To Accurately and Lightly Downscale the Temperature Field by Deep Learning

Abstract

Abstract Downscaling is essential in atmospheric science, aiming to infer the fine-scale field from the coarse-scale field. To obtain the high-resolution temperature field, our team proposed a deep learning–based model, the China Meteorological Administration land data assimilation system statistical downscaling model (CLDASSD). Inspired by some works in computer vision, we proposed the improved version, Light-CLDASSD, which is a lightweight model with fewer parameters. The modified model has the characteristics of light training and fewer parameters. What is more, we introduced station observation data in the model to make the downscaling results more accurate. Taking temperature as the research object, we performed experiments in the Beijing–Tianjin–Hebei region and downscaled the temperature field from 1/16° (0.0625°) to 0.01°. Experiments show that Light-CLDASSD can get robust results. As for spatial distribution, Light-CLDASSD can reconstruct fine and accurate spatial distribution on complex mountains and reconstruct small-scale characteristics in plain areas that other models cannot achieve. As for temporal change, Light-CLDASSD performs better at local noon and warm seasons. Furthermore, Light-CLDASSD achieves better performance than other models and is comparable with High-Resolution China Meteorological Administration’s Land Assimilation System (HRCLDAS). The root-mean-square error (RMSE) of Light-CLDASSD is 0.08°C lower than HRCLDAS, and the bias distribution is more concentrated at 0°C. This article is an upgrade of the CLDASSD model and preliminary exploration of the back-calculation for high-resolution historical data. Significance Statement This work proposes a deep learning–based statistical downscaling model named Light China Meteorological Administration land data assimilation system statistical downscaling model (Light-CLDASSD), which can downscale the temperature field generated by CLDAS from 1/16° (0.0625°) to 0.01°. Introducing observation data improves the performance, and the model results are comparable to HRCLDAS products. Our research is of great significance to developing high-resolution data and the back-calculation of historical assimilation data.