Detecting Deforestation with AI

The Scale of Deforestation

The world loses approximately 10 million hectares of forest every year – an area roughly the size of South Korea. Tropical deforestation accounts for the majority of this loss, driven by agricultural expansion, cattle ranching, logging, and infrastructure development. The Amazon, Congo Basin, and Southeast Asian rainforests experience the most severe pressure, with cascading effects on biodiversity, carbon emissions, and indigenous communities.

Monitoring deforestation at this scale is impossible with ground surveys alone. A single ranger station might oversee hundreds of thousands of hectares of forest, much of it accessible only by river or on foot. By the time illegal clearing is discovered through patrols, the forest is already gone. Satellite monitoring changes this equation fundamentally – every hectare of forest on Earth can be observed from space every few days, and AI can process these observations automatically to detect clearing within days of it occurring.

How Satellite Monitoring Works

Satellite-based deforestation monitoring relies on repeated observations of the same location over time. Each satellite image captures a snapshot of the forest canopy. By comparing images from different dates, analysts (and algorithms) can identify where forest has disappeared. The fundamental signal is simple: forest reflects strongly in the near-infrared spectrum, while bare soil and grassland do not. When a pixel transitions from high near-infrared reflectance to low, something has changed on the ground.

In practice, this simple concept encounters several complications:

• Cloud cover: Tropical forests are cloudy for much of the year. A single Sentinel-2 image may have 50-80% cloud coverage, hiding the very deforestation you are trying to detect. Temporal compositing (combining multiple images over weeks) and SAR data (which sees through clouds) help mitigate this problem.
• Seasonal variation: Deciduous forests naturally lose leaves seasonally, mimicking the spectral signature of deforestation. Models must distinguish seasonal phenology from actual clearing, typically by building a baseline of expected seasonal behavior for each pixel.
• Degradation vs. clearing: Selective logging, fire damage, and canopy thinning are harder to detect than complete clearing because the spectral change is subtler. Radar data is more sensitive to structural changes in the canopy than optical data.
• False positives: Cloud shadows, smoke from nearby fires, and sensor artifacts can be misidentified as forest loss. Machine learning models trained on confirmed deforestation events learn to filter these false signals.

AI Detection Methods

Change Detection

The most straightforward approach compares two images from different dates and flags pixels where vegetation indices have dropped below a threshold. NDVI differencing (subtracting the earlier NDVI from the later NDVI) highlights areas of vegetation loss. AI improves on simple thresholding by learning complex decision boundaries that account for seasonal patterns, sensor noise, and local ecosystem characteristics. Siamese neural networks, which process two images through identical network branches and compare the resulting features, have become the standard architecture for deep learning-based change detection.

Time Series Analysis

Rather than comparing just two dates, time series approaches analyze the full temporal trajectory of each pixel. Algorithms like BFAST (Breaks For Additive Season and Trend) decompose the signal into seasonal, trend, and residual components, then detect abrupt breaks in the trend that indicate deforestation. This approach is more robust to seasonal variation and gradual degradation than two-date comparison. Deep learning variants use recurrent neural networks (LSTMs) or temporal convolution networks to learn temporal patterns from labeled deforestation time series.

Spatial Pattern Recognition

Deforestation often follows recognizable spatial patterns: fishbone patterns along roads in the Amazon, concentric clearing around settlements in the Congo Basin, or systematic plantation establishment in Borneo. Convolutional neural networks learn these spatial patterns from labeled examples, enabling them to detect deforestation even before the spectral change is fully visible. This spatial context helps distinguish planned agricultural expansion from natural disturbances like windthrow or flooding.

Existing Monitoring Systems

GLAD Alerts

The Global Land Analysis and Discovery lab at the University of Maryland produces weekly deforestation alerts from Landsat imagery. GLAD alerts have been the backbone of tropical forest monitoring since 2016, providing publicly accessible, near-real-time alerts at 30-meter resolution. The system processes every new Landsat scene automatically, applying a trained classifier to detect tree cover loss and issuing alerts through the Global Forest Watch platform.

Global Forest Watch

Global Forest Watch (GFW) is a platform built by the World Resources Institute that aggregates multiple deforestation monitoring systems into a single interactive map. It combines GLAD alerts, Hansen tree cover loss data, fire alerts (VIIRS and MODIS), and integrated deforestation alerts (combining optical and radar data) into a comprehensive monitoring dashboard. GFW provides free data access, email alerts for custom areas of interest, and an API for programmatic access.

