Partitioning eddy covariance CO2 fluxes into ecosystem respiration and gross primary productivity through a new hybrid four sub-deep neural network

A data-driven model is commonly employed for partitioning eddy covariance (EC) CO2 fluxes (NEE) into ecosystem respiration (ER) and gross primary productivity (GPP) fluxes. However, current data-driven solely utilizing one sub-neural network to estimate above-ground respiration (ERa) and below-ground respiration (ERb), leading to substantial uncertainty. To address this issue, this research introduces a hybrid four-sub-deep neural network (HFSD) for partitioning NEE into GPP and ER. The HFSD employs dual sub-deep neural networks (DNNs) to independently estimate ERa and ERb. Additionally, the HFSD incorporates GPP and various environmental variables to predict vegetation transpiration (T). The GPP and ER partitioned by the HFSD model are constrained by EC-derived T and NEE. Comparison between the partitioned GPP and ER by the HFSD model and the nighttime (NT) and daytime (DT) temperature-driven methods is conducted across three EC towers. The results indicate that the dual sub-DNNs architecture enhances the accuracy of ER simulations, while integrating EC-derived T as a constraint improves the accuracy of GPP simulations. Furthermore, the HFSD model exhibits the capability to simulate GPP and ER under extreme scenarios and demonstrates strong generalization potential. Correlation analyses suggest that seasonal variations in GPP and ER are primarily influenced by solar radiation (Ra) and air temperature (Ta) during wet seasons, while GPP and ER are highly sensitive to soil moisture (SM) during dry seasons. This study advances the biophysical description of data-driven models for NEE partitioning and enhances the accuracy of GPP and ER estimates.