A Convolutional Neural Network for Lightning Strikes Detection Over the Italian Territory Using SEVIRI@MSG Data

Abstract

The study of lightning strikes (LSs), as a direct expression of atmospheric instability and an indicator of convective phenomena expected to increase with global warming, is essential to analyze how climate change manifests and to assess potential impacts on diverse sectors such as civil protection, power grids, and air transport. To support this objective, this work presents a novel methodological framework, representing the first application of a deep learning model to exclusively spinning enhanced visible infra-red imager (SEVIRI) infrared channels for lightning detection over Italy, coupled with independent validation against an external data source. Specifically, a U-Net convolutional neural network model monitoring infrared Italian lightning-UNET (MIRIL-UNET) has been trained to detect lightning occurrences based on three key channel differences: water vapor difference (6.2– 7.3 µm), infrared difference (3.9– 10.8 µm), and water vapor infrared difference (6.2– 10.8 µm). The selection of these channel combinations was carried out through an expert judgment phase, ensuring the physical relevance and discriminative capacity of the input features. The methodology enables continuous monitoring during both day and night by exclusively utilizing infrared channels. The network was trained and validated using lightning data collected from the LAMPINET ground-based detection network of the Italian Air Force during the period 2020–2023. To address the significant class imbalance inherent in lightning detection, a binary focal cross-entropy loss function with class balancing has been employed. The model achieved an average precision of 0.68, a recall of 0.76, and a F-beta score of 0.72 across the Italian territory, with performance depending on the Italian macro area considered. Overall, our results are consistent with those reported in other studies applying a similar methodology. In addition, by exploiting the characteristic signatures of convective systems in satellite channels, they offer improved coverage relative to ground-based detection systems, which inherently struggle to detect intracloud and cloud-to-cloud flashes. Finally, an independent validation using Blitzortung data underscores the robustness and originality of this approach, while demonstrating how the neural network’s performance is influenced by the spatial granularity of the ground-based detection network. This approach highlights the value of satellite, especially in areas where ground-based networks are limited in time or resolution, offering a complementary means to improve overall detection capability.