Continuous vertical air temperature (AT) from in-situ observation is of crucial importance for understanding the atmospheric environment, but the satellite data that have complete spatial coverage lacked vertical in-situ observation data, and the vertical dropsonde data from in-situ observations only were single-point observations. Therefore, this article introduced machine learning algorithms for fusing in-situ observation and multi-satellite data to achieve spatial continuity of vertical data on a large scale. Specially, random forest (RF), support vector regression (SVR), artificial neural network (ANN) and recurrent neural network (RNN) were employed to capture the non-linear relationships between the variables and estimated AT. The pre-training process and fine-tuning process ensured the prediction of AT spatiotemporal distribution. The four models were implemented for three-dimensional AT estimating across China. Additionally, we used the radiosonde observation data to evaluate the accuracy of estimated AT data under conventional weather and typhoon conditions. Our results revealed that the RF model performed the best with the R of 0.9992, the MAE of 0.70 °C, and the RMSE of 1.04 °C at the national scale, followed by the SVR and ANN models. The RNN model exhibited promising results under typhoon conditions, which will be valuable insights for further research on the applicability of machine learning models under different weather conditions. Besides, having a larger sample size does not necessarily result in reduced errors. For instance, the MAE value for SVR in the pressure height range of 100-200 hPa was larger than that in the pressure height range of 300-400 hPa, but the former sample size was 16,324, which was 7433 higher than the latter.
Keywords: Air temperature fusion; Dropsonde; FY3D; FY4A; RF model.
Copyright © 2023. Published by Elsevier B.V.