The Downscaled GOME-2 SIF Based on Machine Learning Enhances the Correlation with Ecosystem Productivity

Sun-induced chlorophyll fluorescence (SIF) is an important indicator of vegetation photosynthesis. While remote sensing enables large-scale monitoring of SIF, existing products face the challenge of trade-offs between temporal and spatial resolutions, limiting their applications. To select the optimal model for SIF data downscaling, we used a consistent dataset combined with vegetation physiological and meteorological parameters to evaluate four different regression methods in this study. The XGBoost model demonstrated the best performance during cross-validation (R2 = 0.84, RMSE = 0.137 mW/m2/nm/sr) and was, therefore, selected to downscale GOME-2 SIF data. The resulting high-resolution SIF product (HRSIF) has a temporal resolution of 8 days and a spatial resolution of 0.05° × 0.05°. The downscaled product shows high fidelity to the original coarse SIF data when aggregated (correlation = 0.76). The reliability of the product was ensured through cross-validation with ground-based and satellite observations. Moreover, the finer spatial resolution of HRSIF better matches the footprint of eddy covariance flux towers, leading to a significant improvement in the correlation with tower-based gross primary productivity (GPP). Specifically, in the mixed forest vegetation type with the best performance, the R2 increased from 0.66 to 0.85, representing an increase of 28%. This higher-precision product will support more effective ecosystem monitoring and research.