Abstract: High-precision 3D models serve as the core carriers for the digitalization and informatization of water conservancy projects. As a high-fidelity modeling approach, digital twin modeling enables dynamic and precise mapping from physical entities to virtual models. However, existing modeling technologies face dual challenges in complex water conservancy scenarios: single-modal sensors suffer from blind spots and occlusions in complex construction environments, while multimodal sensing data can compensate for these inherent defects but often leads to poor registration and fusion performance due to large density differences and low overlap rates. To address these issues, this paper proposes a method for modeling digital twin underground cavern groups under image and point cloud multimodal sensing. It utilizes 3D image reconstruction and 3D laser scanning to acquire cross-source point cloud data of hydraulic underground cavern groups, compensating for the limitation of single-modal sensors in fully perceiving the spatial information of complex cavern groups. A Dual-path Feature Fusion Attention (DPFFA) module is designed and embedded into the Geotransformer backbone network to enhance the network’s ability to fuse multi-scale features of global structures and local details, overcoming the failure of single-path attention caused by cross-source point cloud density differences and missing data. A complete high-precision digital twin model is then established based on the accurately registered cross-source point clouds. This method was applied to the underground cavern group of the Kala Hydropower Station and tested on point cloud datasets with overlap rates of 20%, 40%, 60%, and 80%. The registration recall (RR) reached 26.90%, 72.50%, 84.40%, and 86.90%, respectively, representing improvements of 3.20%, 11.20%, 8.20%, and 2.50% compared to the unimproved model. Method comparisons demonstrate the superiority of the proposed registration approach, validate the accuracy of the established digital twin underground cavern group model, and provide new effective technical support for digital twin modeling.