Satellite remote sensing is a promising approach for monitoring global CO2 emissions. However, existing satellite-based CO2 observations are too coarse to meet the requirements of fine-scale global mapping. We propose a novel data-driven method to estimate global anthropogenic CO2 emissions at a 0.1° scale, which integrates emissions inventories and satellite data while bypassing the inadequate accuracy of CO2 observations. Due to the co-emitted anthropogenic emissions of nitrogen oxides (NOx = NO + NO2) and CO2, high-resolution NO2 measurements from the TROPOspheric Monitoring Instrument (TROPOMI) are employed to map the global anthropogenic emissions at a global 0.1° scale. We construct the driving features from NO2 data and also incorporate gridded CO2/NOx emission ratios and NOx/NO2 conversion ratios as driving data to describe co-emissions. Both ratios are predicted using a long short-term memory (LSTM) neural network (with an R2 of 0.984 for the CO2/NOx emission ratio and an R2 of 0.980 for the NOx/NO2 conversion ratio). The data-driven model for estimating anthropogenic CO2 emissions is implemented by random forest regression (RFR) and trained using the Emissions Database for Global Atmospheric Research (EDGAR). The satellite-based anthropogenic CO2 emission dataset at a global 0.1° scale agrees well with the national CO2 emission inventories (an R2 of 0.998 with Global Carbon Budget (GCB) and an R2 of 0.996 with EDGAR) and consistent with city-level emission estimates from Carbon Monitor Cities (CMC) with the R2 of 0.824. This data-driven method based on satellite-observed NO2 provides a new perspective for fine-resolution anthropogenic CO2 emissions estimation.