Artificial Intelligence for Convective Regions Segmentation in Radar Images

The WIVERN (WInd VElocity Radar Nephoscope) mission, proposed under ESA’s Earth Explorer 11 programme, aims to deliver global, three-dimensional observations of atmospheric winds, alongside measurements of reflectivity and brightness temperature linked to clouds and precipitation. A critical requirement for fully exploiting WIVERN data is distinguishing between convective and stratiform cloud regions. Convective areas, marked by strong vertical motions (|w| > 1 m/s), are associated with intense precipitation and severe weather, while stratiform regions are broader with weaker vertical motion. This distinction is essential for improving numerical weather prediction (NWP), as WIVERN’s horizontal wind measurements are only valid in stratiform zones. This thesis proposes developing an Artificial Intelligence (AI) model for binary convective/stratiform classification using WIVERN observables such as vertical profiles of radar reflectivity and Doppler velocity- The selected model is the U-Net neural network, widely used in biomedical image segmentation, here adapted for meteorological radar data. The network will be trained on synthetic datasets generated by WIVERN’s end-to-end (E2E) simulator, which mimics radar observations based on high-resolution atmospheric models and accounts for radar physics, geometry, and satellite dynamics. This allows for large-scale, labelled training data without relying on costly field campaigns. The model will perform pixel-wise classification of simulated radar images, labelling each point as convective or stratiform based on the strength of vertical motions. The study will also explore the potential for further classification and analyze the influence of dataset characteristics on performance. Evaluation metrics will include Precision, Recall, F1-score, and Equitable Threat Score (ETS). The goal is to create a reliable tool to support WIVERN’s scientific objectives and enhance global weather prediction.