Simplification of 3D building models is an important way to improve rendering efficiency. When existing algorithms are directly applied to simplify multi-component models, generally composed of independent components with strong topological dependence, each component is simplified independently. The consequent destruction of topological dependence can cause unreasonable separation of components and even result in inconsistent conclusions of spatial analysis among different levels of details (LODs). To solve these problems, a novel simplification method, which considers the topological dependence among components as constraints, is proposed. The vertices of building models are divided into boundary vertices, hole vertices, and other ordinary vertices. For the boundary vertex, the angle between the edge and component (E–C angle), denoting the degree of component separation, is introduced to derive an error metric to limit the collapse of the edge located at adjacent areas of neighboring components. An improvement to the quadratic error metric (QEM) algorithm was developed for the hole vertex to address the unexpected error caused by the QEM’s defect. A series of experiments confirmed that the proposed method could effectively maintain the overall appearance features of building models. Compared with the traditional method, the consistency of visibility analysis among different LODs is much better.

Neural Radiance Fields (NeRFs) are an emerging approach to 3D reconstruction that use neural networks to reconstruct scenes. However, its applications for multi-view satellite photogrammetry, which aim to reconstruct the Earth’s surface, struggle to acquire accurate digital surface models (DSMs). To address this issue, a novel framework, Geometric Constrained Neural Radiance Field (GC-NeRF) tailored for multi-view satellite photogrammetry, is proposed. GC-NeRF achieves higher DSM accuracy from multi-view satellite images. The key point of this approach is a geometric loss term, which constrains the scene geometry by making the scene surface thinner. The geometric loss term alongside z-axis scene stretching and multi-view DSM fusion strategies greatly improve the accuracy of generated DSMs. During training, bundle-adjustment-refined satellite camera models are used to cast rays through the scene. To avoid the additional input of altitude bounds described in previous works, the sparse point cloud resulting from the bundle adjustment is converted to an occupancy grid to guide the ray sampling. Experiments on WorldView-3 images indicate GC-NeRF’s superiority in accurate DSM generation from multi-view satellite images.

Complex 3D building models, because of their huge data volume, almost always result in transmission congestion, which leads to poor user experience. To reduce the real-time transmission pressure, a novel view-dependent progressive transmission method was developed. With this method, only a small amount of transmitted data is necessary to achieve an acceptable rendering effect when the viewpoint changes. The method involves two stages. A preprocessing stage simplifies the building model using a multi-level vertex clustering algorithm. The local mesh in each clustering unit is organized into a node tree where each node includes a vertex and its related triangles. The building model is finally reorganized into a node forest. In the reconstruction stage, all root nodes are transmitted first to build a basic model. Their descendant nodes are then requested and transmitted according to viewpoint information to refine the building model during user interaction. The experimental results show that this method can effectively improve the transmission and reconstruction efficiency of 3D building models.

Progressive encoding and transmission, i.e., a crucial technical foundation of 3D Web Geographic Information Systems (WebGIS), addresses the contradiction between massive 3D building data and limited network transmission capacity. Most progressive encoding algorithms, taking vertices, edges or triangles as encoding units, may break the inherent geometric and topological characteristics of 3D building models. Thus, a novel 3D building model encoding method that can maintain the internal characteristics is proposed, which can be used for high-efficiency progressive transmission. With this method, each building is decomposed into three types of fundamental structures: main structure, independent structure and attached structure. A structural topology graph (STG) was constructed based on the connections among structures. Guided by STG, one or more structures were wrapped as the smallest incremental transmission unit, denoted as transmission node. When requested, the real-time position of viewpoint, orientation and visual importance of nodes are used to pick up expected nodes for responding. The results confirm that the proposed method can better maintain the geometric and topological characteristics while encoding 3D building models. While serving for transmission, the proposed method not only effectively reduces the transmission load, but also provides users with a better consistency experience on the building appearance at different simplification levels.

In a point cloud semantic segmentation task, misclassification usually appears on the semantic boundary. A few studies have taken the boundary into consideration, but they relied on complex modules for explicit boundary prediction, which greatly increased model complexity. It is challenging to improve the segmentation accuracy of points on the boundary without dependence on additional modules. For every boundary point, this paper divides its neighboring points into different collections, and then measures its entanglement with each collection. A comparison of the measurement results before and after utilizing boundary information in the semantic segmentation network showed that the boundary could enhance the disentanglement between the boundary point and its neighboring points in inner areas, thereby greatly improving the overall accuracy. Therefore, to improve the semantic segmentation accuracy of boundary points, a Boundary–Inner Disentanglement Enhanced Learning (BIDEL) framework with no need for additional modules and learning parameters is proposed, which can maximize feature distinction between the boundary point and its neighboring points in inner areas through a newly defined boundary loss function. Experiments with two classic baselines across three challenging datasets demonstrate the benefits of BIDEL for the semantic boundary. As a general framework, BIDEL can be easily adopted in many existing semantic segmentation networks.