A small-scale thunderstorm lightning nowcasting method based on multi-scale feature fusion

Accurate and timely lightning nowcasting has important application value in preventing meteorological disasters and reducing economic losses. We conducted research on the technical difficulties in the nowcasting of small-scale thunderstorm lightning. Considering that most lightning processes in sparse lightning data are associated with small-scale thunderstorms, which have a small spatial scope and evolve rapidly, we proposed an optimized method for small-scale thunderstorm lightning forecasting. A small-scale thunderstorm lightning dataset was constructed through the analysis of the DBSCAN spatio-temporal clustering algorithm. To address the issue of the imbalanced ratio between lightning and non-lightning samples in lightning data, a class-weight adaptive cross-entropy loss function was designed to dynamically adjust the weights and alleviate the class imbalance. A three-dimensional multi-scale refinement module was designed to establish multi-scale spatio-temporal feature associations, enhance the ability to capture lightning activities of small-scale cells, and achieve refined nowcasting of small-scale thunderstorms in sparse lightning data. Experiments have proven that the systematic optimization of the small-scale thunderstorm forecasting model significantly improves the refined forecasting ability for small-scale thunderstorms.