Web Analytics

Forecasting of Tropospheric Delay Using AI Foundation Models in Support of Microwave Remote Sensing

  • 1The Hong Kong Polytechnic University

  • 2Shanghai Astronomical Observatory

  • 3UCAS

  • 4Shanghai Key Laboratory of SNPT

  • 5GeoForschungsZentrum (GFZ)

  • 6Technische Universität Berlin

  • 7Curtin University

  • 8ETH Zurich

  • 9Shanghai AI Laboratory

Global Tropospheric Delay based on AI Models

This work enables users to generate real-time and forecast global tropospheric delay products locally for geodetic and remote sensing applications. It is built on weather forecasting foundation models such as Huawei Cloud's Pangu-Weather, Google DeepMind's GraphCast, and Shanghai AI Lab's FengWu. The solution supports real-time forecasting of global tropospheric delay products with hourly resolution, allowing local inference and computation to be completed in just a few minutes on any user's local device.


Advantages

Comparison Item Proposed AI solution Traditional (e.g. VMF3_FC)
Accuracy Higher Accuracy 🏅🏅 Lower Accuracy
Prediction Epoch 60 (15 days) or more 🏅🏅🏅 4 (1 day)
Release Delay Instant 🏅🏅🏅 ~9.5 hours
Access Method Any user can perform local inference calculation 🏅🏅🏅 Download from official website (only authorized users can get current year's data)
Service Scope Anywhere on the Earth (include overhead) 🏅🏅 Limited specific ground stations and grid points
Model Error Can directly calculate slant path delay locally 🏅 Cannot avoid errors caused by mapping functions, gradients, and grids interpolation

Approach

The solution first obtains the initial field, then feeds it into the foundation model for inference. Using the forecast atmospheric state output by the inference, it computes to generate real-time and forecast tropospheric delay products—such as zenith delay, mapping functions, and gradients—as well as weighted mean temperature and precipitable water vapor.

Products

The product format follows the VMF3 standard and is available in both grid-wise and site-wise versions. The site-wise version includes additional GNSS stations from NGL and radiosonde stations from IGRA. In addition to three AI foundation model-based versions, the product also offers the ERA5-based version.

In addition to providing products on global grid points, it also offers site-wise products, covering stations from four geodetic techniques (GNSS, VLBI, SLR, DORIS), as well as over 22,000 GNSS stations from NGL and more than 2,600 radiosonde stations from IGRA. The geographical distribution of these stations is shown below.

The current version supports the following tropospheric delay parameters:

No. Types Variables Full name
1 Meteorological p Pressure in [hPa]
2 T Temperature in [°C]
3 e Water vapour pressure in [hPa]
4 Tropospheric ZHD Zenith hydrostatic delay in [m]
5 ZWD Zenith wet delay in [m]
6 ah Hydrostatic mapping function coefficient "a"
7 aw Wet mapping function coefficient "a"
8 SHD Slant hydrostatic delay at 3 ° elevation angle  in [m]
9 SWD Slant wet delay at 3 ° elevation angle  in [m]
10 Linear Horizontal Gradients Gnh hydrostatic north gradient in [mm]
11 Gnw wet north gradient in [mm]
12 Geh hydrostatic east gradient in [mm]
13 Gew wet east gradient in [mm]
14 Meteorology Tm Weighted mean temperature  in [K]
15 PWV Precipitable water vapor  in [mm]

A Demo of the Products

轮播GIF

Abstract

This paper presents for the first time a geodetic and remote sensing tropospheric delay correction scheme based on AI weather forecast foundation models. Based on three foundation models, namely Pangu-Weather, GraphCast and FengWu, the tropospheric delay parameters for the entire year of 2022 were calculated, including zenith delay, mapping function and horizontal gradient, and the new scheme was evaluated and tested. The results show that the new scheme can generate high-precision 15-day 6-hour resolution tropospheric delay forecasts for any location on the globe within minutes at the user's local site. The forecast results are superior to the current best forecast model product VMF3_FC in terms of forecast length, forecast delay and accuracy. We suggest using the scheme proposed in this paper to replace the existing forecast product generation methods. In addition, we also suggest that users with the conditions and needs can bypass the mapping function model and gradient model and directly calculate the slant path delay at the user's local site to avoid the errors brought by these two types of models and grid interpolation. The number of pressure levels supported by the model and the accuracy of the initial input NWM are very important for tropospheric delay forecasting based on large models. We hope that more large models supporting 37-layer input will emerge in the future. In addition, increasing the weight of space geodetic observations in NWM numerical assimilation will also help improve the accuracy of this scheme.

BibTeX

@article{ding2024forecasting, title={Forecasting of tropospheric delay using AI foundation models in support of microwave remote sensing}, author={Ding, Junsheng and Mi, Xiaolong and Chen, Wu and Chen, Junping and Wang, Jungang and Zhang, Yize and Awange, Joseph L and Soja, Benedikt and Bai, Lei and Deng, Yuanfan and others}, journal={IEEE Transactions on Geoscience and Remote Sensing}, year={2024}, publisher={IEEE} }

Total Visits: