A Spatially Downscaled TROPOMI SIF Product at 0.005° Resolution With Bias Correction

Abstract

Solar-induced chlorophyll fluorescence (SIF) provides a valuable tool for gross primary production (GPP) monitoring. However, the spatial resolution of satellite SIF products is lower than the kilometer level, which hinders their potential for carbon cycle study at regional scales. This work reconstructed a 0.005° SIF in China during 2019 and 2020 from the 5-km level TROPOMI SIF by a proposed downscaling strategy that corrected the predicted bias when statistics-based machine learning models, such as random forest (RF), were used. Our bias-corrected downscaled SIF (named BCSIF) had an improved capacity of preserving the information of the original TROPOMI SIF than the directly predicted SIF from RF. The BCSIF showed better consistencies with the tower-based SIF than the 0.05° TROPOMI SIF with an averaged R2 increased from 0.590 to 0.798 at two sites since it has a more comparable spatial scale with spectral observations. For the spatial–temporal correlations with FLUXCOM GPP at different biomes in China, BCSIF outperformed the original SIF with the averaged R2 increased from 0.472 to 0.877 due to its reduced noise, also outperformed the near-infrared radiation reflected by vegetation (NIRvP), especially for the savanna type with the R2 increased from 0.828 to 0.889. For the temporal correlations with FLUXCOM GPP, BCSIF gives comparable R2 values as NIRvP in more than half of China (around 65% pixels), not including the needleleaf forest region in the southern Tibetan Plateau and savanna region in Yunnan province where BCSIF greatly outperformed, as well as some alpine meadows regions in Inner Mongolia and Tibetan Plateau where NIRvP outperformed.

Keywords

• Bias correction
• gross primary production (GPP)
• random forest (RF)
• solar-induced chlorophyll fluorescence (SIF)
• spatial downscaling