Global forecasting of carbon concentration through a deep learning spatiotemporal modeling

Given the global urgency to mitigate climate change, a key action is the development of effective carbon concentration reduction policies. To this end, an influential factor is the availability of accurate predictions of carbon concentration trends. The existing spatiotemporal correlation as well as the diversity of influential factors, pose important challenges in accurately modeling these trends. In this work, different strategies based on deep learning are proposed with the aim of predicting global carbon dioxide and methane concentrations. For this purpose, satellite observations are used for six-month projections, covering geographical regions that span the globe. In addition, complementary environmental variables are integrated to improve the predictive capacity of the proposed models. The results obtained demonstrate the high accuracy of the predictions, in particular of models based on graphical neural networks, reaffirming the great potential of deep learning techniques in predicting carbon dioxide and methane concentrations. Likewise, the effectiveness of models based on deep learning to accurately predict carbon concentrations by incorporating dynamic and static information is demonstrated.