LiDAR & Photogrammetry
From 2017 to 2019 we set out to map a tropical forest from the air, combining drones, LiDAR, photogrammetry and deep learning to count and classify trees and estimate the carbon the forest stores. Begun under our founders' earlier company and carried into Tecnologías VM, it ran with the Tecnológico de Costa Rica (TEC), the country's public institute of technology, as a staged series of sponsored Computer Engineering graduation projects.
From drone to point cloud.
The survey drone itself, the LiDAR point clouds and orthomosaics it captured, and the deep-learning detection that counted and classified individual trees, all from the 2017–2019 R&D.
Costa Rica had set a national carbon-neutrality goal, but certifying forest carbon still meant inventorying biomass by hand: 12 to 24 months of plot-by-plot fieldwork, often dangerous or simply unreachable. We saw the opening. Combine drones, LiDAR, photogrammetry and deep learning, and you could estimate forest carbon faster and at far lower cost, letting Costa Rican landowners take part in the carbon market.
The pipeline was designed to run end to end. A drone overflies the forest carrying LiDAR and cameras under centimeter-level RTK positioning; photogrammetry and point-cloud processing build terrain and canopy-height models; deep learning detects individual trees, their species and dimensions; and an allometric model turns those variables into biomass and CO₂.
We ran it as sponsored Computer Engineering graduation projects at TEC, working hands-on inside its UAS Photogrammetry Lab, the university's drone and aerial-imaging research lab in the School of Electronic Engineering. Six projects across three cohorts, each owning one slice, with knowledge passed from one to the next and regular working sessions alongside the lab's researchers. The company financed and supervised the work, and kept the IP.
It is our earliest sustained program in applied deep learning and computer vision, proof that we were working with CNNs, remote sensing and big data years before “AI” became the label. It also set a pattern we still follow: sponsor real R&D, validate honestly, and carry the patterns forward.
From flight to carbon estimate.
- 01
Fly
A UAV overflies the forest carrying LiDAR and cameras, under centimeter-level RTK positioning.
- 02
Capture
Velodyne VLP-16 point clouds logged in the field: passes of up to 16.4M points.
- 03
Model
Photogrammetry (SfM) and point-cloud processing build terrain and surface models, subtracted into a canopy-height model.
- 04
Detect
Deep learning finds and classifies individual trees (species, height, crown) at up to ~95% accuracy.
- 05
Estimate
An allometric model turns per-tree variables into biomass, and biomass into CO₂.
What we built, and proved.
LiDAR capture & georeferencing
A Velodyne VLP-16 + RTK + Raspberry Pi capture chain that produced the first 3D point clouds from that sensor at the institution: passes of up to 16.4M points.
Photogrammetry to biomass
Drone photos to 3D canopy models (SfM), a canopy-height model, and >90% detection of separated crowns, feeding the intended allometric biomass estimate.
Tree-species classifier
A CNN (Inception V3, transfer learning) classifying tree species from photos at 94.4% accuracy, selected over VGG16 after a head-to-head.
Point-cloud tree classification
Algorithms that count and extract individual trees from georeferenced point clouds: about 95% average accuracy on the evaluated samples.
TEC's UAS Photogrammetry Lab
We ran this R&D as sponsored Computer Engineering graduation projects inside the lab, TEC's drone and aerial-imaging research group. A look at what the lab does:
This R&D validated its components but was never assembled into a single deployed end-to-end system, and no forest CO₂ figure was corroborated against ground truth. We present it as what it was: an early, ambitious effort that built real capability, and taught us how hard the last mile is.