Machine Learning for Forest Loss Monitoring
This project uses satellite imagery and machine learning to classify forest loss events in central Tasmania, distinguishing clearcut logging from fire-driven loss. The goal is to identify spatial patterns and detect potential illegal clearing in protected areas.
A convolutional neural network was trained on 263 manually-labelled Sentinel-2 satellite images (2019–2024) to recognise the visual difference between geometric clearcut boundaries and irregular fire scars. Predictions were then cross-referenced with land tenure data to flag clearcuts occurring outside permitted logging zones.
Key Question: Is unauthorised clearcut logging occurring in Tasmania's protected forests?
Answer: In central Tasmania, fire — not logging — is the dominant driver of forest loss. Only 3.5% of detected loss was clearcut logging, and 96% of those clearcuts occurred in permitted zones.
The overwhelming majority of forest loss in central Tasmania (2019–2024) was fire-driven, particularly from the 2019 and 2023 bushfire seasons. Only 26 of 736 analysed sites showed clearcut logging patterns.
Implication: Fire management — not logging enforcement — should be the priority for forest protection in this region.
While overall clearcut rates are low, 2024 shows a sharp increase: 37% of forest loss in that year was clearcut logging, compared to less than 2% in fire-heavy years (2019, 2023). This suggests recent logging activity has intensified.
25 of 26 predicted clearcuts fell within PTPZ (Private Timber Reserve Zones) where logging is permitted under Tasmanian law. One clearcut was detected in a protected reserve and flagged for ground-truthing to confirm whether it represents a genuine violation or a model error.
The classification system uses a convolutional neural network (CNN) — a type of machine learning model designed to recognise patterns in images. The model was trained on satellite imagery with five spectral bands (visible light + near-infrared + short-wave infrared) to distinguish clearcut logging from other forest loss types.
What the model learned: Clearcuts have geometric boundaries (rectangular logging coupes), while fire scars follow terrain contours with irregular edges. The model detects these shape differences in satellite imagery.
Training data: 263 forest loss sites were manually labelled by inspecting Sentinel-2 imagery. Only 24 were clearcuts (10:1 class imbalance), which required special handling during training to prevent the model from simply predicting "not clearcut" for everything.
Validation: Manual review of 12 high-confidence clearcut predictions achieved 83% precision (10 correct, 2 false positives where fire was misclassified as clearcut).
This analysis covers central Tasmania only (145.5–146.5°E, -43.2 to -42.2°S), representing ~18% of Tasmania's total forest area. Findings may not generalise to western or northern regions with different forest types and fire regimes.
Small training dataset: Only 263 labelled samples were available, with just 24 clearcut examples. Additional labelled data would improve model robustness and allow detection of rarer clearing types.
Binary classification: The model distinguishes "clearcut" from "not clearcut" but does not differentiate clearcut types (salvage logging, plantation harvest, native forest clearfelling) or other loss drivers (disease, windthrow).
Fire/clearcut ambiguity: In 2 of 12 validation cases, fire boundaries were misclassified as clearcuts. Ground-truthing is essential before enforcement action.
Annual temporal resolution: Imagery is annual composites (one per year), so intra-annual dynamics are missed. Some clearcuts may occur between composite dates.
To extend this work into operational forest monitoring, the following would be valuable:
The full dataset, interactive map, and analysis code are publicly available:
Open to collaboration on environmental data science projects and actively seeking opportunities in geospatial analysis and remote sensing