy in forested areas and 68.80% in grassland areas. Feature importance analysis revealed that over dense forest cover, the MLP benefits from the indirect biochemical pathway (Cab), while for grasslands, it favors the soil brightness scale (rsoil) during the late summer and fall periods. These results establish a PROSAIL-based workflow for vegetation-covered lithological mapping, demonstrating that the vegetative canopy operates as a decodable biogeochemical lens.