Generating high spatial resolution GLASS FAPAR product from Landsat images

Abstract

The fraction of absorbed photosynthetically active radiation (FAPAR) is an essential biophysical variable for monitoring vegetation growth and quantitatively describing the efficiency of light absorption by vegetation. Several long-term global satellite FAPAR products, such as MODIS, GLASS, GEOV2, and GIMMS3g-FAPAR, have been produced at coarse spatial resolutions, however, they are inclined to criticism over spatial heterogeneous landscapes due to mixed pixel effects. As the successor of GLASS FAPAR products, the high spatial resolution GLASS (Hi-GLASS) FAPAR product suite has been generated in this study. We developed a hybrid algorithm through integration of physically-based radiative transfer models and machine learning to estimate the black-sky, white-sky and blue-sky Hi-GLASS FAPAR products at 30 m resolution from Landsat imagery globally. The coupled soil-leaf-canopy (SLC) radiative transfer model was applied to depict the physical relationship between FAPAR and Landsat surface reflectance data, and its simulations were executed with uniform distributions of input parameters. Separate random forest models were trained with these simulation samples and applied to Landsat data to generate the Hi-GLASS FAPAR products. Extensive validation of the FAPAR products was performed for different sensors and biomes using in situ measurements from the VAlidation of Land European Remote sensing Instruments (VALERI), ImagineS, and Copernicus Ground Based Observations for Validation (GBOV) projects. Results demonstrated that our FAPAR products were highly accurate for blue-sky (R2 = 0.91, RMSE = 0.09, bias = −0.03), black-sky (R2 = 0.94, RMSE = 0.11, bias = −0.02), and white-sky (R2 = 0.88, RMSE = 0.10, bias = −0.02). R2 and RMSE values for different Landsat sensors were slightly different in the order of TM (0.84, 0.13), ETM+ (0.92, 0.12), and OLI (0.95, 0.11), while the corresponding statistics for biomes were ranked from the best to the worst: shrubs (0.78, 0.08), crops (0.93, 0.09), grasslands (0.88, 0.11), and forests (0.85, 0.13). Additionally, Hi-GLASS products displayed reasonable spatial distribution and seasonal variation of vegetation dynamics. The developed method performed better than those for producing the decametric-resolution and coarse resolution FAPAR products, enabling us to generate long-term 30 m FAPAR records from the 1980s.