Contribution Disparity and Key Factor Screening of Lightning Identification and Nowcasting with Multi-Source Data

Radar and geostationary satellite observations provide essential information for lightning identification and nowcasting. However, multi-source channels commonly exhibit correlation and redundancy, and directly using all channels increases computational cost and may degrade model performance. This paper focuses on lightning identification and 0–1 h nowcasting, integrating radar composite reflectivity, Himawari-8/9 AHI multi-channel observations, and VLF-LLN lightning location data to construct a multi-source spatiotemporally aligned dataset. Using baseline deep-learning-based identification and nowcasting models, this paper proposes a workflow for analyzing multi-source channel contribution differences and selecting key factors by integrating permutation feature importance (PFI), SHAP attribution, channel collinearity grouping, and validation-set wrapper search. The results show that radar composite reflectivity has a stable dominant contribution in both tasks. The final selected channels are radar composite reflectivity, $\Delta$Band 13, and Band 14 + Band 09 for the identification task, while they are radar composite reflectivity, Band 14, (Band 11 − Band 13) − (Band 13 − Band 15), and $\Delta$(Band 08 − Band 13) for the nowcasting task. Independent test results show that under fixed thresholds, the three-channel identification model achieves a CSI of 0.433, outperforming the full 28-channel model with a CSI of 0.415, while the four-channel nowcasting model achieves a CSI of 0.284, outperforming the full-channel model with a CSI of 0.274. These results indicate that the proposed method can reduce input dimensionality while maintaining or improving model performance, and it can support multi-source predictor selection, lightweight deployment, and operational interpretation in lightning nowcasting.