AI Turns GNSS Signals Into More Detailed 3D Humidity Maps

Humidity often determines whether a cloud produces a few drops of rain or triggers an intense storm. A team at the Wrocław University of Environmental and Life Sciences (UPWr) in Poland has developed a deep learning system to reconstruct more detailed three-dimensional water vapor maps from navigation satellite signals.

The study, published in Satellite Navigation, offers a way to improve one of the most difficult components to observe in weather forecasting. According to results reported by NVIDIA and the researchers, the method reduced estimation error by 62% in tests conducted in Poland and by 52% in California, even under rainy conditions.

Navigation satellites can measure the atmosphere, too

Global navigation satellite systems, known by the acronym GNSS—the technology family that includes GPS—transmit radio signals toward Earth. As those signals pass through the atmosphere, water vapor causes them to be slightly delayed.

That delay makes it possible to infer how much moisture is present in the air column. The problem is that reconstructions produced with this technique, known as GNSS tomography, tend to have limited resolution. They can describe broad trends, but may blur local variations that matter greatly when a storm, torrential rain or flash flood is forming.

The Polish team's approach is to use those measurements as a starting point rather than a final result. The AI reconstructs a much more detailed version of the humidity field, including its three-dimensional distribution.

A generative network to refine weather data

The model used is an SRGAN, short for super-resolution generative adversarial network. It is a type of neural network originally designed to increase image resolution: it learns to transform a low-quality image into one that better preserves edges, textures and details.

In this case, it does not work with photographs. It takes low-resolution atmospheric maps and learns to generate a finer reconstruction of humidity. To train it, the team used global weather data and NVIDIA GPUs, processors particularly well suited to training neural networks because they can perform many calculations in parallel.

The practical difference lies in gradients: rapid changes in humidity between nearby areas. In meteorology, these contrasts can indicate where energy is building up and where convective clouds—those associated with showers and thunderstorms—are more likely to develop.

The work does not replace the physics-based models that solve equations describing the atmosphere. It can serve as an upstream layer, providing them with a more useful representation of available humidity. It could also feed AI-based weather models, which depend on the quality and coverage of their input data.

The explanation matters as much as the result

The team added interpretability tools, a relevant decision in a field where forecasts can trigger public alerts and emergency decisions. They used Grad-CAM and SHAP, two methods that help visualize which regions of the data carried the most weight in a model prediction.

The visualizations showed attention focused on areas prone to complex weather events, such as western Poland and California's coastal mountains. That alone does not prove that the model is reliable in every region or situation, but it allows meteorologists to check that its results are not based on arbitrary patterns.

Saeid Haji-Aghajany, an assistant professor at UPWr, argues that reliable, high-resolution humidity data is a missing piece in forecasting weather events that disrupt daily life. Transparency, he adds, is necessary for professionals to trust AI systems used in forecasting.

From experiment to weather warnings

The reported improvements reflect evaluations of the method against earlier reconstruction techniques, not a guarantee that every operational forecast will reduce its error by the same proportion. Before it can reach weather services, the system will need to be validated across more climates, seasons and observation networks, as well as tested to determine whether improved humidity maps actually result in earlier, more accurate warnings.

Even so, the approach has a clear advantage: it makes use of satellite infrastructure that is already in operation. If GNSS data can reliably complement radar, radiosondes and weather satellites, AI could help detect the small atmospheric changes that precede a severe storm.