Improving terrestrial net primary productivity estimation through process-consistent FPAR retrieval from multi-source remote sensing data

• Process-consistent FPAR dataset reconstructed from FLUXNET observations. • FAMA-Net integrates multi-source variables for FPAR retrieval. • FAMA-Net mitigates NDVI saturation and improves FPAR estimation. • Improved FPAR enhances CASA-based NPP simulations accuracy. • SHAP reveals how canopy, water, and energy regulate FPAR variability. Fraction of absorbed photosynthetic active radiation (FPAR) is a key input to Light Use Efficiency (LUE) models for estimating net primary productivity (NPP). Conventional FPAR retrieval approaches based on vegetation indices often suffer from signal saturation and limited representation of environmental constraints, introducing uncertainties into terrestrial carbon sink assessments. In this study, an ecosystem-scale, process-constrained FPAR observation was generated from FLUXNET via a CASA-consistent optimization framework to provide physiologically plausible supervisory signals. Using these data, a feature-adaptive multi-head attention network (FAMA-Net) was developed to model nonlinear relationships among vegetation indices, surface reflectance, and climatic variables. FAMA-Net consistently outperforms conventional models across multiple ecosystem types, with R 2 improvements of 0.008–0.512 and RMSE reductions of 0.010–0.085, particularly in evergreen needleleaf forests and open shrublands. When integrated into the CASA model, FAMA-Net derived FPAR improves NPP estimates (R 2 = 0.611–0.843) relative to the classical CASA implementation and MODIS products. SHAP analysis further reveals that, beyond LAI and NDVI, water- and energy-related variables such as evapotranspiration, potential evapotranspiration, and soil moisture substantially contribute to FPAR variability. This study demonstrates that synergistic modeling of multi-source vegetation and environmental data can effectively improve the reconstruction of FPAR within the CASA model and enhance the accuracy of terrestrial carbon sink estimation in remote sensing applications.