Lightning nowcasting with aerosol-informed machine learning and satellite-enriched dataset

Abstract

Accurate and timely prediction of lightning occurrences plays a crucial role in safeguarding human well-being and the global environment. Machine-learning-based models have been previously employed for nowcasting lightning occurrence, offering advantages in computation efficiency. However, these models have been hindered by limited accuracy due to inadequate representation of the intricate mechanisms driving lightning and a restricted training dataset. To address these limitations, we present a machine learning approach that integrates aerosol features to more effectively capture lightning mechanisms, complemented by enriched satellite observations from the Geostationary Lightning Mapper (GLM). Through training a well-optimized LightGBM model, we successfully generate spatially continuous (0.25° by 0.25°) and hourly lightning nowcasts over the Contiguous United States (CONUS) during the summer season, surpassing the performance of competitive baselines. Model performance is evaluated using various metrics, including accuracy (94.3%), probability of detection (POD, 75.0%), false alarm ratio (FAR, 38.1%), area under curve of precision–recall curve (PRC-AUC, 0.727). In addition to the enriched dataset, the improved performance can be attributed to the inclusion of aerosol features, which has significantly enhanced the model. This crucial aspect has been overlooked in previous studies. Moreover, our model unravels the influence of aerosol composition and loading on lightning formation, indicating that high aerosol loading consisting of sulfates and organic compounds tends to enhance lightning activity, while black carbon inhibits it. These findings align with current scientific knowledge and demonstrate the immense potential for elucidating the complex mechanisms underlying aerosol-associated lightning phenomena.

Lightning Nowcasting Using Solely Lightning Data

“…These include the fact that we only utilize lightning data for prediction, we train our model on significantly less amount of data, and the fact that there is a different temporal and spatial granularity for each model. However, by considering the work of Tippett et al [23], Leinonen et al [27], and Song et al [36], we can see that our metrics, namely precision (POD) and recall (1-FAR) lies in the same range as the ones presented by these papers. This is an advantage point for our model as it uses less data for training.…”

“…In [31], a combination of satellite data recorded over Europe and lightning data recorded by ground-based lightning detection networks are used to train a residual U-Net model for lightning activity prediction [32][33][34][35]. In [36], the GLM data in combination with aerosol features are used to train a Gradient-boosted decision tree model called LightGBM to perform hourly forecasts.…”

Leading role of Saharan dust on tropical cyclone rainfall in the Atlantic Basin

“…Therefore, the SHAP values have already exhibited its exploratory power for analyzing complex relationships (47), especially in atmospheric science research, such as tropical cyclone genesis, lightning prediction, etc. (48)(49)(50). In the present study, by combining XGBoost ML and SHAP values, we successfully identify the relationships between TCR and environmental variables including SST and DOD, most of which can be explained by current knowledge of physical processes with TC.…”