ArcheoVLM

Phase 3: Scalable LiDAR Processing

Convert prioritized raw point clouds into analysis-ready visualizations using GKE Autopilot

Goal

Convert prioritized raw point clouds into analysis-ready visualizations in a scalable, robust, and archaeologically-sensitive manner using cloud infrastructure.

Processing Pipeline

Automated LiDAR processing workflow

1. Ground Classification

Point Transformer V3

Semantic segmentation of ground vs. non-ground points

2. DEM Generation

Adaptive Interpolation

1-meter resolution Digital Elevation Models

3. Visualization

RVT Toolbox

Archaeological visualization products

4. Output

Cloud-Optimized GeoTIFFs

Analysis-ready data products

Technical Components

Core processing algorithms and methods

Ground Point Classification

Point Transformer V3 (PTv3)

Fine-tuned model for semantic segmentation preserving low-relief archaeological features

Cloth Simulation Filter (CSF)

Robust fallback method for ground point identification

DEM Generation

Adaptive Interpolation

Applies optimal algorithms (Kriging, Natural Neighbor, IDW) based on local conditions

1-meter Resolution

High-resolution output preserving subtle archaeological features

RVT Visualization Products

Archaeological visualization techniques

Sky-View Factor

SVF

Reveals subtle topographic features

Local Relief Model

SLRM

Highlights local elevation variations

Openness

Positive/Negative

Emphasizes ridges and valleys

Hillshade

Multi-directional

Enhanced terrain visualization

Execution Checklist

Phase 3 tasks and deliverables

Set up GKE Autopilot cluster environment

Containerize the process_lidar_tile.py script and dependencies

Develop and fine-tune the Point Transformer V3 model

Develop and parameter-tune the Cloth Simulation Filter (CSF)

Implement the adaptive interpolation logic in the script

Implement the rvt-py visualization generation logic

Test the full process_lidar_tile.py job on a single high-potential tile

Execute the batch processing job on all prioritized tiles in GKE Autopilot

Verify that all output COGs are correctly generated

Expected Outputs

Classified Point Clouds

Ground/non-ground segmented LiDAR data

High-Resolution DEMs

1-meter resolution elevation models as COGs

RVT Visualizations

Archaeological visualization suite as COGs