Speaker
Description
Commercial Microwave Links (CMLs) constitute an opportunistic sensing network for providing high-resolution rainfall measurements. In CML-based precipitation retrieval, a fundamental preprocessing step is the wet/dry classification. Traditional approaches rely on direct labeling strategies, where wet/dry states are inferred from nearby radar observations. While operationally simple, such methods may introduce labeling inconsistencies due to radar uncertainties, spatial mismatches, and limited sensitivity to light precipitation.
In this work, we propose an AI-based wet/dry classification framework combining XGBoost and a Tiny Transformer architecture. The XGBoost model exploits instantaneous statistical and physical features derived from link attenuation measurements, while the Tiny Transformer captures short-term temporal dependencies within the attenuation time series. We use dataset from the OpenRainER project, which provides large-scale, time-resolved observations of CML signal attenuation alongside collocated weather radar measurements. Performance evaluation focuses on classification reliability under strongly imbalanced conditions dominated by dry periods. Precision, recall, and F1-score metrics are used to quantify detection skill.
The hybrid framework is compared with a Long Short-Term Memory (LSTM) network. Results indicate that the proposed hybrid architecture provides improved classification stability and detection robustness.