Deep learning based XCO2 global map generation using satellite observations

Greenhouse (GHG) gases are the primary driver of climate change. There are two approaches to measuring GHG emissions: bottom-up and top-down. Bottom-up measurement involves collecting data based on local emissions and modeling individual sources and sinks of carbon. This is useful for understanding the specific drivers of GHG emissions; however, there is a time lag for collecting data from each source, and good national statistics are required. Meanwhile, top-down measurement involves estimating GHG emissions based on atmospheric measurements and modeling. In general, the bottom-up approach is considered to be more accurate than the top-down approach, but the top-down approach is helpful for providing broad-scale estimates of GHG emissions, and it allows for spatial mapping on a global scale.

In this study, we use OCO-2 satellite products to generate a XCO2 global map using a deep-learning-based technique. Although OCO-2 measurement provides the CO2 concentrations with the highest spatial resolution on a global scale, there are limitations to the FOV coverage and the low temporal resolution of the low-earth orbit satellite. To solve this problem, we use additional satellite products, which can be a precursor to CO2, and we applied TabNet which is firstly introduced at the International Conference on Machine Learning (ICML) 2020. TabNet is an attention-based network that uses a self-attention mechanism. The preliminary results showed that the XCO2 global map for every half of the month could be provided using integrated satellite observations. The validation results will be discussed in the presentation.