A novel hybrid modelling framework for GPP estimation: Integrating a multispectral surface reflectance based V_cmax²⁵...

Area of Science:

• Ecology
• Remote Sensing
• Biogeochemistry
• Machine Learning

Background:

• Terrestrial gross primary productivity (GPP) is crucial for understanding the global carbon cycle.
• Accurate estimation of GPP relies on the maximum carboxylation rate (V_cmax,0²⁵), but ecosystem-level V_cmax,0²⁵ derivation has high uncertainty.
• Recent research indicates a strong correlation between spectral reflectance and V_cmax,0²⁵.

Purpose of the Study:

• To develop a multispectral surface reflectance-driven simulator for V_cmax,0²⁵ using deep neural networks.
• To construct a hybrid modeling framework for GPP estimation by integrating the data-driven V_cmax,0²⁵ simulator into a process-based model.
• To evaluate the performance of the hybrid GPP model and compare it with existing methods.

Main Methods:

• Developed a fully connected deep neural network to simulate V_cmax,0²⁵ from multispectral surface reflectance.
• Integrated the data-driven V_cmax,0²⁵ simulator into a process-based model to create a hybrid GPP estimation framework.
• Validated the model performance at 95 flux sites and analyzed key spectral bands using Shapley value analysis.

Main Results:

• The V_cmax,0²⁵ simulator showed satisfactory estimation performance across different land cover types (R: 0.34–0.80, RMSE: 14–43 μmol m^-2 s^-1, MdAPE: 21%–66%).
• The hybrid GPP model produced good estimates (R: 0.76–0.89, RMSE: 1.79–6.16 μmol m^-2 s^-1, MdAPE: 27%–90%).
• Multispectral surface reflectance significantly improved V_cmax,0²⁵ and GPP estimates compared to the EVI-driven method, with reduced MdAPE.
• Shapley value analysis identified red, near-infrared, and shortwave infrared bands as crucial for V_cmax,0²⁵ estimation.

Conclusions:

• Multispectral surface reflectance holds significant potential for quantifying ecosystem-level V_cmax,0²⁵.
• The hybrid framework effectively utilizes spectral information via deep learning to reduce parameter uncertainty while maintaining physical realism.
• This approach offers a powerful tool for enhancing the accuracy of global GPP estimation.