Boosting leaf trait estimation from reflectance spectra by elucidating the transferability of PLSR models

Leaf spectroscopy, combined with partial least squares regression (PLSR), is recognized as an efficient and precise tool for measuring plant leaf traits. However, the feasibility of developing a generalizable model remains unclear, primarily due to limited understanding of PLSR model transferability. Here, we collected six key leaf traits along with paired leaf reflectance spectra from 1967 samples of 349 tree species in eight forest sites across China. Using this dataset, we explored the transferability of PLSR models, factors affecting model transferability, and the feasibility of developing generalizable PLSR models for leaf trait prediction. Overall, PLSR models trained at a specific study site demonstrate limited transferability to other study sites. Dissimilarities in plant evolutionary history and environmental conditions between study sites are the primary factors influencing the transferability of PLSR models. Incorporating training data from diverse evolutionary histories and environmental conditions can improve the transferability of PLSR models, achieving accuracy equivalent to that of site-specific models. Our findings provide guidelines for the use of spectroscopy in leaf trait prediction and underscore the urgent need for collaborative efforts to build an open database of leaf traits and reflectance spectra, thereby promoting the development of universal PLSR models for plant leaf trait prediction.