SCIENCE · METHODOLOGY

Forest Integrity Score

A composite 0-100 score integrating 6 satellite sources, a Random Forest ML classifier trained on ESA WorldCover + Hansen GFW, and Artificial Intelligence analysis to quantify forest zone integrity.

EXAMPLE — AMAZON RAINFOREST · OCTOBER 2026

Base satellite score: 80/100 · RF deforestation penalty: -3 pts (4.8% loss 2021-2026) · Final score: 77/100 Good Forest

1. Calculation formula

# Base formula (before RF correction)
Score = pts_NDVI + pts_Trend + pts_Cover + pts_Radar + pts_Fire
Score_max = 25 + 25 + 25 + 15 + 10 = 100 points

# Random Forest correction (deforestation penalty)
if RF deforestation > 15%: Score -= 20
if RF deforestation > 10%: Score -= 15
if RF deforestation > 5%: Score -= 8
if RF deforestation > 2%: Score -= 3
if RF degradation > 20%: Score -= 10
if RF degradation > 10%: Score -= 6
if RF degradation > 5%: Score -= 3

Final Score = max(0, Score - RF_penalty)

2. Score components

Current vegetation (NDVI)

25/25 pts

NDVI from Sentinel-2 (10m) — measures current vegetation density and health.

NDVI ≥ 0.70 → 25 pts (Excellent)
NDVI ≥ 0.50 → 18 pts (Good)
NDVI ≥ 0.30 → 11 pts (Moderate)
NDVI ≥ 0.10 → 5 pts (Low)
NDVI < 0.10 → 0 pts

NDVI Trend (10 years)

25/25 pts

Comparison of current Sentinel-2 NDVI vs Landsat 8/9 from 10 years ago. Indicates whether the forest is improving or degrading.

Δ ≥ +0.10 → 25 pts (Strong improvement)
Δ ≥ +0.05 → 20 pts
Δ ≥ 0 → 15 pts (Stable)
Δ ≥ -0.05 → 8 pts
Δ ≥ -0.10 → 4 pts
Δ < -0.10 → 0 pts (Strong degradation)

Forest cover

25/25 pts

Percentage of vegetated pixels (NDVI > 0.3) in the analyzed zone via Sentinel-2 10m.

Cover ≥ 80% → 25 pts
Cover ≥ 60% → 18 pts
Cover ≥ 40% → 11 pts
Cover ≥ 20% → 5 pts
Cover < 20% → 0 pts

Sentinel-1 Radar (VV)

15/15 pts

ESA Sentinel-1 VV radar backscatter (10m). Measures forest structure independently of clouds and nighttime.

VV ≥ -8 dB → 15 pts (Dense forest)
VV ≥ -12 dB → 10 pts
VV ≥ -16 dB → 5 pts
VV < -16 dB → 0 pts

Fire activity (FIRMS)

10/10 pts

NASA FIRMS/VIIRS active fire pixels (375m) over 12 months. Fewer fires = higher score.

0 pixels → 10 pts (No fire)
≤ 5 pixels → 5 pts
> 5 pixels → 0 pts

3. Random Forest ML correction

Random Forest classifier (50 trees)

A Random Forest classifier trained on ESA WorldCover 2021 and Hansen GFW 2025 classifies each pixel in the zone into 5 categories. Results are used to correct the base score.

Training data:
- ESA WorldCover 2021 (10m) → land cover labels
- Hansen GFW 2025 (30m) → deforestation 2021-2026
- ESA Sentinel-2 (10m) → features: B2,B3,B4,B8,B11,B12,NDVI,EVI,NBR,NDWI

Output classes:
1 → Healthy forest
2 → Degraded
3 → Non-forest
4 → Water
5 → Deforested (2021-2026)

4. Artificial Intelligence Analysis

Artificial Intelligence (AI)

The Artificial Intelligence receives all satellite data and RF results to generate an intelligent analysis of forest type, risks and conservation recommendations.

AI inputs:
- Spectral indices: NDVI, EVI, NBR, NDWI
- RF results: % healthy forest, degraded, deforested
- Geographic location
- Dynamic deforestation period

AI outputs:
- Forest type (Dense Tropical/Temperate/Boreal/...)
- Leaf type (Broadleaf/Coniferous/Mixed)
- Development stage
- Deforestation risk (Low/Moderate/High/Critical)
- Degradation and recovery signs
- Main cause and recommendation

5. Score interpretation

80–100

HEALTHY FOREST

Dense stable vegetation, few fires, no RF deforestation detected

60–79

GOOD FOREST

Good cover, positive trend, moderate risk

40–59

DEGRADED FOREST

Degradation detected, negative trend or significant fires

20–39

CRITICAL FOREST

Active deforestation, strong RF degradation detected

0–19

NON-FOREST

No detectable forest cover

6. Data sources

Source	Resolution	Usage	Component
ESA Sentinel-2	10m	NDVI, EVI, NBR, NDWI, cover	NDVI, Trend, Cover, RF
NASA Landsat 8/9	30m	10-year trend	Trend
NASA Landsat 4/5	30m	40-year trend	Historical
ESA Sentinel-1	10m	VV radar — forest structure	Radar
NASA FIRMS/VIIRS	375m	Active fires	Fire
NASA GEDI LiDAR	25m→10m	Canopy height	Biomass
ESA WorldCover 2021	10m	RF labels	RF
Hansen GFW 2025	30m	RF deforestation	RF
Artificial Intelligence (AI)	—	Intelligent analysis	AI
GBIF / WorldClim / CMIP6	—	Species suitability	BEA

7. Limitations

Important points

The score represents an analysis zone (1-100km) — not the entire forest
GEDI LiDAR limited above 51.6° latitude
RF classifier trained on 100 points/class — larger training would improve accuracy
AI analyzes spectral data — not direct images
RF deforestation dynamically covers the last 5 years (2021-2026) — updated automatically each year

Analyze a forest → Publications →