Territorially-Specialized Machine Learning Models for Wildfire Risk Prediction Across Argentina Using Satellite Data and H3 Hexagonal Grids

federico nicolas sinato, camila rivas

Wildfire risk prediction in large, ecologically diverse countries requires models that account
for regional variation in fire drivers. We present GeoAlertAR-ML, a wildfire risk prediction
system for Argentina that uses an ensemble of regionally specialized Random Forest
classifiers operating over a national hexagonal grid of 13,231 H3 cells. Unlike global fire
danger indices or single-model approaches, our system trains independent models for each
of five ecological regions (Centro, Cuyo, NEA, NOA, Patagonia), routed by a K-Nearest
Neighbors classifier with 99.96% accuracy. The system ingests multisource satellite data —
including GFS meteorological fields, GSMaP precipitation, MODIS vegetation indices,
SRTM topography, Dynamic World land cover, and GHSL population density — and
produces daily hexagon-level risk predictions in four categories (Low, Moderate, High,
Extreme). Cross-validated F1-score averaged 93.2% across regions using GroupKFold to
prevent spatial data leakage. Operational validation against NASA FIRMS hotspots
confirmed 100% detection of active fire clusters in zones predicted as High or Extreme risk,
including the February 2026 Comarca Andina crisis (280+ FIRMS hotspots, 68–72%
predicted risk). A complementary 7-day forecasting model (Model B) based on regional
XGBoost classifiers trained on GFS hindcast data achieved 77.8% F1-score with minimal
degradation across forecast horizons (0.3–3.4% from day+1 to day+7). Additionally, the
system identifies potential anthropogenic ignition anomalies by flagging discrepancies
between low predicted meteorological risk and observed fire activity. GeoAlertAR-ML
demonstrates that regionally specialized models, trained on territory-specific fire histories,
outperform generic global approaches for national-scale wildfire risk assessment. The
system has been operational since late 2025 and won the NASA Space Apps Challenge
2025 Best Mission Concept award.
Keywords: wildfire prediction, machine learning, Random Forest, XGBoost, satellite remote
sensing, H3 hexagonal grid, Argentina, regional models, FIRMS validation