Applied AI for environmental defense

Making the destruction
of the Amazon visible,
measurable, and actionable.

We are a small research lab building computer-vision systems that monitor the Peruvian Amazon from orbit — detecting illegal gold mining, forest clearing, and road incursions within hours of when they happen.

SATELLITE × DEEP LEARNING × GIS · SINCE 2023
FIG. 01 — Live model output. Sentinel-2 imagery, model CV-DEFOR-v4.2, Madre de Dios region. Scanning west → east · cycle ≈ 24s
SENTINEL-2 L2A PLANETSCOPE 3M LANDSAT 8/9 SAR · SENTINEL-1 MAAP ALERTS GFW INTEGRATED MULTI-TEMPORAL DELTA U-NET · SEGFORMER · ViT POSTGIS / GEE SENTINEL-2 L2A PLANETSCOPE 3M LANDSAT 8/9 SAR · SENTINEL-1 MAAP ALERTS GFW INTEGRATED MULTI-TEMPORAL DELTA U-NET · SEGFORMER · ViT POSTGIS / GEE
§ 01 · Research

Four pillars at the intersection of computer vision and remote sensing.

01 / Vision

Segmenting the canopy at sub-hectare resolution

Custom U-Net and SegFormer architectures trained on labeled Sentinel-2 and PlanetScope tiles across the western Amazon. Specialized heads for mine pits, tailings, burn scars, and road scars.

02 / Change

Multi-temporal change detection

Pixel-paired models that compare time-aligned tiles week over week, controlling for cloud cover, seasonality, and atmospheric noise — surfacing only the changes that signal human activity.

03 / Geo

Geospatial intelligence

Detections are joined against concession boundaries, protected areas, Indigenous territories, and known infrastructure. Each alert carries the context an analyst needs to decide what to do next.

04 / Signal

Cloud-piercing SAR fusion

Sentinel-1 synthetic-aperture radar feeds into the same pipeline so detection continues through the rainy season, when optical satellites see only cloud. Loss happens year-round; so does monitoring.

§ 02 · Focus region

Madre de Dios — the front line of illegal gold mining in Peru.

PE-MDD · 85,300 km² monitored
Coordinates · 12.59° S, 69.19° W

A region the size of Portugal, where mining pits expand by the day and most of the loss happens far from any road.

Artisanal and informal gold mining has stripped tens of thousands of hectares from the Madre de Dios floodplain — leaving behind moonscapes of toxic ponds and mercury-contaminated rivers. The activity is fast-moving, remote, and almost invisible from the ground.

That's exactly what satellites and computer vision are good at. Our models tile the region every week, surface what's changed, and route alerts to the conservation NGOs and enforcement agencies who can act on them.

Forest lost to mining (2009–2024)
~95,750ha
Source: MAAP / project archive
Median detect-to-alert
6.4hrs
From Sentinel pass to partner inbox
Active monitored area
85.3kkm²
Across MdD, Loreto buffer, Ucayali
Alerts delivered · 2025
14,210
To 11 partner organizations
§ 03 · Methods

Open methods, replicable pipelines. Every model is benchmarked against ground truth.

M / 01

Tile, normalize, mask.

We ingest Sentinel-2 L2A tiles across the Peruvian Amazon, apply per-band atmospheric normalization, and mask clouds and shadows with a learned classifier (FMask + custom head). Tiles that fail QA are flagged for SAR fallback.

Sentinel-2 10 m/px L2A surface refl. FMask 4.6
M / 02

Segment land-cover and disturbance.

A SegFormer-B3 backbone, finetuned on ~42,000 labeled tiles across MdD, Ucayali, and Loreto, produces per-pixel disturbance masks for five classes: intact forest, secondary growth, recent clearing, mine pit, tailings/pond.

SegFormer-B3 5 classes 42k labels mIoU 0.81
M / 03

Compare across time.

Paired tiles from consecutive Sentinel passes feed a Siamese change-detection model. Differences below threshold are discarded; the rest become candidate detections with provenance, confidence, and bounding geometry.

Siamese U-Net 5–10 day Δ cloud-aware precision 0.93
M / 04

Join geospatial context.

Each candidate is intersected with concession polygons, protected-area boundaries, Indigenous territories, and prior alerts in PostGIS. Detections inside protected or unlicensed terrain are escalated to priority.

PostGIS SERNANP polygons ANP / RC / RN priority routing
M / 05

Route to humans.

Alerts are pushed to partner dashboards and email digests with the imagery, geometry, confidence, and a one-click link to a verification tile. Analysts confirm or dismiss; their labels flow back into the next training run.

REST + GeoJSON verification loop human-in-loop
§ 04 · System

From raw satellite pass to actionable alert in under seven hours.

01 · Ingest

Satellite tiles

Sentinel-2, Sentinel-1 SAR, and PlanetScope tiles are pulled within minutes of acquisition.

02 · Process

Normalize + mask

Atmospheric correction, cloud masking, geometric alignment to a common reference grid.

03 · Infer

CV models run

Segmentation + paired change detection produce candidate polygons with confidence scores.

04 · Contextualize

Geospatial join

Each polygon is enriched with protected-area, concession, and Indigenous-territory context.

05 · Alert

Partners notified

Verified, prioritized alerts are delivered to NGOs, researchers, and enforcement agencies.

§ 05 · Partners

Working alongside the people who can actually do something with the data.

Conservation

Field NGOs receive prioritized alerts for verification and field response.

Including organizations focused on Indigenous land defense, primary-forest monitoring, and ecosystem restoration in the western Amazon.

Government

Enforcement agencies use detections to direct patrols and seizures.

Including the Peruvian protected-area authority and prosecutors investigating environmental crime under specialized units.

Research

Universities and research groups co-author papers and share labeled data.

All datasets the lab produces are made available to academic collaborators under non-commercial research licenses.

Collaborate with the lab

If you work on forest defense, environmental crime, or earth-observation AI — we want to hear from you.

We're a small team. We collaborate with NGOs, government agencies, journalists, and academic groups. We also publish tooling, datasets, and pretrained weights for non-commercial research.

hello@canopyaivisionlab.org → Read the latest report Open-source repository