Karhutla Riau — LightGBM operational checkpoint

Production default model for the Riau peatland-fire early-warning system.

Model

model_lgbm_operational.joblib — LightGBM classifier, operational regime (22 channels, including hotspot_count_lag fire history). Trained 2019–2021, validated 2022, tested 2023 (50k training samples).

Metric	Value
PR-AUC	0.712
F1	0.625
Recall	0.583
ROC-AUC	0.879
F1 threshold	0.420

Files

File	Purpose
model_lgbm_operational.joblib	LightGBM classifier (joblib)
checkpoint_operational.json	sidecar: norm_stats, tab_names, channels, best thresholds
comparison_table_operational.csv	full model comparison on the operational regime
shap_importance_operational.png	SHAP feature importance
predictions.parquet	per-cell, per-horizon risk (columns cell_idx, day, probability) — consumed by the Space backend

Consumption

The HuggingFace Space backend (danishritonga/karhutla space) fetches predictions.parquet at runtime (10-min TTL cache) and falls back to simulation mode if fetch/schema validation fails. Set HF_MODEL_REPO=danishritonga/karhutla-risk-model on the Space.