Physics-constrained deep learning for joint estimation of terrestrial CO2 fluxes and differentiated climatic sensitivities across China (2002–2022)

Accurate joint estimation of gross primary productivity (GPP), ecosystem respiration (Reco), and net ecosystem exchange (NEE) is essential for quantifying terrestrial CO 2 budgets; however, it remains a challenge. To address physical inconsistency from separate modelling, this study developed the Multi-Source CarbonNet (MS-CarbonNet) model, in which dynamically adjustable physiological correlation constraints and carbon balance constraints were embedded into the loss function. It integrates eddy covariance observations from 275 site-year records, multi-source remote sensing, and environmental drivers to achieve physically consistent joint estimation of CO 2 fluxes. Against eight benchmark models, MS-CarbonNet achieved the best overall performance with high accuracy across diverse ecosystem types. Validation R 2 values for GPP, Reco, and NEE reached 0.82, 0.79, and 0.73, respectively, with RMSE reductions of 23.8–45.1%, 16.2–33.2%, and 27.3–40.4%, relative to benchmarks. Spatiotemporal retrieval for China’s terrestrial ecosystems (2002–2022) revealed a spatial gradient of higher fluxes in the southeast and lower values in the northwest. The national multi-year mean GPP, Reco, and NEE were 7.51 ± 0.96, 6.25 ± 0.88, and -1.26 ± 0.31 Pg C yr −1 , respectively, with GPP and Reco showing significant increasing trends and the carbon sink strengthening at -0.021 Pg C yr −1 . The 10–20°N zone served as a relatively stable high-value region, while 20–30°N contributed the largest share to national CO 2 fluxes, decreasing towards higher latitudes. Sensitivity analysis and SHAP modelling jointly revealed a complementary and differentiated climatic regulation mechanism for China’s terrestrial carbon sink. In 61.53% of China’s terrestrial area, the statistical temperature sensitivity of GPP exceeded that of Reco; conversely, in 49.52% of the area, the statistical precipitation sensitivity of Reco exceeded that of GPP. MS‑CarbonNet provides a robust and transferable framework for multi‑flux estimation in China's terrestrial ecosystems, advances understanding of spatiotemporal dynamics and climatic drivers of regional carbon sinks, and offers scientific support for climate‑adaptive carbon management strategies.