Appraising retrieval schemes from spaceborne hyperspectral imagery for mapping leaf and canopy traits in forest ecosystems

Forest ecosystems, covering about one-third of the Earth’s mainland, are vital for providing essential ecosystem services. However, their extent and health are threatened by climate change. Remote sensing has the potential to evaluate the condition and functionality of global forests, but methodological and technological challenges impede the quantitative estimation of forest traits from spaceborne imagery. The development of next-generation sensors and advanced retrieval algorithms offers the chance to overcome these obstacles, though such data and models need development and testing. In this study we investigated and compared the potential of machine learning regression algorithms (MLRA) and hybrid approaches combining MLRA and radiative transfer simulations for the retrieval of forest traits from PRISMA spaceborne imagery. We tested the method in the Ticino Park, a mid-latitude forest located in northern Italy. We conducted an intensive field campaign in the summer of 2022 concurrently with four PRISMA overpasses to collect trait samples for calibrating and validating the retrieval schemes. Our results demonstrated the capability of PRISMA images and hybrid models to accurately quantify Leaf Chlorophyll Content (LCC) (r 2 =0.67, nRMSE=13.5 %), Leaf Nitrogen Content (LNC) (r 2 =0.82, nRMSE=9.5 %), Leaf Water Content (LWC) (r 2 =0.98, nRMSE=3.5 %), Leaf Mass per Area (LMA) (r 2 =0.93, nRMSE=6.6 %) and Leaf Area Index (LAI) (r 2 =0.83, nRMSE=11.7 %) in forest ecosystems. Conversely, Leaf Carotenoid Content (Ccx) yielded a lower accuracy (r 2 =0.43, nRMSE=13.5 %), indicating potential for improvement. The results evidenced a slightly superior performance of hybrid approaches over purely statistical approaches. The application of the models to PRISMA images acquired before and after a severe drought event corroborated the effectiveness of the models to provide reliable estimates in operational conditions. This underscores the valuable role of next-generation models and hyperspectral spaceborne imagery for forest monitoring. To our knowledge, this is the first study to appraise hybrid retrieval schemes with real spaceborne hyperspectral data for mapping multiple forest traits, thereby providing a reference framework for future applications. • PRISMA detected spectral differences related to leaf biochemistry. • At the leaf level (LCC, Ccx, LNC) hybrid models showed better performance than MLRA. • MLRA was more sensitive to the range of variation in the training dataset. • Kernel-based algorithms (GPR, KRR, SVR) outperformed other MLRA (PLSR, NN, PCR). • Hybrid models captured well the trait-specific variability induced by the 2022 drought.