Ground heat flux at daily scale? Estimates from machine learning models and Earth observation products

Energy exchanges between the lower atmosphere and the shallow subsurface are fundamental to understand and quantify processes relevant to the society and ecosystems, such as extreme events, the hydrological cycle, the land carbon cycle, and the Earth heat inventory. Among these energy fluxes, Ground Heat Flux (GHF) corresponds to the conduction of heat through the subsurface. GHF is used for estimating evapotranspiration in order to ensure the conservation of energy in the applied models. Ground heat storage in the continental subsurface is estimated from GHF data, constituting the second largest term of the Earth heat inventory after the ocean. Furthermore, the increase in GHF in recent times is warming permafrost soils in the Arctic, thus enabling the thawing of permafrost and the release of additional carbon into the atmosphere.

Nevertheless, ground heat flux is the term of the surface energy balance with less measurements around the world, hindering the analysis of those processes. There are around 60 Eddy-covariance towers measuring GHF globally, with most sites containing less than a decade of records. Because of this limitation, geothermal data has been used to obtain long-term estimates of GHF. However, these estimates are only able to retrieve long-term changes in surface conditions with decadal to centennial periods, and there are not enough sampling sites to retrieve a global average after the year 2000. Although satellite observations have been recently used to bridge the gap in the heat storage evolution between 2000 and 2020 at the annual scale, data at daily and weekly temporal scales are still necessary in order to analyze the role of GHF on short-term processes such as evapotranspiration and extreme events.

Here, we develop a framework, based on machine learning models and Earth observation products, capable of estimating GHF at daily resolution across several land covers and climate zones. Our framework predicts GHF with a Root Mean Squared Error (RMSE) of 4.79 W m^-2 and a Pearson’s correlation coefficient (R) of 0.65 at the global scale. The performance of the framework improves when predicting 8-day periods, achieving a RMSE of 3.31 W m^-2 and a R of 0.77. A hybrid approach is also evaluated. This method predicts ground surface temperatures and uses them as forcing for a physical model that yields GHF values. Nevertheless, the performance of this hybrid method is lower than the direct approach. We identify several physical processes as the leading features driving the model performance. Given its capability to estimate GHF across several land covers and climate zones, the framework provides the basis for developing a global GHF product, thereby filling a critical gap in the datasets available to study the surface energy balance. Furthermore, this product would enable the characterization of the spatial structure of GHF, contribute directly to monitoring the land component of the Earth heat inventory, and provide a crucial observational reference for developing the land components of global climate models.