Emberwise: An AI solution for forecasting wildfire ignition risk

The problem

Edmonton faces increasingly complex challenges in managing and mitigating wildfire risks, exacerbated by the global climate crisis and progressively dry, arid conditions. These environmental shifts have dramatically lengthened fire seasons and triggered prolonged droughts, elevating the threat within the city's vast natural spaces, such as its extensive river valley wildlands.

Historically, the information pipeline for evaluating fire danger was entirely manual, making it operationally burdensome, resource-intensive, and prone to human error. Edmonton Fire Rescue Services (EFRS) traditionally relied on city-wide, broad safety warnings based on standard macro-level environmental models. This baseline methodology lacked localised specificity, making it highly difficult for command staff to proactively and strategically deploy limited apparatus where they were most immediately needed. Structural and financial constraints prevented the city from adopting expensive, proprietary alternatives, requiring a solution that could be engineered strictly within existing operational budgets using open-source tools.

The solution

To modernise public safety infrastructure, the City of Edmonton’s Data Science and Research team built Emberwise: an in-house, AI-driven predictive analytics application developed iteratively in two distinct operational phases.

Part 1: Automated Environmental Modelling (Phase 1)

The foundational infrastructure automates the ingestion of localised weather and geographic variables to map out standard spatial hazards. It interfaces with Environment Canada weather station feeds, interpolating missing points automatically, to extract core inputs: raw temperature, relative humidity, wind speed, and 24-hour precipitation.

Using the open-source cffdrs R package, the application calculates data indices from the Canadian Forest Fire Danger Rating System (CFFDRS), mapping out variables from the Fire Weather Index (FWI) Subsystem, such as Fine Fuel Moisture Code (FFMC), Initial Spread Index (ISI), and Buildup Index (BUI). Emberwise layers these indexes over topography maps and the city's urban Primary Land and Vegetation Inventory (uPLVI) database. The uPLVI supplies hyper-local forest fuel metadata (identifying specific combustible tree and grass species) across the complete localised wildland study area, which encompasses roughly 16% of Edmonton's total geographic footprint. The output provides daily maps detailing structural fire descriptions, head fire intensity, and rate of spread from the Fire Behaviour Prediction (FBP) Subsystem to help assess wildfire intensity and spread dynamics after ignition.

Part 2: Behavioural Risk Integration & Granular Forecasting (Phase 2)

Phase 2 expanded the tool beyond passive weather observations into active ignition forecasting by blending spatial data with human behaviour risk modelling. It integrates standardised municipal "calls for service" data from Unison, an internal public safety deployment technology that aggregates cross-departmental demand data.

The system programmatically filters Unison data using fire-related keywords, tracking indicators of human activity strongly correlated with accidental blazes, such as reported illegal encampments, open-flame cooking, or improper disposal of flammables. By combining historical spatial fire logs (previous 30 days) with real-time human behaviour metrics, Emberwise projects future ignition vulnerabilities. The platform translates this integrated data array into an operationally actionable, highly granular 1 km² forecasting grid map, indicating exactly where a wildfire is most probable to ignite within the next day.

The impact

1. Actionable spatial resolution

Transformed broad city-wide warnings into granular 1 km² risk zones, providing command staff with a clear operational picture to guide targeted assets before peak burn hours.

2. Exceptional model refinement

Successfully optimised the machine learning prediction algorithms during development iterations, significantly slashing the false alarm rate down from 40 to 14 "false alarms" for every real fire caught while sustaining an 80% catch rate.

3. Maximised area protection

Modernised wildland fire tracking to secure a localised ecosystem covering approximately 16% of Edmonton's total land area.

4. Strategic resource efficiencies

Automated critical calculation processes, delivering an executive email brief directly to EFRS leadership daily at 8 AM to optimise early scheduling and structural asset distribution.

5. Cross-disciplinary FireSmart alignment

Provided a unified operational foundation supporting multiple municipal FireSmart initiatives, such as prioritising vegetation clearance zones for technicians, providing evidence-based justification for trail closures, and coordinating targeted drone overwatch patrols.

Lessons learned

• Leveraging public safety demand data proved to be a powerful predictive breakthrough. The implementation team discovered that localised call volume and keyword-filtered 311 activity logged within a neighbourhood over a 30-day window are among the most critical spatial predictors for identifying future wildland ignition threats. Municipalities looking to replicate this system should focus heavily on breaking data silos between traditional emergency response logs and civic behavioural lines.
• Open source maximises fiscal innovation. By utilising preexisting in-house expertise and established, open-source technological toolkits (such as the cffdrs library in R), Edmonton successfully launched a complex predictive platform without incurring high-cost procurement or technical vendor dependencies. This highlights that mid-sized public organisations can engineer sophisticated AI architectures entirely within standard parameters of operating budgets.
• Target operational utility over pure data density. A critical transition step for stakeholder buy-in was formatting data explicitly for user workflows. Rather than overwhelming field chiefs with raw mathematical models or abstract statistics, the data team embedded risk rankings into highly structured 1 km² visual segments matching standard dispatch conventions. This ensures predictive analysis translates seamlessly into active tactical planning, directed drone paths, and smart deployment adjustments.

Launch year: 2024