Deep Neural Networks for Refining Vertical Modeling of Global Tropospheric Delay

Abstract

Kinematic airborne platforms are becoming increasingly vital for Earth observation. They highlight the critical need for accurate tropospheric delay corrections across varying altitudes, especially as most existing models are limited to Earth's surface. Although analytical functions have been used to model vertical reductions in tropospheric delays, they struggle to capture the intricate vertical variations of atmospheric state. In response, we introduce a novel approach that utilizes deep neural networks (DNN) to reconstruct global three-dimensional zenith hydrostatic delay (ZHD) and zenith wet delays (ZWD) derived from numerical weather models (NWM). Our method reconstructs NWM-derived ZHD and ZWD globally up to 14 km above the Earth's surface, with average precision levels of 0.4 and 0.8 mm, respectively. Compared to the analytical third-order exponential model, the DNN approach demonstrates substantial improvement with global average root-mean-square reductions of 63% for ZHD and 36% for ZWD.