Developing a new combined model of zenith wet delay by using neural network

Abstract

The motivation of this research is to develop a more accurate global empirical model of zenith wet delay (ZWD) based on a neural network model, which was carried out by introducing a new combined modeling strategy. This strategy was based on two important features about ZWD: the ZWD seasonal variations and ZWD being relevant to site meteorological measurements. The water vapor pressure from GPT2w (eS-GPT2w), the water vapor decrease factor from GPT2w (λGPT2w), and the weighted mean temperature of GTrop (Tm-GTrop) can provide the approximate site seasonal ZWD and then were set as input variables of neural network. With a combination of approximate site seasonal ZWDs and site meteorological measurements, more accurate ZWD estimates can be provided. Based on this new combined modeling strategy, the new ZWD neural network (N-ZWDNN) model was developed. On the other side, the author followed conclusions of the previous studies about zenith tropospheric delay models using neural network, then extended modeling scope to a global scale and developed the traditional ZWD neural network (T-ZWDNN) model. We compared N-ZWDNN with T-ZWDNN on the condition that they were modeled with the same data set. Other two state-of-the-art ZWD models. i.e. the site-augmented GPT2w-ZWD (SAGPT2w-ZWD) model, and the GridZWD model are also included for the comparison. The results show that N-ZWDNN has a global accuracy of 17.8 mm and shows respective 13.6%, 20.7% and 10.9% improvements in accuracy compared with the SAGPT2w-ZWD, GridZWD and T-ZWDNN models for global ZWDs from the earth’s surface to 10 km. N-ZWDNN can maintain its superiority in different latitude belts, height ranges and seasons. The new combined modeling strategy and the great non-linearity fitting capability of neural network mostly contributed to the accuracy improvement of N-ZWDNN. In addition, a simplified version of N-ZWDNN (N-ZWDNN-B) was given and site meteorological measurements were not used in this model.