Creating 1-km long-term (1980-2014) daily average air temperatures over the Tibetan Plateau by integrating eight types of reanalysis and land data assimilation products downscaled with MODIS-estimated temperature lapse rates based on machine learning

Air temperature (Tair) is critical to modeling environmental processes (e.g. snow/glacier melting) in high-elevation areas of the Tibetan Plateau (TP). To resolve the issue that Tair observations are scarce in the TP western part and at high elevation, many studies have estimated daily air temperatures by using MODIS land surface temperature (LST) and various reanalysis datasets. These estimates are however inadequate for supporting high-resolution long-term hydrological simulations or climate analysis due to the high cloud cover, short time span or low spatial resolution. To improve the Tair estimation, this study develops a novel machine-learning based method that uses the Gradient Boosting model to efficiently integrate observations from high-elevation stations with eight widely used air temperature reanalysis and assimilation datasets (i.e., NNRP-2, 20CRV2c, JRA-55, ERA-Interim, MERRA-2, CFSR, ERA5 and GLDAS2) downscaled with remote sensing-based temperature lapse rates (TLR). This method is used to generate a new dataset of daily air temperature with the 1-km resolution for the period of 1980-2014. To overcome the problem that TLR derived from limited stations may be unreliable, a new TLR estimation method is developed to first estimate spatially continuous monthly TLRs from MODIS LST and then downscale daily mean Tair from eight reanalysis and assimilation datasets to obtain Tair at the 1-km resolution using the MODIS-estimated TLRs. The Gradient Boosting (GB) model is selected for integrating the eight downscaled Tair and five other auxiliary variables. The models are trained and validated using observations from 100 common stations (i.e. China Meteorology Administration stations) and 13 independent high-elevation stations (4 on glaciers). The results show that the proposed TLR estimation method can efficiently reduce exceptional TLRs in the meantime keeping acceptable downscaling accuracy. The downscaled Tair from JRA-55 is the best among the eight downscaled datasets followed by ERA-Interim, MERRA-2, CFSR and others. Finally, the GB-integrated Tair further outperforms the downscaled JRA-55 Tair with the mean root-mean-squared-deviation (RMSD) of 1.7 degrees C versus 2.0 degrees C, especially in high-elevation stations with mean RMSD of 1.9 degrees C versus 2.7 degrees C. Both the MODIS-estimated TLR and the high-elevation training observations are demonstrated to significantly improve the air temperature estimation accuracy of the GB model in high-elevation stations. This study also provides a framework for integrating multiple reanalysis and assimilation temperature data with elevation correction in mountainous regions that is not restricted to the TP.

Automated summary

A novel machine-learning based method was developed to accurately estimate daily air temperature in high-elevation areas of the Tibetan Plateau, utilizing remote sensing data and reanalysis datasets. By integrating observational data and multiple data sources, the accuracy of air temperature estimation was significantly improved.