Machine Learning · 2026 · complete
SIWES Oil Spill Risk Classifier
Severity classifier for Niger Delta oil spills built on NOSDRA incident records. Three XGBoost setups were compared on a 96%-Minor dataset; class weighting won on macro F1 and minority recall, at a large cost in precision.
Specs
- Source
- NOSDRA Oil Spill Monitor
- Raw data
- 21,107 rows x 42 cols
- Usable
- 13,202 rows (Unknown severity dropped)
- Split
- 10,561 train / 2,641 test, stratified 80/20
- Class balance
- Minor 96.0% / Medium 3.6% / Major 0.4%
- Final model
- Class-weighted XGBoost (lr=0.1, depth=7, 100 trees)
- Macro F1
- 0.420 (baseline 0.346, SMOTENC 0.371)
- Medium recall
- 0.40 (baseline 0.03, SMOTENC 0.08)
- Major recall
- 0.20, i.e. 2 of 10 test cases
- Precision
- 0.12 for both Major and Medium
- Accuracy
- 86.6% (baseline 95.9%)
Features
- -Predicts spill severity (Minor / Medium / Major) from 9 incident features available at reporting time
- -Streamlit app: enter an incident profile, get a severity assessment from the class-weighted model
- -Severity label engineered from NOSDRA barrel thresholds and habitat type
- -Honest imbalance study: unweighted baseline vs SMOTENC (inside a CV pipeline) vs class weighting
- -Five executed notebooks, one per CRISP-DM phase, plus reusable src/ modules
Architecture & tech stack
- 01 IngestNOSDRA CSV, placed in data/raw/ (not tracked in git).
- 02 PreprocessDrop 34 columns (>20% missing, IDs, free text); extract year and month; label-encode categoricals; engineer the severity target.
- 03 SplitStratified 80/20 with random_state=42. The test set holds only 10 Major cases, so Major metrics are indicative.
- 04 ImbalanceBaseline vs SMOTENC (capped at 600 rows per minority class, resampled on training folds only) vs balanced sample weights.
- 05 SelectClass-weighted XGBoost chosen on macro F1 and minority recall; weighted F1 is ignored because the Dummy classifier scores 0.94 on it.
- 06 ExplainSHAP TreeExplainer to be re-run on the final weighted model; earlier driver findings are marked unverified in the README.
- 07 Servestreamlit_app.py loads models/xgboost_weighted_model.pkl and the saved encoders.
Pythonpandasscikit-learnXGBoostGridSearchCVimbalanced-learn (SMOTENC)SHAPStreamlit
Known limitations
- !Major and Medium precision are both about 0.12: roughly 8 in 9 high-severity flags are false alarms. Recall was prioritised on purpose, because a needless investigation visit costs less than missing a serious spill.
- !SMOTENC barely helped (Medium recall 0.03 to 0.08, Major stayed 0.00). Nine features that cannot observe spill volume cap what any resampling can recover.
- !The model is a screening aid, not a decision-maker. Severity is derived from reported volume, so noise in that field carries into the target.
- !SHAP driver findings (cause, habitat, company, year) still need re-verifying against the final model, and results come from a single split.
Local development guide
1. Get the code
$ git clone https://github.com/shalombebx/oil-spill-risk-classifier.git
$ cd oil-spill-risk-classifier2. Virtual environment and dependencies
$ python -m venv venv
$ source venv/bin/activate # Windows: venv\Scripts\activate
$ pip install -r requirements.txt3. Add the dataset
# Not in the repo. Download from nosdra.oilspillmonitor.ng
$ cp ~/Downloads/nosdra_*_complete.csv data/raw/4. Reproduce the pipeline
# Run notebooks 01 -> 05 in order; each depends on the last
$ jupyter notebook notebooks/5. Launch the app
# Needs models/xgboost_weighted_model.pkl and label_encoders.pkl (from notebooks 03-04)
$ streamlit run streamlit_app.py