Fine-Tuning Lightning Nowcasting for a New Domain

Abstract

Accurate lightning nowcasting is critical for mitigating weather-related risks, yet adapting existing predictive models to new spatial domains remains challenging due to computational demands and data requirements. Transfer learning offers a promising solution, but its application in weather nowcasting, particularly for tasks framed as semantic segmentation problems, is still underexplored. In this study, we employed transfer learning techniques to fine-tune the U-Net architecture of LightningCast, originally developed for the contiguous United States (CONUS) region, to predict lightning for the Brazilian domain. Given the distinct meteorological characteristics of Brazil, particularly in regions dominated by tropical systems, there is a compelling motivation to explore fine-tuning LightningCast for this new spatial domain. The methodology involved investigating the impact of fine-tuning different architectural components, comparing fine-tuned models with those trained from scratch, and analyzing the benefits of transfer learning across varying data availability scenarios. The fine-tuned model consistently outperformed the model trained from scratch, achieving superior performance even with limited data—surpassing the original model's results with just 10% of the available training data—9.3% of improvement in the Area Under the Curve for Precision and Recall (AUC-PR) and 12.8% in the Critical Success Index (CSI) at a 35% probability. Spatial analysis revealed improvements in the Critical Success Index (CSI) across most regions, with an average of 5.2%, and significant reductions in false alarms—with a mean decrease of 10.5%, addressing the original model's overestimation issue. These findings highlight the effectiveness of transfer learning in adapting a lightning nowcasting model to new domains, reducing computational demands while improving performance. The publicly available fine-tuning framework developed in this study offers a versatile tool for extending LightningCast or similar U-Net-based models to other spatial regions.