Explainable machine learning for tracking spatial variation in leaf chlorophyll fluorescence within temperate deciduous forest canopies

Abstract Monitoring forest ecosystems is essential for understanding ecological processes, assessing biodiversity, and guiding sustainable forest management. Chlorophyll fluorescence (ChlF) parameters provide valuable, physiologically relevant insights into the state of the photosynthetic apparatus and serve as indicators in forest monitoring. However, effectively describing the spatial distribution of ChlF parameters within the canopy and developing convenient methods to track their spatial variations remains challenging for scale extension and remote sensing applications. This study focused on several typical temperate deciduous forest species, for which the ChlF parameters, spectral reflectance, and canopy leaf coordinates were synchronously recorded. The results indicated that there were significant spatial variations in ChlF parameters within the canopy. The development of different machine learning (ML) algorithms for estimating ChlF parameters using different combinations of spectral reflectance, species, and leaf coordinates demonstrated that the random forest and extreme gradient boosting (XGBoost) models using leaf spectral reflectance have considerable potential for capturing spatial variation in ChlF parameters. Furthermore, the inclusion of additional information, such as species and leaf coordinates, significantly improved the performance of these models (explaining over 85% of the variance for all ChlF parameters) as well. This study supports the assessment of spatial patterns of ChlF parameters and demonstrates the potential of ML to characterize their variation across forest canopies. By directly modeling ChlF parameters at high spatial resolution, this approach provides new insights into forest physiological processes and advances biodiversity and environmental research.