Project aim

Detect artisanal ("galamsey") mining inside the Atewa Forest Reserve reliably enough to support monitoring, even under cloud cover and despite mining being a tiny fraction of the landscape.

  • 82%Recall
  • 0.79F1 on held-out tiles
  • <4%Share of landscape mined
  • S1 + S2Sensors fused
Illustrative map for Atewa Illegal Mining Detector Illustrative graphic

Overview

A dual-stream model combines a U-Net for segmentation with YOLOv8 for detection, fusing Sentinel-1 radar with Sentinel-2 optical imagery. Labels come from UAV imagery. On held-out tiles the model reaches 82% recall and an F1 score of 0.79.

The challenge

Mining scars cover under 4% of the landscape, so a model can look accurate while missing most of them. Persistent cloud cover limits optical imagery, and bare soil, roads and settlements look similar to fresh mining pits.

Data & inputs

  • Sentinel-1 SAR
  • Sentinel-2 MSI
  • UAV imagery (labels)

Approach

  1. 01

    SAR + optical inputs

    Sentinel-1 supplies cloud-independent radar texture and Sentinel-2 supplies spectral information; both feed the model.

  2. 02

    UAV-derived labels

    High-resolution UAV imagery provides training labels that satellite data alone could not.

  3. 03

    Dual-stream architecture

    A U-Net segments mining extent while YOLOv8 localises individual sites.

  4. 04

    Imbalance handling

    Hard negative mining and focal loss keep the rare mining class from being ignored.

Results

On held-out tiles the model reaches 82% recall and an F1 score of 0.79 for a class occupying under 4% of the landscape.

Validation

Evaluated on held-out tiles not used in training.