High-resolution GNSS troposphere tomography through explainable deep learning-based downscaling framework

Abstract This study presents the first High-Resolution (HR) Global Navigation Satellite System (GNSS) troposphere tomography, which utilizes a Super-Resolution Generative Adversarial Network (SRGAN) in combination with the outputs from the Weather Research and Forecasting model. The resulting HR tomography products have a good potential to enhance the capability of both physics-based and Artificial Intelligence (AI)-based weather forecasting models in capturing small-scale weather phenomena with assimilation techniques. The proposed method is evaluated in two case studies in Poland and California, representing diverse geographical and meteorological conditions. The effectiveness of SRGAN is assessed in both normal and rainy weather conditions, demonstrating significant improvements in downscaled wet refractivity values. SRGAN consistently outperforms the Lanczos3 interpolation method, reducing the Root Mean Square Error (RMSE) by up to 62% in Poland and 52% in California compared to the original tomography data. Additionally, the study incorporates eXplainable AI (XAI) techniques, such as Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP), to analyze the model's decision-making process. The findings indicate that SRGAN focuses on the areas with high atmospheric variability, such as open to weather fronts western part of Poland and the lee side of Transverse Ranges in California, which are affected by dynamic Pacific Ocean weather systems. Overall, the proposed SRGAN method not only enhances spatial resolution in different weather conditions but also provides insights into critical atmospheric regions through XAI analysis.