Mitigating Spatiotemporal Error Clustering in SIF Downscaling: A Novel Zoning Strategy Achieves High-Accuracy Reconstruction

Spatiotemporally refined solar-induced chlorophyll fluorescence (SIF) information over large regions has become a crucial data source for terrestrial carbon cycle studies and agricultural drought stress monitoring. Various machine learning (ML) methods have been employed to reconstruct satellite-based SIF retrievals, addressing issues such as spatial gaps and coarse resolution. However, the spatiotemporal heterogeneity in the statistical relationships between explanatory variables and SIF introduces potential errors in these empirical models, leading to spatiotemporal error aggregation. This study proposes a novel spatiotemporal zoning strategy to reconstruct GOME-2 SIF. The Spatiotemporal Ensemble (STE) model integrates multi-source geospatial data and diverse ML models, demonstrating superior SIF estimation performance. The STE model achieved an R2 of 0.925 with 5-fold cross-validation and an RMSE of 0.058 mW m−2 sr−1 nm−1, outperforming six classical ML models by 14%. The STE model significantly mitigates spatiotemporal error aggregation. Temporally, the seasonal variability in model performance was reduced, with the RMSE gap between months with peak and trough errors decreasing to merely 0.04 mW m−2 sr−1 nm−1. Spatially, error clustering trends were weakened. We successfully generated a reconstructed SIF product for China at 0.01°/8-day spatiotemporal resolution spanning 2008 to 2018. Reducing spatiotemporal error aggregation in SIF estimation will enhance its application potential.