GeoForest: Satellite Monitoring of Deforestation and Galamsey in Ghana

A geospatial machine learning pipeline that combines multispectral imagery, vegetation indices, and transformer based object detection to track forest loss.

Why this problem

Illegal mining, known locally as galamsey, and unregulated logging have reshaped large parts of Ghana's forest cover over the last decade. Rivers turn ochre from sediment, canopy cover disappears in weeks, and the affected areas are often remote enough that ground surveys arrive after the damage is done.

Satellite imagery closes that gap. A single Sentinel 2 pass covers hundreds of square kilometres at ten metre resolution, revisits the same location every five days, and is free. The engineering question is not whether the data exists. It is whether we can turn a stack of GeoTIFF tiles into a signal that says, for a given location and a given week, the forest here is under stress.

GeoForest is my attempt at that pipeline. The source is at github.com/i-ninte/geoforest, and the dataset it builds on is on Kaggle at ghana-sentinel2-forest.

The pipeline

Sentinel-2 imagery
        ↓
Preprocessing & tiling
        ↓
NDVI computation
        ↓
RF-DETR object detection
        ↓
Temporal change analysis
        ↓
Forest loss / galamsey detection

Two signals feed the final decision. The first is a per pixel measure of vegetation health computed from spectral reflectance. The second is a learned detector that finds visual patterns associated with disturbance, such as cleared land, exposed soil, and small scale mining sites. Neither signal alone is sufficient. NDVI drops for reasons that have nothing to do with mining, and object detectors trained on limited labels miss subtle transitions. Combined, they produce a more reliable signal than either does in isolation.

Data acquisition

The dataset is built with Google Earth Engine. The builder notebook queries Sentinel 2 Level 2A scenes over Ghana, filters by cloud cover, tiles each scene into 512 by 512 patches, exports the patches as GeoTIFFs, and records metadata for reproducibility.

Four spectral bands are retained per tile.

Band 1 → Blue
Band 2 → Green
Band 3 → Red
Band 4 → Near Infrared (NIR)

Blue, green, and red give a natural colour composite that a human can visually verify. Near infrared is the band that matters for vegetation. Chlorophyll reflects strongly in NIR and absorbs in the red, and the ratio between the two is the basis for the vegetation index used downstream.

Every tile is exported as a georeferenced GeoTIFF at 512 by 512 pixels. That size is a compromise. Larger tiles reduce boundary artefacts during inference but consume more memory per forward pass. 512 sits inside the memory budget of a single consumer GPU and matches input sizes used by common detection backbones.

The vegetation index

Vegetation health is estimated with the Normalized Difference Vegetation Index.

$$ \mathrm{NDVI} = \frac{\mathrm{NIR} - \mathrm{Red}}{\mathrm{NIR} + \mathrm{Red}} $$

Values fall between minus one and one. Dense healthy vegetation returns values above 0.6. Bare soil, water, and built up areas fall toward zero or below. NDVI is old, simple, and remarkably durable. It is not a substitute for a learned model. It is a strong prior that constrains the space the learned model needs to explore.

For each tile the pipeline records the mean and standard deviation of NDVI, along with an estimate of vegetation cover. A typical row looks like this.

tile_id,date,ndvi_mean,ndvi_std,vegetation_cover
0001,2025-01-15,0.82,0.03,0.91

The temporal component matters more than any single reading. A tile with an NDVI mean of 0.82 in January and 0.31 in March tells a story that a single snapshot cannot.

Detection with RF-DETR

The learned component of the pipeline is a transformer based object detector, RF-DETR. Choosing a DETR variant rather than a YOLO family model is deliberate.

DETR treats detection as a set prediction problem. There is no anchor grid, no non maximum suppression, no hand tuned matching heuristic. The model outputs a fixed set of predictions and a bipartite matching loss aligns them with the ground truth. For satellite imagery, where target instances are often small, densely packed, and spatially irregular, that formulation removes several sources of error that plague grid based detectors.

The detector is trained to find features that correlate with forest disturbance.

Forest canopy clusters that remain intact.
Cleared land, distinguished from natural clearings by shape and edge structure.
Exposed soil and sediment, which is the visual signature most closely associated with active mining.
Regions with rapid vegetation loss over the observation window.

The detector's output is a set of bounding boxes with class scores. Those boxes feed the temporal analysis stage, not a final decision.

Temporal change detection

A single frame cannot separate seasonal variation from real disturbance. The pipeline compares detections and NDVI statistics across timestamps for each tile. Three patterns are surfaced.

Persistent detections in the same spatial region across multiple dates. These are the most reliable signal for genuine disturbance.
NDVI trajectories that drop below a threshold and do not recover within the expected seasonal window.
Spatial expansion of detections across adjacent tiles, which is the visual signature of an active site growing over time.

Each pattern by itself has false positives. Persistent detections can be roads or existing settlements. NDVI drops can be dry season effects. Spatial expansion can be agricultural clearing. The pipeline flags a tile when at least two of the three patterns agree.

What GeoForest is not

It is worth being explicit about the limits. GeoForest is not a legal instrument. A detection is not proof of illegal activity. Ground truth requires field verification, and the pipeline's role is to prioritise where verification should happen, not to replace it.

Ten metre resolution is coarse for the smallest artisanal mining sites, which can operate under a few pixels of canopy. Sub metre commercial imagery would improve small target recall considerably. The project chose Sentinel 2 because it is free, revisits frequently, and covers the entire country. That trade off is defensible for a monitoring pipeline. It would not be defensible for evidence collection.

What comes next

Several extensions are planned.

Trained detection weights. The current release focuses on the data pipeline and the detection framework. A trained RF-DETR checkpoint on Ghana specific labels is the next milestone.

Time series change detection models. Rather than treating temporal comparison as a rule based step, a sequence model over NDVI trajectories and detection histories would produce a calibrated probability of disturbance per tile.

Near real time ingestion. Sentinel 2 has a five day revisit. Wiring the pipeline to ingest new scenes as they land, rather than in batch, moves the system from retrospective analysis to active monitoring.

Independent verification loop. A small labelled set of confirmed sites and confirmed non sites, refreshed quarterly, would give the pipeline something to measure itself against. Without that loop, every improvement claim is anecdotal.

Closing

The interesting part of GeoForest is not any single model. It is the composition of a physically grounded index that has worked for decades, a modern learned detector that catches what the index misses, and a temporal layer that filters out noise neither of them can resolve on its own. The design principle throughout is that satellite monitoring works when independent signals are made to agree, and fails quietly when any single signal is trusted alone.

Source: github.com/i-ninte/geoforest Dataset: kaggle.com/datasets/kwabenaobeng/ghana-sentinel2-forest

Kwabena Obeng

Comments (0)

Leave a comment