To address the urgent rescue demands during ship cabin flooding accidents, this paper propose a model for rapid flooding time prediction in emergency flooding scenarios. The study utilizes towing tank-based real ship experimental data. The proposed dual-stream four-dimensional feature decoupling method resolves the limitations of traditional models in characterizing nonlinear features of flooding processes. The model employs a dual-stream architecture to extract four-dimensional key information from flooding processes: spatiotemporal dynamic features, motion and appearance features, and optical flow features, constructing a multi-source data fusion framework. Cross-attention mechanisms enable efficient fusion of multidimensional features, with experimental comparisons of different fusion strategies. Results demonstrate that the model achieves an average accuracy of 93.8% in flooding time prediction, outperforming current models such as TimeSformer and ViViT by 1.4% in multi-cabin comprehensive prediction capability. The model enables balanced prediction of flooding times across all cabins to support decision-making by authorities. This study provides an extensible technical paradigm for intelligent ship emergency systems.