A deep learning method for predicting lower troposphere temperature using surface reanalysis

Abstract

Planetary boundary layer (PBL) is formed by the interaction of the atmosphere and the surface and surface conditions have an important influence on the PBL state. This study developed a method to retrieve temperatures in the lower troposphere using temperature reanalysis at 2 m (T2) with a U-net based deep learning model (briefly, lower troposphere retrieval with U-net, LTRUnet). High-resolution four-dimensional data assimilation (FDDA) reanalysis data over Southern China was selected to demonstrate the ability of the deep-learning network. The normalization method and the hyper-parameters of LTRUnet were determined through experiments. The results show that LTRUnet can project the temperatures in the layers below about 1.3 km with good accuracy, with a Mean Absolute Error (MAE) <0.5 K on the test set compared to the reanalysis. The performance of LTRUnet deteriorates with height and has a strong diurnal variation. By utilizing auxiliary physical information of the terrain height and the first-guess temperature (Tb) as additional channels of the LTRUnet input, the error (MAE) of the projected temperature was further reduced by 19.5%. Using the first guess background fields also allows LTRUnet to achieve the same prediction performance with much less training data. Sensitivity tests of the input variables indicate that the topographic information helps the model to better predict the topography-induced temperature distribution, and Tb plays a more significant role in higher levels than T2. Finally, by training LTRUnet independently for different phases of PBL with a physics-guided data grouping strategy, the model results are further improved.