基于随机森林和反向传播神经网络机器学习方法的区域ZTD建模精度分析

Accuracy Analysis of Regional ZTD Modeling Based on Random Forest and Back Propagation Neural Network Machine Learning Method

针对常用的GPT2w和UNB3m两种区域(经验)对流层天顶总延迟(ZTD)模型精度不高的问题,探讨基于机器学习方法进行区域ZTD建模的可行性。以GAMIT软件解算的美国加州13个IGS测站2021年连续31 d的ZTD数据(ZTD_GAMIT)为例,构建以经度、纬度、大地高、年积日、每日小时数、GPT2w或UNB3m经验ZTD模型估计的ZTD值(ZTD_GPT或ZTD_UNB)为输入,以ZTD_GAMIT为输出的随机森林(RF)和反向传播神经网络(BPNN)区域ZTD改进模型。实验结果表明,相较于GPT2w和UNB3m模型,两种基于机器学习方法的区域ZTD改进模型的预测精度均有所提高,能有效改善系统偏差。以ZTD_UNB为输入的BPNN和RF改进模型的预测均方根误差(RMSE)分别为15.14 mm和19.48 mm,以ZTD_GPT为输入的BPNN和RF改进模型的RMSE分别为15.32 mm和20.74 mm。BPNN模型的预测精度总体上优于RF模型,具有较高的可靠性。

Aiming at the low accuracy of GPT2w and UNB3m regional tropospheric zenith total delay(ZTD)model,we discuss the feasibility of regional ZTD modeling based on machine learning method.Taking ZTD data calculated by GAMIT software(ZTD_GAMIT)for 31 consecutive days in 2021 from 13 IGS stations in California as an example,we propose an improved ZTD model using ZTD values estimated by longitude,latitude,geodetic height,day of year,daily hours,GPT2w or UNB3m empirical ZTD model(ZTD_GPT or ZTD_UNB)as inputs and ZTD_GAMIT as outputs based on random forest(RF)and back propagation neural network(BPNN).The experimental results show that compared with the GPT2w and UNB3m models,the prediction accuracy of the two improved regional ZTD models based on machine learning methods is improved,and the system bias is effectively improved.The root mean square error(RMSE)of BPNN and RF improved models with ZTD_UNB as inputs are 15.14 mm and 19.48 mm,respectively.The RMSE of BPNN and RF improved models with ZTD_GPT as inputs are 15.32 mm and 20.74 mm,respectively.The prediction accuracy of BPNN model is generally better than that of RF model,and has higher reliability.