An RTM-driven machine learning approach for estimating high resolution FAPAR from Landsat 5/7/8/9 surface reflectance

Fraction of absorbed photosynthetically active radiation (FAPAR) is an essential indicator of vegetation productivity and carbon uptake. Though coarse resolution FAPAR products from MODIS and VIIRS sensors have supported global vegetation monitoring, their limitations in characterizing heterogeneity at ecosystem scales necessitate Landsat-derived FAPAR at the 30 m resolution. This study developed a practical approach integrating radiative transfer (RT) modeling and machine learning to estimate 30-m FAPAR from Landsat surface reflectance. A coupled land-atmosphere RT model and shuffled complex evolution (SCE) optimization algorithm was first implemented at globally distributed VIIRS pixels for realistic simulation of land surface reflectance corresponding Landsat spectral bands and FAPAR under various conditions. The simulated dataset encompassing diverse canopy structural and atmospheric states is utilized to train a random forest model relating Landsat surface reflectance bands to FAPAR. The trained RF model was applied to estimate long-term FAPAR from Landsat 5/7/8/9 surface reflectance data and demonstrated reliable performance in capturing FAPAR dynamics across vegetation types. Comprehensive validation of the Landsat FAPAR was conducted using global field measurements from three observation networks, including the VALERI (VAlidation of Land European Remote Sensing Instruments), ImagineS, and Copernicus GBOV (Ground Based Observations for Validation) projects, which involved shrubland, forest, grassland, and cropland, coarse resolution FAPAR products were also used for comparison to evaluate our estimates. Validation results demonstrate good agreement of the estimated Landsat FAPAR with ground measurements, achieving R2 of 0.876, RMSE of 0.108 and bias of 0.032. Consistent accuracy was attained across different Landsat sensors. The validation performance of the Landsat FAPAR over various land cover types was: shrubland (R2: 0.848; RMSE: 0.092), forest (R2: 0.876; RMSE: 0.101), grassland (R2: 0.739; RMSE: 0.118), and cropland (R2: 0.839; RMSE: 0.125). Comparison against GLASS/ MODIS products shows improved accuracy over existing FAPAR datasets, benefiting from the ability of Landsat to resolve within-pixel heterogeneity. This study synergized the strengths of radiative transfer model and machine learning algorithm while overcoming the limitations of RTM parameterization and generalization, providing an efficient and robust Landsat FAPAR estimation appro