Deforestation Analysis via API

SciRouter's geospatial endpoints enable programmatic deforestation monitoring. Here is how to check for forest loss in a specific region:

import requests

API_KEY = "sk-sci-your-api-key"
BASE = "https://api.scirouter.ai/v1"
HEADERS = {"Authorization": f"Bearer {API_KEY}"}

# Monitor a region in the Amazon for deforestation
response = requests.post(f"{BASE}/geospatial/change-detect",
    headers=HEADERS,
    json={
        "latitude": -3.47,
        "longitude": -62.22,
        "radius_km": 25,
        "source": "sentinel-2",
        "date_before": "2025-06-01",
        "date_after": "2026-03-01",
        "change_type": "deforestation",
        "min_area_ha": 1.0
    })

result = response.json()
print(f"Region analyzed: {result['area_km2']:.0f} km2")
print(f"Forest cover (before): {result['forest_pct_before']:.1f}%")
print(f"Forest cover (after):  {result['forest_pct_after']:.1f}%")
print(f"Forest loss detected:  {result['loss_ha']:.1f} ha")
print(f"Loss hotspots: {len(result['hotspots'])}")
for h in result["hotspots"][:5]:
    print(f"  ({h['lat']:.4f}, {h['lon']:.4f}) - {h['area_ha']:.1f} ha")

Region analyzed: 1963 km2
Forest cover (before): 87.3%
Forest cover (after):  84.1%
Forest loss detected:  628.4 ha
Loss hotspots: 23
  (-3.4212, -62.1847) - 142.3 ha
  (-3.5103, -62.3021) - 87.6 ha
  (-3.3876, -62.1543) - 64.2 ha
  (-3.4921, -62.2765) - 51.8 ha
  (-3.4456, -62.1932) - 43.1 ha

Building a Deforestation Alert Pipeline

For continuous monitoring, you can set up a pipeline that runs periodically and sends alerts when new deforestation is detected:

import requests
from datetime import datetime, timedelta

API_KEY = "sk-sci-your-api-key"
BASE = "https://api.scirouter.ai/v1"
HEADERS = {"Authorization": f"Bearer {API_KEY}"}

# Define monitoring areas (protected forests, supply chain regions)
watch_areas = [
    {"name": "Xingu Indigenous Park", "lat": -11.5, "lon": -53.0, "radius": 50},
    {"name": "Tapajós National Forest", "lat": -4.0, "lon": -55.0, "radius": 30},
]

today = datetime.now()
two_weeks_ago = today - timedelta(days=14)

for area in watch_areas:
    response = requests.post(f"{BASE}/geospatial/change-detect",
        headers=HEADERS,
        json={
            "latitude": area["lat"],
            "longitude": area["lon"],
            "radius_km": area["radius"],
            "source": "sentinel-2",
            "date_before": two_weeks_ago.strftime("%Y-%m-%d"),
            "date_after": today.strftime("%Y-%m-%d"),
            "change_type": "deforestation",
            "min_area_ha": 0.5
        })
    result = response.json()
    if result["loss_ha"] > 0:
        print(f"ALERT: {area['name']} - {result['loss_ha']:.1f} ha lost")

From Detection to Action

Detecting deforestation is only valuable if it leads to intervention. The most effective monitoring systems connect satellite detections to enforcement agencies, supply chain compliance teams, and conservation organizations through automated alert pipelines. Here is how different stakeholders use deforestation intelligence:

• Law enforcement: Real-time alerts direct ranger patrols to active clearing sites, enabling intervention before large areas are lost.
• Supply chain compliance: Companies committed to zero-deforestation supply chains monitor their sourcing regions and flag suppliers operating in recently deforested areas.
• Carbon markets: Forest carbon credit projects use satellite monitoring to verify that protected forests remain intact, providing the measurement and verification layer that carbon credits require.
• Research: Scientists use long-term deforestation records to study the relationship between forest loss, climate change, biodiversity decline, and human development.

Next Steps

To explore satellite-based environmental monitoring further:

• Try the Satellite Analyzer Studio – see land cover segmentation on sample satellite scenes
• SAM for Satellite Images – use the Segment Anything Model for land cover segmentation
• How to Analyze Satellite Images with AI – comprehensive guide to satellite analysis methods
• Satellite Analyzer – free online tool for satellite image analysis

Every day that deforestation goes undetected is a day of irreversible loss. AI-powered satellite monitoring makes detection faster, cheaper, and more comprehensive than ever before. Get a free API key and start monitoring the forests that matter to you.