# How Can Deep Learning Improve 3D Humidity Forecasting? GNSS, SRGAN, and NVIDIA GPUs - Link: https://digiosveny.hu/en/hogyan-javithatja-a-melytanulas-a-paratartalom-3d-s-elorejelzeset-gnss-srgan-es-nvidia-gpu-k/ - Published: 2026-10-11T10:37:26+02:00 - Author: Márton **In Brief:** Accurate mapping of atmospheric humidity is key to forecasting weather extremes. A Polish research team used deep learning to improve the resolution of humidity maps derived from GNSS measurements. The result is a promising research direction, but it does not in itself prove that the method already yields more accurate operational forecasts or earlier warnings. To forecast severe thunderstorms and sudden downpours, it is 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. Researchers at the **University of Environmental and Life Sciences in Wrocław** processed GNSS measurements using deep learning methods and generated higher-resolution, three-dimensional humidity fields . The study was published in the journal _Satellite Navigation_. DOI: ** DOI: 10.1186/s43020-025-00177-6**. The approach centers on a super-resolution generative adversarial network, or **SRGAN**. The researchers trained the model on weather data and used NVIDIA GPUs for the computations. The goal was to convert GNSS-based, lower-resolution data into more detailed humidity maps. ## Why is humidity important in weather forecasting? The moisture content of the air is not merely a comfort variable. 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 region, it can affect how a thunderstorm or precipitation system develops. More accurate measurements can help better describe the state of the atmosphere. However, this does not automatically mean that a specific downpour can be reliably predicted. Forecasting requires multiple data sources and complex modeling. ## How do GNSS systems measure atmospheric moisture? Signals from GNSS systems—including satellite navigation systems—pass through the atmosphere before reaching ground receivers. Water vapor in the atmosphere alters the propagation of these signals. By analyzing these measurement variations, it is possible to infer atmospheric moisture content. These observations are useful because the GNSS infrastructure can provide atmospheric information beyond navigation purposes. However, the measurements do not provide the same level of detail everywhere or at all times. Researchers therefore sought a method capable of reconstructing finer spatial structures from the available information. ## What does the SRGAN do with humidity data? The SRGAN is a deep learning model that is originally used to increase the resolution of images as well. In this research, it was not used to sharpen photographs. The model learned from weather data what more detailed patterns could be fitted to lower- resolution humidity observations. Simplified, the process consists of three steps: 1. **GNSS measurements:** The researchers use data derived from the atmospheric effects on satellite signals . 2. **SRGAN processing:** The deep learning model generates a higher-resolution humidity field . 3. **Verification and interpretation:** The results are compared with data from ground-based instruments, and the areas on which the model relies are examined. NVIDIA GPUs accelerate the computations required to train and run the model. The use of 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 better matched the observations of ground-based instruments. Based on the figures presented, the model also depicted spatial differences that were rendered more indistinctly by more traditional methods. The abstract accompanying the article mentions a 62% error reduction for Poland and a a 52% reduction in error, 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 every region, weather situation, or operational forecasting system. The key finding of the research, therefore, is not that the model can already predict storms on its own. Rather, it is that the detail of 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 to simply see what kind of map a model produces. Experts must also understand what information the model used to arrive at its result. The researchers used the **Grad-CAM** and **SHAP** methods 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, attention was focused, among other places, on the western regions of Poland and the coastal mountains of California. Such visualizations serve as interpretive aids, not standalone proof that the model functions correctly in every situation. Interpretability must be supplemented with performance metrics, testing of various weather conditions, and independent data. ## What practical applications might this method have? The more detailed humidity fields could be used in the future as input data for physically based or artificial intelligence-driven weather models. If further tests confirm the method’s reliability, it could improve the description of the current state of the atmosphere. This could support 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 understandable warnings, and decision-makers are able to take appropriate measures. The research result is therefore one possible component of a longer-term process. It is not equivalent to a fully developed warning system. ## Main Limitations and Future Research Questions 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 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. The moisture content of the atmosphere is also a crucial part of how storms form and evolve. GNSS measurements, SRGAN-based processing, and explainability tools together offer a new opportunity for the spatial mapping of humidity. 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.