Abstract: Geographic information system(GIS) visualization analysis is an important method for disaster research, and mobile augmented reality (MAR) brings new vitality into GIS visualization.MAR visualization of disaster scenes usually needs to process multi-source and massive data, which challenges the performance of MAR device. To achieve smooth and stable loading capability for MAR visualization of disaster scenes, we proposed a flow-based optimization method for disaster scene loading. First, a MAR scene was split and compressed by the level of detail(LOD) method, then the scene data were stored on a data server. When a MAR client browsed data in a determined area, our optimized method can help to request the data from the server, and control the data loading by using our designed transmission strategy and alleviate the bandwidth pressure through our designed caching mechanism. A prototype system was built to test the performance of the optimized method. Experimental results show that the optimized method can achieve fast loading performance for MAR visualization of disaster scenes, and breakthrough the data volume limit on HoloLens device, and succeed stably under different data volume.