Deep learning model for predicting daily IGS zenith tropospheric delays in West Africa using TensorFlow and Keras

Samuel Osah, Akwasi Afrifa Acheampong, Collins Fosu, Isaac Dadzie

Advances in Space Research · 03 May 2021

In one sentence. Deep learning model (DLztd) predicts daily GNSS zenith tropospheric delay at four West African stations from surface meteorology, filling data gaps.

Abstract

Accessibility and precise modelling of tropospheric delay play a significant role in the precise Global Navigation satellite system (GNSS) positioning applications as well as meteorological studies and weather forecasting. However, if in the event that a GNSS Continuously Operating Reference Station (CORS) is inaccessible due to power outages, poor internet connectivity, equipment failure, and firmware issues, gaps are created in the data archive, and the quality of the tropospheric delay estimation is degraded. Over the years, several modelling approaches and methodologies have been proposed towards the precise estimation of tropospheric delay, owing to the spatiotemporal variability of water vapour content in the atmosphere. This study employs Deep learning (DL) approach with TensorFlow and Keras to develop a predictive model (DLztd) for predicting daily IGS final ZTDs over four selected IGS stations in West Africa. Daily surface meteorological parameters (Pressure (P), Temperature (T), and Water vapour partial pressure (e)), as well as daily ZTD and stations’ coordinates (latitude, and ellipsoidal height) obtained from the site-wise VMF3-ZTD products for the period 2015–2018, were used as input variables to train and test the model, while data from 2019 were used to evaluate the predictive performance of the developed model. Statistical performance indicators such as Mean Bias (MB), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), coefficient of determination (R2), Nash-Sutcliffe coefficient of Efficiency (NSE), and the fraction of prediction within a Factor of Two (FAC2) were employed to determine the degree of agreement between the DLztd model predictions and IGS final ZTD data. The results from the various analyses indicate exceptionally good prediction capability of the DLztd model with average MB, RMSE, MAPE, R2, NSE and FAC2 of 3.25 mm, 9.62 mm, 0.30%, 0.959, 0.947, and 1.00 respectively. This demonstrates that the DLztd model provides a remarkable alternative for improving the availability of the ZTD data over the IGS stations under study should the stations' data be inaccessible or unavailable.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.