Automated 3D Terrain and Building Mapping from LiDAR Scans

LiDAR aerial survey produces a 3D point set: 50–200 points per square meter, each with XYZ coordinates and return intensity. From this data we need to extract structured information—buildings, trees, roads, terrain—and build an accurate 3D model of the territory. We solve this problem with AI: we automate classification, segmentation, and reconstruction, reducing processing time by 5–10 times compared to manual methods.

Problems We Solve

Typical issues: noise (birds, scanner artifacts), overlaps (power lines, bridges), complex terrain with sharp elevation changes. Our pipelines handle these through self-attention mechanisms and data augmentation. Parameters—point density, LAS/LAZ format, coordinate system—we adapt without loss of accuracy, which is especially important for large territories.

How We Do It

Technology: pre-trained models RandLA‑Net and PointNet++ fine-tuned to your landscape, achieving mIoU = 0.927 on reference benchmarks. The result—a classified point set and 3D model in CityGML with LoD1–LoD3 detail level—in 1–3 days for a pilot area.

Stages of LiDAR Data Processing

1. Normalization and noise filtering—outlier removal (birds, artifacts), height normalization relative to DTM. Standard tools: Open3D, PDAL. AI at this stage: a neural network outlier detector outperforms statistical methods (LOF, SOR) on complex terrain.
2. Point classification—each point assigned a class: ground, vegetation, building, water, power lines, noise. Classic algorithm: LASTools, CloudCompare. AI approach: PointNet++, RandLA-Net, KPConv.
3. Object segmentation—extracting individual buildings, trees, poles from same-class point clouds.
4. Surface reconstruction—from segmented points we build a mesh: LoD1 (box model), LoD2 (with roof), LoD3 (detailed facade).

How Point Classification with Neural Networks Works

RandLA‑Net—current state-of-the-art for large-scale outdoor scenes. Key: random point sampling + local feature aggregation (LFA). Works with 10M+ points directly without voxelization. As described in the original paper RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds, the model achieves mIoU of 0.927 on Semantic3D.

import torch
from randlanet import RandLANet

model = RandLANet(
    num_classes=8,   # ground, veg_low, veg_med, veg_high, building,
                     # water, powerline, noise
    decimation=4,
    num_neighbors=16
).cuda()

# Input: (B, N, 3) - xyz coordinates
# Optional: (B, N, F) - additional features (intensity, RGB)
point_coords = torch.rand(2, 100000, 3).cuda()
features = torch.rand(2, 100000, 1).cuda()  # intensity
logits = model(point_coords, features)  # (B, N, num_classes)

On Semantic3D benchmark mIoU = 0.927. On real aerial LiDAR data after fine-tuning—0.89 across 7 classes (ground, vegetation levels, building, infrastructure). We guarantee accuracy not lower than 85% on any type of landscape.

Comparison of Neural Network Architectures for Point Clouds

Automatic Building Extraction and LoD Construction

After classification: all points of class "building" are split into individual structures via DBSCAN clustering. Each structure is a separate point cluster.

• LoD1 (Block model): minimum oriented bounding box + height from roof points. Fully algorithmic, 100% automation.
• LoD2 (Roof model): roof shape detection (flat, gabled, hipped, mansard). PointNet++ classifier on normalized roof point cloud: 93% accuracy on 8 roof types. Geometric reconstruction of sloped surfaces via RANSAC plane fitting.
• LoD3 (Detailed facade): requires additional data (ground laser scans or photos). Only LoD3 shows windows, doors, architectural details.

Why LoD2 Matters for Urban Planning

LoD2 allows precise calculation of building volumes, insolation, and shadow zones. We automatically build LoD2 for an entire territory in 2–3 days per 100 km², which is 80% faster than manual modeling. Budget savings up to 40% due to elimination of manual labor. Processing of a 100 km² area starts at €8,000, with a pilot area of 1 km² from €1,000.

Digital Models of Terrain and Surface

• DSM (Digital Surface Model): height of first return—everything including vegetation and buildings.
• DTM (Digital Terrain Model): only ground, after object filtering.
• Difference DSM - DTM = nDSM (normalized Digital Surface Model)—"height above ground." This is the basis for calculating building volumes, tree heights, biomass.

Classic algorithm for DTM: Progressive Densification (Axelsson). Neural network filter CLOTH_NET: better on complex terrain (bridges, slopes, dense vegetation).

Formats and Integration

• LAS/LAZ—standard point cloud format
• CityGML—standard for 3D city models (LoD1–LoD3)
• 3D Tiles (Cesium)—for web visualization of large territories
• IFC—for BIM integration (construction)

What's Included in the Work

• Analysis of source LiDAR data and selection of model architecture
• Pipeline deployment on GPU server or cloud (AWS, GCP)
• Fine-tuning the model on your area with delivery of a model card
• Classification, segmentation, and LoD reconstruction
• Documentation in PDF format (accuracy metrics, pipeline description)
• Engineer consultations at all stages
• 3 months of support after project delivery

Get an engineer consultation: we will analyze your data and propose an optimal pipeline. Order a pilot project on an area from 1 km²—we set up the pipeline in 3–5 weeks.

Timelines

The cost is calculated based on area and required level of detail (LoD). Our experience—10+ years in AI/ML, 50+ projects in LiDAR processing. Contact us to evaluate your project.