Field work · ArborMeta · 2026

Canopy cover from orbit

LiDAR measures structure honestly, but only over flightlines. Satellites cover whole stations but cannot see in 3D. This learns the mapping: canopy-cover labels from LiDAR, self-calibrated per station, then run across the property from satellite alone. Identifiers stripped; numbers from a real production run.

Ground resolution

0.5 m

Cover layers predicted

11 · 1.0–2.0 m

Self-calibration val loss

0.0133 → 0.0096

One station, predicted

975 tiles · 7 m 24 s

The pipeline

Two instruments, one model. Labels from LiDAR; imagery from satellite; they meet in a wavelength-conditioned encoder.

The two instruments

Labels: airborne LiDAR → smoothed canopy-height model → eleven cover-fraction rasters (vegetation taller than 1.0–2.0 m). Imagery: Jilin-1 PAN + four multispectral bands at 0.5 m. Nothing enters as raw pixels — per-band gain and bias from scene metadata, each band with wavelength, bandwidth, sun and satellite geometry.

A wavelength-aware model

Swin-UNet with DOFA-style dynamic patch embedding: a hypernetwork generates input-layer weights from band wavelengths, so the network is conditioned on what its channels physically are. Angles, location, and capture-to-prediction date delta enter as learned embeddings. Same weights, different constellations, different band counts.

Self-calibration, per station

Before predicting a station, the base model fine-tunes on that station's own LiDAR chips. Every dot is a saved checkpoint from one production run. Hover for the numbers.

The last mile

768-pixel tiles, 128 px overlap, mixed precision, georeferenced outputs. Thousands of tiles merge through GDAL VRT into one ZSTD-compressed station raster. One delivery: 18 rasters, 119.6 GB.

One run, end to end

The production log, trimmed to its skeleton.

The platform around it

The pipeline feeds a geospatial platform: single-handed build used by the team, government stakeholders and visiting researchers. FastAPI + PostGIS (~35 route modules, ~40 tables) under React 18 + TypeScript (14 pages), serving vector tiles and cloud-optimised point clouds and rasters from terabyte-scale archives.

The map

MapLibre workspace over every domain (ALS, TLS, satellite, carbon-project boundaries) as filterable vector-tile layers with AOI drawing, zonal statistics and QGIS export.

The 3D globe

A token-free Three.js globe that streams terrestrial-LiDAR point clouds of individual trees in one WebGL context: eye-dome lighting, classification filters, rescan-vs-parent growth comparison.

Per-tree science

DBH, height, volume, biomass and QSM model uncertainty for every reconstructed stem; colour any plot by any metric.

Pipeline operations

Ingestion timeline, processing queue, throughput and backlog, storage health, data-quality checks, and an admin audit log.

Collaboration

Typed, shareable workspaces, threaded comments, snapshot messaging, and real-time notifications over Postgres LISTEN/NOTIFY → Server-Sent Events.

FieldLink

An air-gap-friendly delivery channel for field crews: X25519 + Ed25519, sign-then-encrypt, replay protection, per-device sealed packages. The transport only ever sees ciphertext.

Why it matters

These rasters sit under Australian carbon-credit policy advice: canopy growth measured between repeat LiDAR captures, extended across whole stations by satellite. Same discipline as the demos here, pointed at country.

The ArborMeta case study →The pretraining report →Harvey Houlahan · ArborMeta · 2026