Cloud-Native Geospatial for Country-Scale Ecosystem Risk Assessment

CNG Forum 2026 · Cloud-Native Geo in Practice

Tyler Erickson

Tyler Erickson

VorGeo · Founder

Radiant Earth · CTO

linkedin.com/in/tylere

github.com/tylere

analyze.earth

“Ecosystems are complexes of organisms and their associated physical environment within a specified area (Tansley, 1935). They have four essential elements: a biotic complex, an abiotic environment, the interactions within and between them, and a physical space in which these operate (Pickett & Cadenasso, 1995).”

To assess ecosystem risk we need to measure what’s happening on the ground. Field ecologists collect data: species counts, vegetation surveys, photographs, and physical samples — combined with satellite imagery and remote sensing to scale up to the country level.

Camera traps are triggered by motion and record wildlife continuously — including nocturnal behaviour invisible to daytime field surveys.

Wildlife Insights is an AI-powered platform that helps researchers process and share camera trap data at scale — using machine learning to identify species in images, reducing the manual work from months to hours.

How at risk

are a country’s

ecosystems?

IUCN Red List of Ecosystems

The global standard for ecosystem risk

The ecological counterpart to the Red List of Threatened Species. Global Ecosystem Typology gives the shared classification.

Countries have done this. Colombia 2017, Myanmar 2020.

A patchwork of tools and ad hoc scripts

hard to share · reproduce · maintain

Integrate national-scale spatial data, compute metrics, produce a formal report meeting IUCN criteria.

460,350ecosystem polygons

87ecosystem types

1.7 GBas GeoParquet

IDEAM · 1:100,000 · 2024

An open-source,

cloud-native

assessment workflow

Document as code

This page is a notebook

code · narrative · maps

one repo, version-controlled, reproducible

Live site: https://tylere.github.io/rle-tyler-colombia/ If the network fails, the next slides are screenshots.

Backup

Backup

Backup

[Screenshot: interactive Lonboard map from the site]

The hard part wasn’t the geospatial computation.

It was setup.

Assessors are ecologists, not software engineers

Python environment GitHub account Google Cloud project gcloud auth login application-default login MFA IAM roles service account Earth Engine registration billing bucket CORS GitHub secrets Pages settings

Setup friction as a first-class problem

One command

uv run …/init_repo.py \
  --country-name Colombia
pixi shell

Fork a template

Config, not code

ecosystem_source:
  data: …/colombia.parquet
  ecosystem_name_column:
    ecos_general

github.com/rle-assessment

Where the tooling still falls short

• Source data behind logins and manual downloads
• Cloud auth and permissions are still the wall
• Format details matter: GeoParquet row groups
• CI builds that need retries

Cloud-native formats make country-scale maps static

Document-as-code keeps the science reproducible

For non-engineers, setup is the product

Thank you

iucnrle.org

github.com/rle-assessment

tylere.github.io/rle-tyler-colombia