New mean tropospheric temperature models based on machine learning algorithms for Brazil

Precipitable Water Vapour (PWV) plays an essential role in atmospheric science. Integration between meteorology stations and Global Navigation Satellite Systems (GNSS) receivers has enabled high-resolution space-time PWV retrieval. However, the quality of PWV values from GNSS depends on the availability of a Weighted Mean Temperature Model (Tm). Various Tm models have been developed using statistical-based methods. In this contribution, we present new Tm models for the Brazilian region based on Machine Learning techniques, leveraging radiosonde data provided by the Brazilian Institute of Space Research (INPE). We used radiosonde data from 1961 to 1993 for model training and data from 1999 to 2002 for the evaluation phase to assess model performance under usage conditions. Our study employs the following Machine Learning (ML) methods: Random Forest Regression, Support Vector Regression, Recurrent Neural Networks, Gated Recurrent Unit Neural Networks, and Long Short-Term Memory Neural Networks. We conducted a comparative analysis with the traditional Multiple Linear Regression method. The results reveal that there is no universally superior method; the choice of method depends on the region. Furthermore, our findings suggest that integrating classical statistical methods with ML approaches may enhance existing Tm models.