A Deep Learning Method for Statistical Downscaling of CLDAS Relative Humidity with Different Sources of Data: Sensitivity Analysis

High-resolution relative humidity (RH) data are essential in studies of climate change and in numerical meteorological forecasting. However, because high-resolution meteorological grid data require a large number of stations, the sparse distribution of ground meteorological stations in China before 2008 has limited the development of long-term and high-resolution RH products in the China Meteorological Administration’s Land Assimilation System (CLDAS) dataset. To retrieve high-quality and high-resolution RH data before 2008, we propose a statistical downscaling model (SDM) based on a generative adversarial network (GAN) to transform the original RH data from a resolution of 0.05° to 0.01°. The GAN-based SDM (GSDM) is trained with the RH of the CLDAS (0.05°) dataset after 2008 as its input, and the RH of the high-resolution CLDAS (HRCLDAS, 0.01°) dataset after 2008 as its target for training. The 2-m air temperature data from the HRCLDAS dataset are also included in the input, and the station observations of RH are incorporated in the target for training. To select the optimum data combination for the model, we compared three methods: (1) incorporating without auxiliary data (GSDM), (2) incorporating air temperature as an additional input (GSDM_T), and (3) incorporating air temperature as an additional input and the RH data at stations as an additional target for training (GSDM_TO). Taking the Beijing–Tianjin–Hebei region as an example, we trained the GSDM by using data from 2018 and tested the model performance in 2019. The experimental results showed that the GSDM_TO algorithm achieved the lowest root-mean-square error (3.85%), followed by the GSDM_T (4.01%) and GSDM (4.95%) algorithms. The proposed models showed a competitive performance and captured more local details of the RH fields than other deep learning models and traditional bilinear interpolation. In general, the GSDM_TO algorithm using a combination of different sources of data (air temperature and observed RH) achieved the best results among the various deep learning approaches, indicating that more auxiliary data and more accurate observations are beneficial in downscaling. This may be helpful for the statistical downscaling of other meteorological data.