Addressing geographical bias in global CO2 flux predictions.

Abstract

Carbon fluxes play an important role in our climate model on Earth. These fluxes have also been shown to be related to global warming. However, these fluxes are only measured at specific locations. Therefore, to obtain a global prediction, these local observations need to be upscaled to a global data product. This has previously been done by combining half-hourly flux measurements with globally available meteorological data and using machine learning to predict the flux based on the measurements alone. When doing so, an important distinction between tropical and extra-tropical regions was not considered. This is important as vegetation cycles are very different in the tropics which has a large impact on the carbon fluxes. These differences are hard to detect for a model because there is a very large imbalance in data availability for the tropics versus the extratropics. In this paper, this distinction is examined and multiple methods are proposed to make the model spatially aware. These methods include a tropic boolean variable, the latitude and longitude coordinates of the measurement, and a separate model for tropics and extratropics. The results showed that a non-spatially aware model does indeed struggle to predict correct diurnal cycles. The predictions improved by introducing spatial variables to the model with the best performing approach being the two separate models. But with more data, the latitude longitude model might perform the best as the model can figure out the tropic to extratropical transition itself. This showed that current approaches are indeed lacking due to spatial bias, and this paper addresses multiple possible solutions.