针对利用全球导航卫星系统(global navigation satellite system,GNSS)反演高精度大气可降水量(precipitable water vapor,PWV)时需要获取大气加权平均温度(Tm)从而影响结果精度的问题,在充分探究PWV与对流层天顶湿延迟(zenith wet delay,ZWD)等诸多因子相关性的基础上,利用中国南方地区40个探空站在2015-2017年的探空数据,基于多层感知器(multi layer perceptron,MLP)神经网络及多元回归拟合算法分别建立预测PWV的MLP模型、线性回归(linear regression,LR)模型与非线性回归(non-linear regression,NLR)模型.为充分探究2种建模方法对PWV精度的影响,利用2018年探空数据为参考值进行模型精度检验,并与传统PWV预测模型(PWV-SC2模型)进行精度对比分析.结果表明:MLP模型的年均均方根误差(RMSE)、偏差(bias)和相对误差(RE)分别为0.66 mm、0.06 mm和2.18%,相比LR模型和NLR模型年均RMSE分别降低了0.11 mm(14.6%)和0.17 mm(20.5%),年均bias分别降低了0.04 mm(43.7%)和0.28 mm(82.3%),年均RE分别降低了50.7%和57.3%;相比PWV-SC2模型,年均RMSE和bias分别降低了0.17 mm(20.5%)和0.15 mm(71.4%),年均RE降低了47.7%.因此,MLP模型在中国南方地区有较好的精度及适应性,可应用于中国南方地区高精度PWV预测.
To address the problem related to the effects on the precision in the inversion of precipitable water vapor(PWV)using global navigation satellite system(GNSS)arisen by the necessity of obtaining key parameters such as atmospheric weighted mean temperature(Tm),the correlation between PWV and based on the correlation between PWV and the tropospheric zenith wet delay(ZWD)and other factors was investigated.The 2015-2017 sounding data from 40 sounding stations in southern China to establish the MLP model,the linear regression model(LRM)and the multiple regression fitting algorithm to predict PWV based on the multi-layer perceptron(MLP)neural network.MLP model,linear regression(LR)model and nonlinear regression(NLR)model were developed respectively based on the multilayer perceptron(MLP)neural network and multiple regression fitting algorithm.To fully investigate the influence of the two modeling methods on the accuracy of PWV,the accuracy of the model was examined using the 2018 sounding data as the reference value and compared with the traditional PWV prediction model(PWV-SC2 model).The results show that the average annual RMSE,bias,and RE of the MLP model are 0.66 mm,0.06 mm,and 2.18%,respectively,which are 0.11 mm(14.6%)and 0.17 mm(20.5%)lower compared to the LR model and the NLR model in terms of average annual RMSE,0.04 mm(43.7%)and 0.28 mm(82.3%)in terms of average annual bias,and 50.7%and 57.3%lower annual mean RE,respectively;compared to the PWV-SC2 model,the annual mean RMSE and bias are reduced by 0.17 mm(20.5%)and 0.15 mm(71.4%),respectively,and the annual mean RE is reduced by 47.7%.Therefore,the MLP model has better accu-racy and adaptability and can be applied to high-precision PWV prediction in southern China.