Deriving gapless CO2 concentrations using a geographically weighted neural network: China, 2014–2020

In recent years, China has been experiencing increasing carbon dioxide (CO2) concentration. Spaceborne satellites provide important support to monitor CO2, but the current instruments typically observe with narrow swaths and are frequently influenced by clouds and aerosols, resulting in extensive gaps in the satellite-derived CO2 estimates. To this end, we aimed to reconstruct the CO2 concentration products of the Orbiting Carbon Observatory-2 (OCO-2) to generate a monthly gapless CO2 dataset (2014–2020) for China and investigate the spatio-temporal distribution of CO2 concentration. The geographically weighted neural network (GWNN) model, which can consider temporal and spatial heterogeneity, was developed to establish the complicated relationships between OCO-2 CO2 and the related variables, including CO2 reanalysis data, meteorological variables, radiance data, and satellite normalized difference vegetation index (NDVI) data. Results showed a high correlation in R2 and only a slight deviation in RMSE and MAPE in both simulated validation (R2, RMSE, and MAPE are 0.936, 1.360 ppm, and 0.242 %, respectively) and ground-based validation (R2, RMSE, and MAPE are 0.898, 1.685 ppm, and 0.317 %, respectively). Based on the gapless dataset, the CO2 changes in China were investigated, and a periodically increasing trend with the growth rate of 2.517 ppm/yr was revealed. In addition, the anomaly analysis found that higher CO2 concentrations exist in major cities, especially in highly developed regions. This study provides a gapless reconstruction approach for satellite-derived CO2 and generates a long-term high-resolution CO2 dataset for China, which will be valuable for China’s CO2 emission reduction policy making.