Global retrieval of canopy chlorophyll content from Sentinel-3 OLCI TOA data using a two-step upscaling method integrating physical and machine learning models

Authors: LI, DONG - Technical University Of Munich; CROFT, HOLLY - University Of Sheffield; DUVEILLER, GREGORY - Max Planck Institute For Biogeochemistry; Schreiner-Mcgraw, Adam; BELWALKAR, ANIRUDH - Technical University Of Munich; CHENG, TAO - Nanjing Agricultural University; ZHU, YAN - Nanjing Agricultural University; CAO, WEIXING - Nanjing Agricultural University; YU, KANG - Technical University Of Munich

Submitted to: Remote Sensing of Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/30/2025
Publication Date: 6/3/2025
Citation: Li, D., Croft, H., Duveiller, G., Schreiner-Mcgraw, A.P., Belwalkar, A., Cheng, T., Zhu, Y., Cao, W., Yu, K. 2025. Global retrieval of canopy chlorophyll content from Sentinel-3 OLCI TOA data using a two-step upscaling method integrating physical and machine learning models. Remote Sensing of Environment. 328. Article 114845. https://doi.org/10.1016/j.rse.2025.114845.
DOI: https://doi.org/10.1016/j.rse.2025.114845

Interpretive Summary: Global estimates of vegetation growth from remote sensing products remain elusive. A reliable remote sensing model of vegetation productivity at global scales could be extremely useful in assessing crop growth as well as crop responses to environmental conditions. Canopy chlorophyll content (CCC) is related to vegetation photosynthesis and may provide a way to estimate vegetation productivity from satellite measurements. Previous studies have estimated CCC using satellite measurements, but their approaches rely on complicated atmospheric models that increase the uncertainty of estimates of CCC from space. In this study, we test an approach to estimate CCC using top of atmosphere (TOA) measurements and machine learning that does not require the use of atmospheric models. We find that this simple, two-step method provides reliable estimates of CCC when compared to ground observations. Additionally, we show that the final CCC product can be related to gross primary productivity (GPP) measured at 4 flux tower locations. GPP is equivalent to plant growth and while this approach requires additional testing and validation, it may be a method to estimate plant and crop growth from remote sensing measurements.

Technical Abstract: Canopy chlorophyll content (CCC) is tightly related to vegetation photosynthesis and has been a promising indicator of photosynthetic capacity. However, a long-term, global operational CCC product has not yet been available. To fill this gap, we estimated global CCC from Sentinel-3 OLCI top-of-atmosphere (TOA) reflectance using a two-step method: Generating high-resolution CCC, followed by upscaling and then estimating by machine learning. In the first step, a physically based inversion model produced accurate CCC maps from high-resolution (1 m) hyperspectral imagery obtained from the National Ecological Observatory Network (NEON). The validation against ground measurements showed an R² of 0.89 and RMSE of 0.30 g/m². We then upscaled these high-resolution CCC maps and used them as training data in the second step, where a random forest (RF) model linked Sentinel-3 OLCI TOA reflectance to CCC at the spatial resolution of 300 m. The RF model demonstrated robust performance with a 10-fold cross-validation, yielding an R² of 0.97 and RMSE of 0.1 g/m² for both Sentinel-3A and 3B. In addition, the two-step approach showed minimal sensitivity to angular effects and land cover variations, underscoring its robustness. Our results also revealed that (1) the direct inversion method (named the one-step method) led to underestimation and overestimation of CCC, and (2) the relationships between chlorophyll index (e.g. MTCI and CIre) and CCC varied with land cover types, making it unfeasible to estimate global CCC using a single empirical relationship. These findings suggest that our two-step method had great potential in estimating global CCC due to its accuracy and efficiency. Applying the two-step method to Sentinel-3 OLCI TOA reflectance, we generated a long-term CCC product spanning from 2016 to 2024 that has the capability to be continuously updated to the most recent dates. This product can potentially capture the spatial and temporal photosynthetic patterns, thereby advancing research on vegetation photosynthesis and carbon cycles at a global scale.