Physics-Guided Machine learning approach for estimating spatiotemporally seamless FAPAR from Landsat surface reflectance

Fraction of Absorbed Photosynthetically Active Radiation (FAPAR) is a key biophysical parameter for characterizing vegetation photosynthetic efficiency, playing a crucial role in understanding terrestrial ecosystems, global carbon cycles, and climate change impacts. While satellite observations enable global-scale FAPAR estimation, existing medium to low-resolution products (250 m-1 km) cannot adequately capture surface heterogeneity, particularly in regions with complex terrain and diverse vegetation types. This study presents a novel physics-guided machine learning approach for high-resolution FAPAR retrieval. First, we established a coupled soil–vegetation–atmosphere radiative transfer model to simulate Landsat surface reflectance and FAPAR at global representative pixels. Second, we integrated simulated reflectance-simulated FAPAR and actual reflectance-actual FAPAR to construct a Physics-Guided Light Gradient Boosting Machine (PGLGBM) for retrieving clear-sky FAPAR from Landsat surface reflectance. Third, we incorporated spatio-temporal constraints and environmental variables to build a second LGBM model for reconstructing missing Landsat FAPAR under cloudy conditions. The method was validated against ground measurements, including Ground Based Observations for Validation (GBOV) and ImagineS, achieving high spatial–temporal consistency (GBOV: R 2 = 0.793, RMSE = 0.113 and ImagineS: R 2 = 0.856, RMSE = 0.110) and demonstrating robust performance across diverse landscapes. This approach effectively addresses the limitations of traditional methods regarding insufficient physical constraints and cloud contamination, while providing continuous 30 m-resolution FAPAR products at 16-day intervals. The resulting high-resolution FAPAR dataset offers valuable support for studying ecosystem light use efficiency, simulating vegetation productivity, and assessing carbon cycle dynamics.