Research on regional zenith tropospheric delay based on neural network technology

Abstract

Tropospheric delay is a primary error source in earth observations and a variety of radio navigation technologies. However, the main problem still present is that not all strategic points around the world will have a GPS receiver. To overcome the shortcoming, a Fusion model was proposed to compensate for errors in prediction of zenith tropospheric delay (ZTD), by incorporating back propagation neural network (BPNN). The input parameters include the surface meteorological data (altitude, atmospheric pressure, absolute temperature and water vapour partial pressure) and Hopfield model predicted ZTD, the output is the difference between the Hopfield predicted and high altitude balloon derived ZTD (also called true ZTD). The datasets covered ∼250 meteorological observation stations within 8∶00 and 20∶00 for May 1st, 10th, 30th, and July 1st, August 1st and September 1st, 2010. Data from 16 uniformly distributed stations were used for training at 8∶00 on May 1st, and the remaining for validation. Modelling results from the Hopfield and current model were compared by reference to the true ZTD. It shows that the current model generates an average RMSE value of 0.0040 m, compared with 0.3445 m for the Hopfield model. In overall, the current model can improve ZTD prediction accuracy by more than 90%. In addition, ZTD predictions from the current model were compared to those obtained directly from GPS data, indicating that our model provides a good alternative for ZTD prediction when the GPS receiver at a specific location is absent.