Hourly mapping of surface air temperature by blending geostationary datasets from the two-satellite system of GOES-R series

Spatio-temporally continuous surface air temperature (SAT) is a valuable input for many research fields such as environmental studies and hydrological models. Substantial efforts have been focused on mapping daily SAT using land surface temperature (LST) derived from polar-orbiting satellites such as NASA's Terra and Aqua. However, reconstruction of SAT at very high temporal scales such as sub-daily or hourly was carried out in very few studies, most of which are based on the Meteosat Second Generation (MSG) satellites for small-scale areas with limited stations. In this study, we developed estimation schemes based on random forest for mapping hourly SAT over a large-scale region (CONUS) by blending LST datasets from the two-satellite system of the Geostationary Operational Environmental Satellite-R (GOES-R) Series with the aim of reconstructing the hourly SAT maps with spatially uniform resolutions of about 2 km over CONUS. Estimation schemes with different variables were compared to evaluate the contributions of reanalysis variables and the hour of day (HOD) variable to model predicative performance, and the impacts of station density on model performance were also explored. We found that the model scheme with reanalysis variables and the scheme with HOD achieved higher predicative performance with the mean RMSE of 2.0 K and 2.5 K, respectively. Systematic predictive errors across the synoptic hours of a diurnal cycle are significantly reduced when considering HOD in the modeling of hourly SAT. The model scheme considering both reanalysis variables and HOD achieved the highest performance with a mean RMSE of 1.9 K. We observed that the influence of station density on model performance is related to types of cross-validation and model schemes, and that the uncertainty levels associated with the evaluation of predicative performance for the model schemes decrease with station density, which increases our confidence in the evaluation results for the schemes. Therefore, the mapping accuracy of hourly SAT over the large-scale area CONUS could be conservatively estimated to be 1.9-2.5 K using site-based cross-validation. We expect that this study will facilitate the further mapping of SAT at high-temporal scales over large-scale areas based on abundant and sustained observations from global systems of meteorological geostationary satellites.

Automated summary

This study developed estimation schemes for mapping hourly SAT over a large-scale region by blending LST datasets from two-satellite system. The model scheme with reanalysis variables and HOD achieved the highest performance with a mean RMSE of 1.9 K, indicating the importance of including these variables in the modeling process.