A Scalable Machine Learning Model for Continental-Scale Wildfire Risk Prediction

This whitepaper presents the Traverse Analytics wildfire risk prediction model, a production-grade machine learning platform designed for continental-scale wildfire risk assessment across the Canadian landmass. Developed jointly by Traverse Analytics and GNO-SYS Technology Ltd. (February 2026), the system generates spatially explicit fire risk predictions across 1,576,727 unique grid cells covering approximately 1.58 million km² of Canadian territory.

The model addresses four critical limitations of existing wildfire risk systems: data complexity, temporal dynamics, geographic scale, and false positive contamination over non-flammable surfaces (e.g., lakes, rivers, and wetlands). It does so through four principal contributions:

Temporal sequence preservation — retains the chronological evolution of environmental conditions per grid cell, capturing drought progression, vegetation curing, and cumulative heat stress.

Domain-driven feature engineering — creates 40–62 wildfire-specific features grounded in fire science, spanning fire weather indices, drought indicators, vegetation stress, topographic risk, and seasonality.

Automated water body discrimination — a two-stage pipeline combining training-time features with post-prediction rule-based filtering to eliminate false positives over non-flammable surfaces.

Scalable architecture — capable of processing millions of observations on commodity hardware (minimum 8 GB RAM), with support for GPU acceleration.

The core algorithm is LightGBM (Light Gradient Boosting Machine), trained on 3,491,146 observations integrating terrain, hydrology, vegetation, land cover, climate, fire history, and infrastructure data spanning early 2002 through end of 2024. On held-out test data, the model achieves a ROC-AUC of 95.71% and a precision of 96.81%, substantially exceeding typical gradient boosting benchmarks reported in the wildfire prediction literature (0.81–0.87 AUC). Out-of-sample evaluation against the 2025 Dryden fire season (NW Ontario) confirms the model's temporal adaptability, correctly distinguishing early-season stability from peak-season escalation with strong spatial alignment to observed fire activity.

Model outputs are mapped to five operational risk categories (Very Low to Very High) using configurable probability thresholds, supporting integration into fire management decision-support systems.