For more than a century meteorologists have advanced forecasting with equations and supercomputers, yet water vapor remains a persistent blind spot. A team at Wrocław University of Environmental and Life Sciences published a paper in Satellite Navigation describing how deep learning can convert blurry global navigation satellite system based atmospheric snapshots into sharp three dimensional humidity maps. The authors trained a super-resolution generative adversarial network, or SRGAN, on global weather data and ran the model on NVIDIA GPUs to produce upscaled humidity fields from GNSS tomography inputs.
The approach substantially reduced retrieval errors in demonstrations. In Poland the technique cut errors by 62 percent, and in California it delivered a 52 percent reduction, including in rainy conditions when forecasts are most challenged. Compared with older tomographic methods that blurred structure, the SRGAN produced sharp humidity gradients that matched observations from ground instruments. The paper notes that the model was trained to preserve meteorological detail rather than smear it into a watercolor effect, improving the representation of the invisible moisture structures that feed storms, flash floods, and hurricanes.
The team also embedded explainability into the workflow. Using visualization tools such as Grad-CAM and SHAP the researchers showed where the model focused when making predictions, and those regions aligned with known storm-prone features like Poland´s western borders and California´s coastal mountains. Lead author Saeid Haji-Aghajany, assistant professor at UPWr, said the method not only sharpens GNSS tomography but also reveals how the model reaches decisions, a step intended to build trust as Artificial Intelligence enters operational forecasting. Sharper humidity fields can be fed into physics based or Artificial Intelligence driven weather models to improve the detection of sudden downpours and flash floods and provide communities with earlier warnings. Reference: DOI: 10.1186/s43020-025-00177-6.
