The tropospheric delay is a major error source for space geodetic techniques, and the performance of its modeling is significantly limited due to the high spatiotemporal variability of the moisture in the lower atmosphere. In this study, global modeling of the tropospheric zenith wet delay (ZWD) was realized based on surface meteorological data obtained from radiosondes and Global Positioning System (GPS) radio occultation (RO) measurements through the random forest (RF) and backpropagation neural network (BPNN) regression analysis. The modeling performance was further validated based on two kinds of global atmospheric profiles for the year 2020. Our results show that the ZWD modeling accuracy gained by two machine learning regression approaches is significantly improved by taking into account surface meteorological parameters, especially the surface water vapor pressure when compared to the Global Pressure and Temperature 3 (GPT3) model. When surface meteorological data are available, the RF-B model yields ZWD estimations with an overall agreement of 3.1 cm in comparison with the sounding profiles and 2.4 cm in contrast to the GPS RO atmospheric profiles. The RF-B is superior to other models based on surface meteorological parameters for ZWD calculation, e.g., the accuracy improves by 21.8–23.8% against the approach by Saastamoinen and 7–12.2% against the formula by Askne and Nordius.

对流层延迟是空间大地测量技术的主要误差源，由于低层大气中水分的时空变化较大，其建模性能受到很大限制。在本研究中，基于无线电探空仪和全球定位系统（GPS）无线电掩星（RO）测量获得的表面气象数据，通过随机森林（RF）和反向传播神经网络实现了对流层天顶湿延迟（ZWD）的全局建模。 BPNN）回归分析。基于2020年两种全球大气廓线进一步验证了建模性能。结果表明，通过考虑地表气象参数，特别是地表水汽，两种机器学习回归方法获得的ZWD建模精度显着提高与全局压力和温度 3 (GPT3) 模型进行比较时的压力。当表面气象数据可用时，RF-B 模型产生的 ZWD 估计与探测剖面相比总体一致性为 3.1 厘米，与 GPS RO 大气剖面相比总体一致性为 2.4 厘米。RF-B 优于其他基于地面气象参数进行 ZWD 计算的模型，例如，与 Saastamoinen 的方法相比，精度提高了 21.8%~23.8%，与 Askne 和 Nordius 的公式相比，精度提高了 7%~12.2%。