In today's deep learning-dominated era, real-time classification of public emergencies is a critical research area. Existing methods, however, often fall short in considering both temporal and spatial aspects comprehensively. This study introduces GEDNAS, a novel model that combines atrous convolutional neural network (DCNN), gated recurrent unit (GRU), and neural structure search (NAS) to address these limitations. GEDNAS utilizes DCNN to capture local spatio-temporal features, integrates GRU for time series modeling, and employs NAS for overall structural optimization. The approach significantly enhances real-time public emergency classification performance, showcasing its efficiency and accuracy in responding to real-time scenarios and providing robust support for emergency response efforts. This research introduces an innovative solution for public safety, advancing the application of deep learning in emergency management and inspiring the design of real-time classification models, ultimately enhancing overall societal safety.