Using Multisource Data and Time Series Features to Construct a Global Terrestrial CO₂ Coverage by Deep Learning

Carbon dioxide (CO2) is the most important atmospheric contributor for global warming. Satellite remote sensing is a commonly used method for high-precision CO2 detection, but it often suffers from striping issues, which hinders its ability to achieve full coverage. Therefore, the limited availability of data poses challenges for global carbon accounting. In this article, we generate a global seamless and high-resolution dataset of column-averaged dry air CO2 mole fractions (XCO2) from 2017 to 2020 by integrating multiple data sources using deep learning techniques feedforward neural network (FNN). The data sources primarily include satellite, ground-based, and reanalysis XCO2 products, satellite vegetation index data, and meteorological data. The spatial resolution of the dataset is 0.1°, and the temporal resolution is one day. Moreover, this article also investigated the importance of features in deep learning models and examined the spatiotemporal variations of global XCO2. The results demonstrate that the FNN approach with time series features yields an improved dataset ( ${R} =0.98$ and RMSE =0.82 ppm) compared to the other methods. In an FNN, the known model data [Carbon Tracker (CT) XCO2] are considered as the most important features. We find that the global XCO2 has increased by approximately 7.5 ppm from 2017 to 2020. This seamless and fine-scale dataset provides valuable support for understanding global carbon cycling and formulating carbon emission reduction policies.