Hyperspectral surface reflectance improves GPP estimation in terrestrial biosphere modeling using model-data fusion

Gross Primary Productivity (GPP) estimates from terrestrial biosphere models (TBMs) are often uncertain due to limited constraints on vegetation biochemical and biophysical properties. Remote sensing offers promising opportunities to reduce these uncertainties, yet its full potential remains understudied. Here, we conducted model-data fusion experiments, including Observing System Simulation Experiments (OSSEs), and Observing System Experiments (OSEs) at the Harvard Forest site, using the Terrestrial Ecosystem Carbon cycle simulator (TECs) with an embedded spectral invariant theory-based radiative transfer model. In OSSEs, we assimilated synthetic hyperspectral reflectance, multispectral reflectance, and Leaf Area Index (LAI) into TECs to evaluate their effect under the ideal conditions. In OSEs, we assimilated PRecursore IperSpettrale della Missione Applicativa (PRISMA) hyperspectral reflectance (620–1000 nm), MODerate resolution Imaging Spectroradiometer (MODIS) multispectral reflectance (broadband red and near-infrared), and MODIS-derived LAI to optimize model parameters, including several key vegetation traits such as leaf chlorophyll content (Cab), maximum carboxylation rate at 25 °C (V_cmax25), and LAI. Results show that hyperspectral reflectance consistently outperforms multispectral reflectance and LAI in improving GPP estimates and reducing uncertainties, with RMSE decreasing from 2.68 to 1.18 μmol CO₂ m⁻² s⁻¹ in OSSEs, and from 6.74 to 5.42 μmol CO₂ m⁻² s⁻¹ in OSEs. This is because hyperspectral information better constrains seasonal variations in canopy structure and Cab. Meanwhile, both hyperspectral and multispectral reflectance outperform LAI, with information from both canopy structural parameters and leaf biochemical properties, thus offering a joint constraint on GPP simulations. Our findings highlight that remotely sensed reflectance data, particularly hyperspectral reflectance, have great potential to improve photosynthesis modeling and reduce uncertainties in GPP estimates within TBMs.