As a key indicator of the occurrence of severe convection, convective initiation (CI) exhibits characteristics such as fragmentation, scale heterogeneity, and susceptibility to confusion with other cloud systems in single-temporal remote sensing imagery, posing significant challenges for accurate CI detection. Traditional threshold-based methods inadequately capture spatial representations and have limited generalization capabilities, while existing deep learning approaches fail to fully utilize the temporal correlation features of the same target cloud cluster, resulting in a high false alarm rate. To address these challenges, based on the physical laws of convective development, we propose a spatiotemporal feature fusion-based CI detection model, namely Ti-UHRNet. The model integrates three core designs: integrating digital elevation model geographic information at the input layer to quantify the topographic modulation on convective development and enhance the physical consistency of features; adopting U-HRNet embedded with attention-gated feature fusion as the backbone to extract multi-scale features efficiently, filter critical information dynamically, and retain high-resolution spatial details of convective clouds; and designing a multi-head self-attention-based TransTrack module with multi-temporal inputs to capture the dynamic evolution information of convective clouds within a 15-minute window, thereby distinguishing them from other cloud systems. Experimental results show that compared with several advanced 2D and 3D convolutional segmentation methods, Ti-UHRNet achieves the best performance in extracting the spatiotemporal features of rapidly developing convective cloud clusters. On the test set, it attains a probability of detection of 0.954, a false alarm rate of 0.082, and a critical success index of 0.879. Verified against ground-based radar echoes, the model enables effective early warning of severe convective weather at 15–30 minutes in advance.
Tao et al. (Thu,) studied this question.