Downscaling top-down CO2 emissions and sinks in China empowered by hybrid training

Atmospheric CO₂-based top-down approaches enable objective evaluation of climate mitigation efforts but face dual constraints: sparse monitoring limits spatial resolution, while emission heterogeneity hampers downscaling. To enhance downscaling accuracy, we present a hybrid training method integrating multi-resolution inverse fluxes—national-scale coarse grids with fine-scale grids (Shanxi/Jiangsu). Experiments show hybrid training outperforms conventional approaches, increasing R² from 0.56 to 0.61 with 2.7%-area fine grids and reducing prediction biases compared to data fusion without high-resolution inputs while vastly exceeding nearest-neighbor interpolation (R² = 0.39). By combining gap-filled CO/NO₂ columns, nighttime lights, population density, vegetation indices, and meteorological data, we downscaled national 45 km eight-day CO₂ fluxes to daily 10 km resolution. The derived dataset reveals emission inequities: top 20% cities contribute more than 50% of national emissions, exposing regional capacity disparities. This framework leverages expanding CO₂ monitoring networks to progressively refine spatiotemporal resolution, enabling city-level verification of mitigation actions.