Explaining neural networks for detection of atmospheric surface fronts in gridded atmospheric simulation data

Atmospheric features, including tropical cyclones, atmospheric rivers, extratropical cyclones, and atmospheric fronts (AFs), are important for understanding and predicting the weather. Hence, automated detection of atmospheric features in gridded datasets is used for weather forecasting, statistical and climatological studies, and visual data analysis. Typically, rule-based detection systems are used; however, in recent years, numerous studies have demonstrated the use of machine learning, especially convolutional neural networks (CNNs), to detect atmospheric features. While it was shown that CNNs can detect atmospheric features similarly well to human experts, they are “black box” systems. Therefore, whether the features are detected based on physically plausible patterns is unknown. In a recent study (Radke et al. 2025, Geosci. Model. Dev.) we showed how the explainable artificial intelligence technique “Layer-wise Relevance Propagation” (LRP) can be used to understand the patterns used by a CNN detecting tropical cyclones and atmospheric rivers. In this presentation, we build upon this study as well as recent work in AF detection with CNNs and use LRP to understand the detection patterns learned by a CNN detecting AFs. As hand-labeled data of AFs is only sparsely available, CNNs for AF detection are only trained on local regions.

We find that the patterns used by a CNN detecting AFs are not entirely plausible when compared to rule-based detection systems. This leads to erroneously detected AFs, especially outside the regions used for training the CNN. Taking a closer look at regions outside the training regions, we find that using CNNs to detect AFs outside their training region, for example, in the tropics, leads to poor results. To increase the detection quality, we explore an extension of the data to the tropics and show that additional data improves detection results; however, as a dataset with globally labeled AFs does not exist, further research is required before CNNs can be used to reliably detect AFs globally.