poaching-risk-maps

Geospatial ML pipeline from my MSc dissertation: risk-mapping for orangutan poaching using forest-loss, access proxies (roads/settlements), and conservation context (protected areas).

Results — Borneo Risk Maps

Interpretation: darker/filled areas indicate higher predicted poaching risk. Maps are relative risk surfaces, not counts of confirmed incidents.

Figure 1 — High-Risk Poaching Zones (overview)

What to notice: The overview highlights candidate high-risk regions against provincial boundaries to give geographic context for the predictions that follow.

Figure 2 — Predicted Poaching Incidents (Logistic Regression)

What to notice: Black points mark locations predicted as incidents at a chosen decision threshold. Clusters tend to align with areas of greater access (roads/settlements) and recent forest-loss signals.

Figure 3 — Predicted Poaching Risk Zones (Surface)

What to notice: Contiguous high-risk patches correspond to recent forest-loss clusters and access corridors, while lower-risk regions align with intact/protected forest.

How to read these maps

• Scale: higher intensity/filled area = higher relative risk (not absolute incident counts).
• Inputs (example): forest-loss history (GFW), distance to roads/settlements (OSM), protected-area buffers (WDPA).
• Thresholding: Figure 2 uses a logistic-regression threshold to display predicted incident points; Figure 3 shows a continuous risk surface.
• Caveats: labels are sparse; avoid spatial leakage by splitting train/test by region; access proxies are imperfect.