Python Deforestation Detection — Core Concepts

Why deforestation detection matters

Tropical forests store 250 billion tons of carbon and host 80% of terrestrial biodiversity. Deforestation contributes 10-15% of global greenhouse gas emissions — more than the entire transport sector. Near-real-time detection enables rapid response: Brazil’s DETER system reduced Amazon deforestation by 80% between 2004 and 2012 by giving enforcement agencies weekly alerts.

Satellite data sources

Deforestation detection relies on frequent, consistent satellite observations:

Landsat (NASA/USGS) — 30m resolution, 16-day revisit since 1972. The backbone of long-term forest change mapping. The Hansen Global Forest Change dataset uses Landsat to map annual tree cover loss worldwide from 2000 to present.

Sentinel-2 (ESA) — 10m resolution, 5-day revisit since 2015. Higher resolution enables detection of smaller clearing events and selective logging.

Sentinel-1 (ESA) — Synthetic Aperture Radar (SAR). Operates at 10m resolution and penetrates clouds — critical in tropical regions where cloud cover obscures optical satellites 60-80% of the time.

MODIS/VIIRS (NASA) — 250m-1km resolution, daily revisit. Coarse but fast. MODIS powers near-real-time fire and deforestation alerts.

Planet (commercial) — 3-5m resolution, daily revisit. Enables detection of small-scale and selective logging events invisible at Sentinel/Landsat resolution.

Detection approaches

Threshold-based change detection — Compare NDVI (vegetation greenness) between two dates. When NDVI drops below a threshold, flag as potential deforestation. Simple but prone to false positives from seasonal variation and cloud shadows.

Time-series analysis (BFAST, LandTrendr) — Fit a model to the historical NDVI time series for each pixel, capturing seasonal patterns and long-term trends. A break in the expected pattern indicates disturbance. More robust than two-date comparison but computationally intensive.

Machine learning classification — Train classifiers on examples of deforestation vs. non-deforestation change. Random forests and CNNs learn complex patterns combining spectral, textual, and temporal features.

SAR-based detection — Radar backscatter changes when forest structure changes (trees → bare ground). Works through clouds, enabling near-real-time monitoring in the tropics. Sentinel-1 C-band SAR is freely available and widely used.

The cloud masking challenge

In tropical forests, the biggest obstacle isn’t detecting deforestation — it’s seeing through clouds. The Amazon is cloud-free less than 30% of the time. Python handles this through:

• Cloud masks — Sentinel-2’s Scene Classification Layer (SCL) and the s2cloudless Python library flag cloudy pixels.
• Temporal compositing — Combine multiple cloud-free observations over a period (e.g., median composite over 3 months) to create a clean baseline.
• SAR fusion — Use radar data (cloud-independent) alongside optical data. When optical is cloudy, rely on SAR alone.

Key Python libraries

Library	Role
rasterio	Read and write satellite rasters (GeoTIFF, COG)
xarray	Multi-temporal image stacks and lazy computation
stackstac	Access STAC-cataloged satellite imagery as xarray
scikit-learn	Classification models for change detection
scipy.signal	Time-series breakpoint detection
geopandas	Vector operations on deforestation polygons
dask	Parallel processing for large satellite archives
leafmap / geemap	Interactive mapping and Google Earth Engine integration

Alert systems

Modern deforestation alert systems like GLAD (University of Maryland) and DETER (Brazil’s INPE) update weekly. They balance speed against accuracy:

• Speed: SAR-based systems can detect clearing within 1-6 days. Optical systems take 1-4 weeks depending on cloud conditions.
• Accuracy: Faster alerts have more false positives. A two-stage system issues preliminary alerts quickly and confirms them once clearer imagery is available.

Common misconception

“Deforestation is easy to detect from space because trees are green.” In reality, selective logging (removing individual high-value trees) barely changes the canopy from above. Degradation — where the forest thins without being completely cleared — is much harder to detect than complete clearing. These subtle changes can reduce a forest’s carbon stock by 30-50% while remaining nearly invisible to standard NDVI analysis. Detecting degradation requires texture analysis, canopy height models from LiDAR, or long-term trend analysis.

One thing to remember: Deforestation detection in Python means managing messy satellite time series — fighting cloud cover with multi-sensor fusion, distinguishing real change from seasonal variation, and balancing alert speed against accuracy to give responders actionable information while the damage is still fresh.