Abstract Hailstorms pose a significant hazard, although how hail frequency may respond to climate change is still uncertain and varies by region. In addition, hail is spatially and temporally sporadic and depends on ground‐based reports, so observational samples of hail events are difficult to collect. To address this issue, this study proposes a deep‐learning‐based hail‐diagnosis model (DHDM). We developed a seven‐layer convolutional neural network with residual blocks and a masking mechanism. This model enhances hail‐detection capability under limited sample conditions. The results show that the DHDM has a hit rate of 80.7% and a false‐alarm rate of 50.9% for hail, under the condition of a dataset constructed with a 1:4 ratio of hail to non‐hail samples. Model interpretability analysis reveals that the incorporation of the masking mechanism enhances its focus on circulation patterns at 500‐ and 850‐hPa levels, thereby improving its capability to learn key features at these critical pressure levels. This suggests that the masking mechanism acts as a regularization and data augmentation strategy that encourages the model to learn more robust and distributed representations, rather than relying on fixed local patterns. This study highlights the potential of deep learning in severe convective weather forecasting and provides a reference for integrating masking mechanisms into meteorological applications to address the issue of learning with limited samples.
Zeng et al. (Sun,) studied this question.