Near-surface air temperature estimation for areas with sparse observations based on transfer learning

Abstract

Near-surface air temperature (NSAT) data is essential for climate analysis and applied research in areas with sparse ground-based observations. In recent years, machine learning (ML) techniques have been increasingly used to estimate NSAT, delivering promising results. However, in regions with limited observational samples, machine learning-based NSAT estimations may encounter challenges, potentially resulting in reduced accuracy. Therefore, this study introduces a novel model – TranSAT. TranSAT utilizes a transfer learning (TL) framework, deep neural network (DNN) and U-shape convolutional network (U-Net) to enhance the accuracy of NSAT estimations in regions with sparse observational data. Considering the scarcity of observation stations in certain regions, the Tibetan Plateau within the China's landmass (CNTP) is selected as the study region. The majority of observational stations are concentrated in the eastern and southeastern parts of CNTP, with a significant lack of stations in the northern and western regions. The scarce observations in the CNTP affect NSAT estimation accuracy in recent studies, thus limiting practical applications. The estimated NSAT (i.e., TranSAT NSAT) by TranSAT is evaluated by measurements of 10 independent meteorological stations from the Meteorological network in China's cold region (MSC). Evaluation results indicate an average coefficient of determination (R2) of 0.92 and a root mean squared error (RMSE) of 2.29 °C. The TranSAT NSAT exhibits an overall decrease of at least 7 % on RMSE compared to existing NSAT datasets, with a more significant enhancement of over 40 % in regions with sparse ground observations. These results highlight the good and consistent performance of TranSAT NSAT, further confirming that the proposed TranSAT model effectively improves NSAT estimation in areas with limited observational data.