In short: Accurate mapping of atmospheric humidity is crucial for forecasting extreme weather events. A Polish research team used a deep learning method to improve the resolution of humidity maps generated from GNSS measurements. The result points to a promising direction for research, but it does not, on its own, prove that the method already leads to more accurate operational forecasts or earlier warnings.
To forecast severe thunderstorms and sudden downpours, it’s not enough to know where the clouds are moving. The moisture content of the atmosphere also matters. Water vapor influences how precipitation-bearing systems form and intensify. Yet it is difficult to measure with high spatial resolution.
A Polish research team investigated how this could be improved. The University of Environmental and Life Sciences in Wrocław The researchers processed the GNSS measurements using deep learning methods and generated higher-resolution, three-dimensional humidity fields. The study Satellite Navigation It was published in a journal. Its identifier is: DOI: 10.1186/s43020-025-00177-6.
At the heart of this approach is a super-high-resolution generative adversarial network, that is, SRGAN The researchers trained the model using weather data and used NVIDIA GPUs for the calculations. The goal was to convert GNSS-based, lower-resolution data into a more detailed humidity map.
Why is humidity important in weather forecasting?
The humidity of the air is not merely a comfort factor. Water vapor in the atmosphere plays a role in cloud formation and the development of precipitation. The spatial and vertical distribution of moisture can therefore provide important information when analyzing weather processes.
If humidity changes rapidly in a given area, it can affect how a thunderstorm or precipitation system develops. More accurate measurements can help provide a better description of atmospheric conditions. However, this does not automatically mean that a specific downpour can be predicted with certainty. Forecasting requires multiple data sources and complex modeling.
How do GNSS systems measure atmospheric humidity?
Signals from GNSS systems—including satellite navigation systems—pass through the atmosphere before reaching receivers on the ground. Water vapor in the atmosphere alters the propagation of these signals. By analyzing measurement discrepancies, it is possible to infer the moisture content of the atmosphere.
These observations are useful because the GNSS infrastructure can provide atmospheric information beyond navigation purposes. However, the measurements do not provide images of the same level of detail everywhere or at all times. Researchers therefore sought a method capable of reconstructing a finer spatial structure from the existing information.
What does SRGAN do with the humidity data?
SRGAN is a deep learning model that is originally used to increase the resolution of images. In this study, it was not used to sharpen photographs. The model learned from weather data which finer-scale patterns could be superimposed onto lower-resolution humidity observations.
In simple terms, the process consists of three steps:
- GNSS measurements: Researchers use data derived from the atmospheric effects on satellite signals.
- SRGAN processing: The deep learning model generates a higher-resolution humidity field.
- Verification and Interpretation: The results are compared with data from ground-based instruments, and researchers examine which areas the model relies on.
NVIDIA GPUs accelerate the computations required to train and run the model. Using a GPU alone does not guarantee more accurate results. Performance is determined by a combination of data quality, model architecture, the training process, and independent validation.
What did the results show?
Compared to previous methods, the study presented more detailed humidity patterns. According to the authors, the generated fields were a better fit for observations from ground-based instruments. Based on the figures presented, the model also depicted spatial differences that were represented less clearly by more traditional methods.
The summary accompanying the article mentions a 62% reduction in Poland and a 52% reduction in California, even under rainy conditions. These values must be interpreted within the context of the data and comparison methods used in the study. They cannot be automatically generalized to all regions, weather conditions, or operational forecasting systems.
An important finding of the research, therefore, is not that the model can already predict storms on its own. Rather, it is that the level of detail in humidity maps derived from GNSS data can be improved, and this warrants further investigation.
Why is explainable artificial intelligence important?
The results of weather models can influence decisions. Therefore, it is not enough simply to see what kind of map a model produces. Experts must also understand what information the model used to arrive at that result.
The researchers Grad-CAM and SHAP methods were used to analyze the model's attention. These techniques can help reveal which input areas or features contributed to the model’s output. In the study, the focus was on, among other areas, the western regions of Poland and the coastal mountains of California.
Such visualizations serve as interpretive aids, not as standalone evidence that the model works correctly in every situation. Interpretability must be supplemented by performance metrics, testing under various weather conditions, and independent data.
What practical benefits might this method offer?
These more detailed humidity fields could be used in the future as input data for physics-based or artificial intelligence-driven weather models. If further tests confirm the reliability of the method, it could improve the description of the current state of the atmosphere. This could aid in the analysis of heavy precipitation and other rapidly changing weather conditions.
For communities, more accurate forecasts provide a real benefit only if the information reaches them in time, is translated into a clear warning, and decision-makers are able to take appropriate action. The research findings are therefore one possible component of a longer-term process. It is not the same as a fully developed early warning system.
The main limitations and follow-up questions of the research
The model’s results must always be evaluated in the context of the data used, the geographic areas, and the testing conditions. Future studies should clarify how the method performs in other regions, during different seasons, and under various weather conditions. Another important question is to what extent higher resolution improves the results of the entire forecasting chain.
Before operational deployment, it is also necessary to verify how the model behaves when faced with incomplete or noisy data, and how stable its results are in cases not encountered during training. These questions do not diminish the value of the research. They determine what additional evidence is needed for practical implementation.
The bottom line: More detailed measurements of humidity can improve our understanding of the atmosphere
It’s not just thunder and lightning that matter. Atmospheric moisture content is also a key factor in how storms form and evolve. GNSS measurements, SRGAN-based processing, and interpretability tools together offer a new way to map humidity spatially.
The results from the Polish research team are promising, but the practical value of the method must be confirmed by further, extensive validation. The next step is to investigate whether more detailed maps measurably improve the performance of weather models and warning systems.
Source: Satellite Navigation, DOI: 10.1186/s43020-025-00177-6.