It remains largely unknown how seafloor surface stretches in extremely deep sea around Mariana region. In this paper, we make an attempt to establish online computing strategies for high-resolution multi-beam bathymetry survey to characterize seafloor stretching morphology via GhostNet, combined with topological attributes, which hopefully would help Remotely Operated Vehicles (ROV), Autonomous Underwater Vehicles (AUV), to develop embedded smart capacities. It has been demonstrated from the experiment results that the seafloor stretching classification accuracy of our proposed scheme could reach 87.3% averagely, GPU processing time 56 ms, achieving comparable performances with the state-of-art deep learning frameworks, including MobileNet, Mnasnet, SEnet, which permits us to delicately and adaptively distinguish the specific seafloor stretching morphology categories and topographic mapping.