DTSIF: A machine learning-based downscaling TROPOMI solar-induced chlorophyll fluorescence dataset with high spatiotemporal resolution during 2002–2024 in Huang-Huai-Hai region, China

Solar-Induced Chlorophyll Fluorescence (SIF) provides a new method for large-scale vegetation photosynthesis monitoring. However, the spatiotemporal resolution of existing SIF products is still relatively rough, which cannot meet the needs of precision agricultural monitoring and production, especially in the Huang-Huai-Hai (HHH) region, China. In this study, we reconstructed the TROPOspheric Monitoring Instrument (TROPOMI) SIF dataset with 4 days and 8 days, 500 m resolution from 2002 to 2024 in the HHH region (DTSIF), using eight trained machine learning models based on MODIS reflectance, fraction of photosynthetically active radiation (FPAR) and leaf area index (LAI) data and ERA5 reanalysis data. Results indicate that the ExtraTrees model can simulate and reconstruct TROPOMI SIF dataset in highest simulation accuracy, with training and testing R 2 and RMSE values were 1 and 5.4950E −16 W·m −2 ·μm −1 ·sr −1 , and 0.896 and 0.0787 W·m −2 ·μm −1 ·sr −1 , respectively. The SHAP analysis indicated that Tem, LAI, Band1, FPAR and Band5 were the top five predominant drivers influencing SIF. The spatial residuals between DTSIF and TROPOMI SIF ranged −0.2148–0.4486 W·m −2 ·μ −1 ·sr −1 , with an average value of 0.0061 W·m −2 ·μ −1 ·sr −1 , and followed a normal distribution. From site scale, DTSIF showed significant linear correlations ( p < 0.05) with gross primary productivity (GPP) observed at 8 ChinaFLUX sites (with R 2 values ranged 0.360–0.848) and SIF data observed at 3 ChinaSpec sites (with R 2 and root mean square error (RMSE) values ranged 0.553–0.641 and 0.1378–0.1668 W·m −2 ·μm −1 ·sr −1 , respectively). From spatiotemporal scale, DTSIF showed good spatial correlations with CSIF, GOSIF and MODIS GPP, with the average R 2 value between DTSIF and CSIF (4 days) was 0.750, and that between DTSIF and GOSIF, GPP (8 days, monthly and annual scales) were 0.811 and 0.786, 0.885 and 0.877, 0.522 and 0.494, respectively. Moreover, DTSIF also exhibited significant linear relationships with CSIF, GPP and GOSIF, with the R 2 values larger than 0.950 for each long-term time scale. The new reconstructed DTSIF can provide high-precision data support for climate change research such as vegetation photosynthesis monitoring, carbon flux estimation and agrometeorological disaster monitoring in the HHH region. • A new TROPOMI SIF dataset (DTSIF) in HHH region was reconstructed. • DTSIF was reconstructed using ExtraTrees model based on MODIS products and ERA5 reanalysis data. • The spatiotemporal resolution of DTSIF has been improved to 4 and 8 days, 500 m. • Performance of DTSIF has been verified by both satellite and tower-based GPP and SIF data.