Across Canada, wildfire seasons are becoming more unpredictable, more intense, and more difficult to model using traditional methods. To support decision makers who depend on accurate, forward‑looking intelligence, Traverse Analytics and GNO‑SYS have developed a scalable, production‑grade wildfire risk prediction model capable of analyzing more than 1.58 million km² of terrain and millions of historical environmental data points.
The result is a system that delivers high‑fidelity, spatially explicit wildfire risk assessments, aligned with real fire activity and validated against the 2025 Dryden fire season in Northwest Ontario.
As Trevor Miller, Chief Executive Officer of GNO‑SYS Technology Ltd., explains:
“Wildfire risk is evolving quickly, and our tools must evolve with it. Together with Traverse Analytics, we built a platform that brings engineering rigor and fire‑science expertise together to address the scale and complexity of Canada’s wildfire challenges. This work strengthens national preparedness and supports the agencies protecting communities across the country.”
Amber Rushton, MPS, MBCP, Chief Executive Officer of Traverse Analytics Incorporated, relays:
“Co‑creating this wildfire model with GNO‑SYS has shown what’s possible when shared purpose and complementary expertise come together. By uniting advanced analytics with engineering and fire‑science insight, we’ve developed a model designed to keep pace with the rapidly shifting realities of wildfire risk in Canada. Integrating this model into ResiliSight, our Collective Resilience Platform, elevates it into a cornerstone of our resilience ecosystem—empowering clearer foresight and more unified action to protect communities nationwide.”
Why a new approach was needed
Wildfire behavior is driven by a complex interaction of climate, vegetation, hydrology, terrain, and fuel conditions. Traditional models struggle with:
- Temporal complexity, where months of environmental change get collapsed into averages
- Vast geographic scale, stretching computational limits
- Heterogeneous datasets that require intensive preprocessing
- False positives, particularly over water bodies and non‑flammable terrain
The Traverse Analytics model is engineered specifically to eliminate these constraints through a combination of temporal sequence modeling, domain‑driven wildfire feature construction, and automated water‑body handling.
How the model works through PDAMS
Temporal intelligence
The model preserves multi‑step historical environmental profiles for every grid cell. This allows it to detect patterns such as:
- Multi‑month drought progression
- Vegetation curing cycles
- Seasonal heat accumulation
- Shifts in weather stability and wind conditions
Domain‑driven feature engineering
Rather than relying solely on raw data, the system integrates 40–62 engineered, fire‑science‑based features, including:
- Fire weather indices
- Drought and vegetation stress indicators
- Topographic and seasonal fire risk factors
- Fuel dryness and heat‑drought interaction metrics
- Hydrological indicators that differentiate flammable vs. non‑flammable surfaces
Water‑body discrimination
A two‑stage process (feature‑level awareness plus post‑prediction filtering) ensures that large lakes, rivers, ice surfaces, and coastal zones are not incorrectly flagged as high risk.
Scalable architecture
Built using LightGBM, the platform can:
- Process millions of records on commodity hardware
- Train approximately 20 times faster than traditional boosting models
- Fit within a 10 GB memory constraint
- Scale across national‑level datasets
This makes continental‑scale wildfire forecasting computationally feasible for operational environments.
Performance
On held‑out test data, the model achieves:
- ROC AUC: 95.71 percent
- Precision: 96.81 percent
- Accuracy: 98.51 percent
- Recall: 68.68 percent
The model prioritizes high‑confidence fire‑risk predictions, ensuring reliable identification of high‑risk zones with minimal false alarms.
Real‑world validation: The 2025 Dryden Fire Season
Out‑of‑sample evaluation across June, July, and August 2025 demonstrated that the model:
- Detected early season stability
- Identified regions of environmental instability ahead of ignition events
- Scaled appropriately as conditions intensified mid‑season
- Aligned spatially with verified fire locations across multiple reporting sources
The system captured landscape susceptibility, not just ignitions, which is critical for proactive decision making and resource allocation.
What’s next
Ongoing development will focus on:
- Enhancing recall for greater event sensitivity.
- Integrating real‑time weather forecasts for forward‑looking predictions.
- Improving water‑boundary precision.
- Increasing temporal resolution for rapidly evolving fire conditions.
- Conducting full quantitative validation once Canada’s 2025 fire records are finalized.
- Advancing feature engineering and adding new variables that more accurately reflect on‑the‑ground conditions to further strengthen predictive performance.
- Continuously updating the training dataset and retraining the model annually to incorporate the most current environmental and fire‑season data.
This collaboration shows that continental‑scale wildfire intelligence is now within reach, giving agencies and partners the insight they need to anticipate and prepare for emerging fire risks. For more details, please download the recently published Whitepaper: A Scalable Machine Learning Model for Continental-Scale Wildfire Risk Prediction
