Nowcasting Echo Top for Aviation Operations Using CNN-Transformer
Shijin Wang, Min Sun, Y P Li · IEEE Transactions on Intelligent Transportation Systems · 2025
In recent years, convective weather has become one of the leading factors causing flight delays. Accurate nowcasting of the location and altitude of severe convective weather is essential for reducing delays, optimizing airspace utilization, and enhancing the operational efficiency of air traffic systems. To meet the operational needs of civil aviation, this study focuses on Echo Top (ET)—a key vertical parameter reflecting the impact of severe convective weather on flight operations. We developed a spatiotemporal ET nowcasting model using a Convolutional Neural Network- Transformer (CNN- Transformer) deep learning architecture. The model integrates observational variables such as ET, Vertically Integrated Liquid (VIL), and Basic Reflectivity (BR) from Doppler weather radar, indicative of precipitation intensity, along with global meteorological ERA5 reanalysis data provided by the European Centre for Medium-Range Weather Forecasts, including relative humidity, temperature, wind components (U and V), and vertical velocity. The model was trained on 23,881 samples of severe convective weather. Parameter ablation experiments were conducted to enhance the model’s ability to predict severe convective weather affecting flight operations up to 6 hours in advance. Compared with other nowcasting models such as Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), CNN-CNN, and CNN-LSTM, the CNN-Transformer achieved superior performance across five accuracy metrics: Relative Quantile Error (RQE), Anomaly Correlation Coefficient (ACC), Root Mean Squared Error (RMSE), Probability of Detection (POD), and Accuracy. These metrics reflect prediction performance from different perspectives. Consequently, the developed model provides effective decision-making support for altitude-based rerouting.