Gamuchirai-Magamba/Bindura-gold-prospectivity

Machine learning for gold prospectivity mapping in Bindura, Zimbabwe # Gold Prospectivity Mapping in Bindura, Zimbabwe **Evaluating machine learning models for gold prospectivity mapping using integrated geological and geochemical data in Bindura District, Zimbabwe.** MSc Computer Science dissertation — University of Zimbabwe, Faculty of Computer Engineering, Informatics and Communications, Department of Computer Science, 2026. **Author:** Gamuchirai Nomsa Magamba (R2425841) · **Supervisor:** Dr C. Gombiro 🔗 **Project site:** gamuchirai-magamba.github.io --- ## Overview This study asks whether machine learning can produce usable mineral prospectivity maps under the severe data constraints typical of underexplored terrains — in this case, the Bindura greenstone belt in Zimbabwe. Three algorithmically distinct methods were compared under identical conditions using **only four verified gold deposits** as training examples and **fourteen integrated predictor variables** across **51,134 pixels (≈46 km²)**: | Model | Approach | Outcome | |---|---|---| | **One-Class SVM** | Novelty detection on positive-only samples | Failed — near-uniform output (0.4622–0.4705), no spatial discrimination | | **Isolation Forest** | Anomaly detection | **Best performer** — clear discrimination; 6.9 km² of high-priority targets (14.98% of study area) | | **XGBoost** | Gradient boosting with synthetic negatives | Overly conservative (mean probability 0.0705); 3,413 unique values but no definitive high-confidence targets | **Key finding:** performance differences were driven by *data quality and quantity*, not by inherent algorithmic capability. Four training deposits, sparse geochemical sampling requiring heavy interpolation, and incomplete spatial coverage explain the failures far better than the algorithms do. ## Results | One-Class SVM | Isolation Forest | XGBoost | |:---:|:---:|:---:| | | | | | No discrimination | 6.9 km² targets | Conservative | Georeferenced outputs (GeoTIFF, EPSG:4326, 274 × 557 px) …