Nowcasting Thunderstorms to Protect Lives in Africa

Central Africa is widely recognized as the most active region for thunderstorms globally, with the highest frequency of lightning strikes occurring near Kifuka in the Democratic Republic of Congo, where over 150 lightning flashes per square kilometer are recorded annually. The absence of accessible early warning systems in many developing countries significantly amplifies the risks associated with lightning. For instance, on August 28, 2020, a catastrophic lightning strike near the Uganda-Democratic Republic of Congo border resulted in the deaths of nine children, with a tenth succumbing while being transported to the hospital. Moreover, the detrimental effects of lightning on critical sectors—such as livestock, forestry, power utilities, aviation, high-tech industries, and public safety—are increasingly evident. A discernible rise in lightning-related fatalities has been observed, potentially attributable to population growth, which increases exposure to thunderstorms, or to changes in thunderstorm frequency driven by climate change. Regardless of the underlying causes, the risk posed to the African population remains significant and appears to be intensifying.

Building on the recent advancements of Denoising Diffusion Probabilistic Models (DDPMs)—which have demonstrated superior performance over adversarial and autoencoder-based frameworks in applications such as image generation, text-to-image synthesis, precipitation nowcasting, and weather forecasting—this research introduces an innovative nowcasting system. The proposed system predicts lightning probabilities up to six hours in advance, with 30-minute intervals, offering a probabilistic and life-saving early warning mechanism tailored for Central Africa.

Specifically, we investigate the potential of DDPMs for lightning nowcasting by adapting spatiotemporal frameworks originally developed for precipitation nowcasting. In essence, diffusion models learn the underlying data distribution Ρ(Χ), where Χ represents the spatiotemporal probability density function of lightning. This is achieved by training the model to reverse a predefined noising process that progressively corrupts the target data with Gaussian noise. Here, the diffusion process has been extended to condition on auxiliary data Υ, such as satellite-derived wavelength imagery, constituting the approach suitable for spatiotemporal conditional nowcasting Ρ(ΧΥ).

As a data source, we leverage recent datasets from the Meteosat Third Generation (MTG) Lightning Imager (LI) over Africa and the Earth Networks Total Lightning Network (ENTLN) to train the model that locally characterizes the stochastic nature of lightning events.