3D Digital Twin / Geospatial 3D Visualization Engineer

Profile browser memory first, then inspect geometry construction and React state usage for duplicated arrays. Apply one controlled change at a time such as indexed BufferGeometry, typed arrays, terrain chunking, LOD, or progressive loading rather than reducing the source dataset.

Reproject Sentinel-2 bands from EPSG:32760 to EPSG:2193, then run the existing fuse_sentinel_buildings.py pipeline. The fused building_fused.vtp output attaches BuildingID, GroundElevation, Height, RoofElevation, RGB, and NDVI attributes to LiDAR-derived geometry.

No. The processed Sentinel-2 outputs and the 2 m terrain mesh are validated baselines that must not be regenerated to solve rendering problems. Frontend visualization optimization should always be attempted first.

The terrain contains 346,801 vertices and 691,200 triangles, and freezing usually comes from storing large arrays in React state, duplicating geometry buffers, or recreating BufferGeometry every render. Use typed arrays, memoized geometry, and proper disposal instead.

The dataset uses EPSG:2193 (NZTM2000) with the NZVD2016 vertical datum. Sentinel-2 source data arrives in EPSG:32760 and must be reprojected before any spatial fusion with LiDAR data.

Never commit myvenv/, ml/dataset/ (approximately 27 GB), .env files, or node_modules/. Large geospatial files such as LAZ, TIF, and VTP are tracked through Git LFS and must not be converted into normal Git blobs.