AbstractAbstract
This study evaluates the performance of two deep learning architectures, U-Net and convolutional long short-term memory (ConvLSTM), for predicting cloud-to-ground (CG) lightning probability in the central Amazon basin. Utilizing data from S-band weather radar and very low-frequency (VLF) lightning detection networks, we explored nowcasting capabilities at 12- and 6-min intervals. Various data configurations were tested, specifically examining the integration of volumetric radar reflectivity (dB
Z
) and echo-top 35-dBZ (ET35). The evaluation encompassed objective skill metrics and spatial error analyses, accounting for sensor uncertainties. Results reveal a critical divergence in multimodal data assimilation. While the U-Net struggled with volumetric inputs, the ConvLSTM architecture successfully leveraged dB
Z
data, achieving the highest critical success index (CSI) of 0.546 and an effective spatial scale of 6 km, notably outperforming persistence baselines. Conversely, U-Net performed best using only quality-controlled lightning data or lightweight ET35 inputs. Furthermore, 6-min experiments demonstrated that extending the temporal lookback window notably enhances the skill of recurrent models. Spatial error analyses indicated well-collocated clusters with average centroid errors below 6 km at optimal probability thresholds. These findings suggest that while U-Net provides a robust baseline, ConvLSTM is superior for integrating high-dimensional radar variables in very short-term lightning nowcasting.
Significance StatementThis exploratory study evaluates and compares emerging computational approaches for meteorological forecasting, specifically examining the effectiveness of U-Net and convolutional long short-term memory (ConvLSTM) architectures in predicting cloud-to-ground lightning probabilities in the central Amazon basin. By integrating data from weather radar and lightning detection networks, the research investigates methodological solutions for very short-term lightning prediction (6–12 min) in one of Earth’s most climatically complex regions. The comparative analysis reveals a critical distinction in data assimilation capabilities. While the ConvLSTM successfully leverages volumetric radar reflectivity to achieve effective spatial scales of 6 km, the U-Net performs optimally with simplified inputs like echo tops. This work identifies robust configurations that surpass traditional persistence baselines, laying the foundation for next-generation early warning systems capable of mitigating the socioeconomic impacts and risks of human fatalities caused by lightning.
