SIGAP: Peatland Wildfire Predictive Early Warning System
High-resolution (1x1 km) machine learning framework predicting peatland wildfire flammability at H-7 and H-30 horizons in Indonesia.
THE QUESTION
How can multi-source satellite telemetry and peatland hydrological features be synthesized into a predictive machine learning model to forecast local flammability risks 7 to 30 days before ignition?
BACKGROUND
Tropical peatland fires in Indonesia release disproportionate carbon emissions and transboundary haze. Existing national monitoring systems (e.g. SIPONGI, BMKG Spartan) either detect fires post-ignition via satellite thermal anomalies or rely strictly on meteorological drought indices. SIGAP delivers predictive flammability forecasting per 1x1 km grid cell at operational H-7 and H-30 planning horizons.
APPROACH OVERVIEW
SATELLITE & WEATHER INGESTION
Ingested 6.75 million historical NASA FIRMS hotspot records (2000–2026) paired with CHIRPS precipitation and ERA5-Land reanalysis.
PEATLAND GRIDDING & STATIC MAPPING
Constructed an 82,170-cell spatial grid (1x1 km) across Riau, linking CIFOR peat depth maps with OpenStreetMap infrastructure buffers.
FEATURE STORE & LEAKAGE CONTROLS
Engineered 28 spatial-temporal features (fire history, neighborhood recurrence, drought deficits) across 7.58 million pixel-weeks.
TEMPORAL WALK-FORWARD MODELING
Trained XGBoost classifiers with Optuna hyperparameter search, isotonic probability calibration, and strict walk-forward backtesting.
OPERATIONAL DISSEMINATION
Automated end-to-end delivery via FastAPI GeoJSON services, interactive MapLibre dashboard, and guardrailed field alert briefings.
METHODS
- •Spatial Resolution & Scope: 82,170 grid cells (1x1 km resolution), covering 49,935 peatland cells in Riau Province, Indonesia.
- •Data Sources: NASA FIRMS (5 satellites), CIFOR Peat Depth (Gumbricht et al.), CHIRPS precipitation, ERA5-Land climate, and OpenStreetMap.
- •Machine Learning Architecture: Dual-horizon XGBoost classifiers (H-7 and H-30) optimized via Optuna (40 trials) with MLflow tracking.
- •Validation Framework: 10-year walk-forward backtesting (2016–2025) with zero future-data temporal leakage.
- •Explainability & Delivery: Subdistrict SHAP feature importance vectors, FastAPI backend, and Next.js / MapLibre v5 frontend.
KEY DATA SNAPSHOT
RESULTS
SIGAP outperforms conventional persistence baselines (ROC-AUC 0.55) and meteorological drought indices (0.63), achieving test ROC-AUC scores of 0.842 (H-7) and 0.813 (H-30). Fire history features and peat depth spatial constraints served as the strongest drivers of predictive flammability.
DISCUSSION
Fire recurrence and historical burn intervals emerged as the dominant predictive drivers, elevating test ROC-AUC from 0.72 to over 0.84.
Walk-forward temporal backtesting across 10 annual folds (2016–2025) confirmed model robustness, outperforming baselines in 100% of evaluation folds.
Coupling machine learning probability outputs with automated, guardrailed operational briefings bridges the gap between raw spatial predictions and practical field ranger patrols.
LIMITATIONS
- •Dynamic high-resolution canal blockage datasets remain scarce, requiring OpenStreetMap waterway distances as a structural proxy.
- •Extreme climate shifts (e.g. abrupt El Niño onset) require quarterly probability recalibration to maintain optimal recall thresholds.
IMPACT & APPLICATION
DATA & REPRODUCIBILITY
Analytical code and specific target coordinates are currently held under institutional review and confidential protocol.
REFERENCES
- 01Gumbricht, T. et al. (2017). An expert system model for mapping tropical wetlands and peatlands. Mitigation and Adaptation Strategies for Global Change, 22(5), 709–727.
- 02Chen, T., & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.
- 03Lundberg, S. M., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems (NeurIPS).