The tropospheric mapping functions (MFs) play a crucial role in estimating the delay of electromagnetic waves as they traverse the troposphere, particularly in space geodetic techniques like GNSS and VLBI, where these waves are the primary measurement tool. Currently, empirical models that rely on non-meteorological parameters are commonly used to calculate MFs, offering convenience but often yielding less accurate results in regions with volatile climates. Such inaccuracies can lead to errors in slant total delays reaching the meter-level, subsequently affecting other data products. Although authoritative bodies periodically release essential meteorological parameters for MF calculation or provide ready-made MF products, these often come with a delay of several days or require specific authorization. To address this, we introduced a method leveraging the random forest (RF) model to rapidly derive key MFs from surface observations. Experimental results indicate that the RF model significantly enhances the accuracy of MFs, with improvements exceeding 60% for the hydrostatic component and over 20% for the wet component, effectively eliminating seasonal systematic deviations; although it is somewhat inferior to the MFs from VMF3-FC, RF does not require an internet connection and can independently train and construct models, making it very suitable for independently operated systems. We further trained a series of RF models using varied sample spaces and analyzed their performance under different combinations of feature dimensions and time spans. This analysis suggests that an optimal balance exists when considering the challenges in acquiring sample data, training time, model size, computational efficiency, and accuracy.
Read full abstract