The reconstruction for the monthly surface air temperature over the Tibetan Plateau during 1901–2020 by deep learning

Abstract

The Tibetan Plateau (TP) has experienced a prominent warming in recent decades, and there are many datasets to help people to understand the variations and distributions of climate change over the TP. However, the observed stations, satellite images, simulation models, and reanalysis products have weaknesses of sparse spatial distribution, short-term span, inconvenient access, and coarse resolution, respectively. Therefore, it is scarce for the high-resolution and long-term temperature dataset over the TP. In this study, based on the Climatic Research Unit (CRU) gridded data, the monthly surface air temperature dataset with a spatial resolution of 0.1° over the TP during 1901–2020 was reconstructed by a deep learning method called the Generative adversarial networks (GAN). Compared with the observation, the accuracy of temperature from the GAN (with Root Mean Square Error and Correlation Coefficients of 1.35 °C and 0.97) is improved compared to the CRU (with Root Mean Square Error and Correlation Coefficients of 3.30 °C and 0.82). In comparison with the CRU, Global Land Precipitation and Temperature (GLPT), ERA5, and NCEP/NCAR, the temperature from the GAN have the best robustness, especially over the western TP, suggesting the generated dataset has the best performance. Besides, compared to CRU, the temperature of GAN has an improved resolution and can present a more detailed texture structure of the temperature. Compared to Downscaling CRU with the same resolution, the temperature of GAN inherits the spatial and temporal distribution characteristics of CMFD and therefore better reflects the variation of temperature with longitude. Overall, the reconstruction for the monthly surface air temperature over the TP during 1901–2020 has a long-term temporal span and high spatial resolution, which provides an alternative basis for the temperature change over the TP.