Adaptive spatial partitioning under the GeoSOT-3D global grid for efficient low-altitude environment modelling

Abstract Allocating finite representational capacity across spatial domains where geometric complexity and semantic importance vary independently is a recurring problem in spatial modelling. We address this problem in the context of low-altitude airspace, where representational fidelity must trade off against computational efficiency. Building on the GeoSOT-3D global grid encoding system, we propose TSDP (Terrain-Semantic Dual-driven Partitioning), an adaptive partitioning framework that fuses terrain roughness and semantic importance into a multi-criteria weighted decision function. The current implementation operates as a horizontal-plane quadtree subdivision with elevation as a voxel attribute, complemented by the TS-AdaptiveSubdivide algorithm that formalises adaptive mesh generation as spatial-complexity minimisation under fidelity constraints. Experiments on a synthetic scene and a real-world Xi'an district show that TSDP achieves a favourable accuracy–efficiency trade-off: 60% voxel compression with semantic accuracy (F1 = 0.918) comparable to a uniform grid (F1 = 0.933) and competitive with single-driver baselines that excel on either accuracy or efficiency alone but not both. Path-planning time decreases by approximately 40% while mean safety distance to obstacles increases roughly threefold. More broadly, the results indicate that spatial discretisation efficiency in such environments is governed jointly by geometric and semantic factors rather than by either alone, supporting the design of digital-twin systems for low-altitude airspace and urban air mobility management.