Information-Aware Reduction of Construction LiDAR Point Clouds: Scene-Level Fidelity, Minority-Class and Registration Trade-offs

Abstract Dense construction point clouds are often reduced before storage, transfer or analysis, but global distortion metrics can conceal loss of sparsely represented classes and instances. This study evaluates information-aware point-cloud reduction using 51 terrestrial laser-scan scenes from two Rohbau3D construction sites. Every result is conditional on a class-aware 150,000-point reference sample per scene (7.65 million sampled points in total), rather than the native full-resolution scans. We compare a naïve NPZ/Deflate storage baseline, voxel grids of 0.05, 0.10 and 0.20 m, and five archived fixed-rate selection policies at 25%, 10% and 5% retention. Scene-level bootstrap confidence intervals and repeated-measures tests replace point-pooled inference. The naïve storage baseline reduced sampled payload size by 10.13%. A 0.05 m voxel grid removed 69.83% of sampled points while producing 1.470 cm mean nearest-neighbour error and 99.32% nearest-neighbour class agreement. All voxel outcomes changed monotonically across settings. In controlled ICP, 0.05 m voxelisation did not significantly change the archived control-RMSE measure relative to the sampled-reference condition, but fitness within 10 cm fell by 13.62 percentage points ( $$p<.001$$ p < . 001 ). At 0.20 m, 83.77% of rare class–scene records retained less than 10% of their sampled points, although no evaluated class disappeared completely. Balanced information-aware selection improved small-instance presence over class-aware selection by 2.10, 7.21 and 15.47 percentage points at 25%, 10% and 5% retention, respectively, while sharing identical class quotas and therefore identical rare-class retention. These measures indicate how much information is retained after point-cloud reduction, on top of segmentation or object-detection performance. The study therefore provides a construction-specific, scene-level benchmark and a transparent basis for selecting appropriate reduction settings.