FireSight | Devpost

Inspiration: Wildfires are worsening every year, threatening lives, communities, and ecosystems across Canada. Ontario, in particular, faces mounting challenges with limited resources and rising response costs.

We wanted to create an AI-powered tool that can forecast next month’s wildfire risk and help authorities optimize their response plans proactively.

What it does: Fire Sight predicts the number of wildfires likely to occur across Ontario in the upcoming month.

Users—especially government agencies or emergency response teams—upload recent monthly weather data, and the system instantly returns:

✅ The estimated wildfire count for the coming month ✅ A risk tier (e.g., Low / Moderate / High) ✅ An interactive map showing Ontario wildfire monitoring stations and displaying prediction results

While the prediction currently applies province-wide, the map sets the stage for more localized insights as more granular data becomes available.

How we build it:We collected and cleaned 22 years of monthly weather data from Environment Canada, focusing on summer months (May–August).

•    We paired this with historical monthly wildfire occurrence data in Ontario from the National Forestry Database.
•    We trained a Random Forest Regressor model in Python to learn the relationship between temperature, precipitation, and cooling degree days and the number of fires.
•    The backend API was built in FastAPI for quick predictions.
•    A frontend in React and Leaflet.js displays results in a clean, interactive map.

Challenges we ran into: • Granular data gaps

We hoped to train our model on historical fire counts by individual weather station to enable truly localized predictions. However, this data wasn’t publicly available in time, so we had to aggregate at the province level. • AI assistant feature scope We initially aimed to create an AI-powered assistant that could automatically suggest specific action plans—like recommended budget allocations or evacuation alerts—based on predicted fire severity. As the deadline approached, we made the difficult decision to focus on core model training and prediction instead. • Capital constraints Accessing richer datasets—including air pressure, lightning strike records, and additional meteorological features—requires paid subscriptions. With more funding, we could train an even more robust forecasting model.

Accomplishments that we're proud of: • End-to-end working solution

We built a functional pipeline from data cleaning and ML training to a web app demo, all in under 36 hours. • Meaningful use case Our model addresses a real and urgent problem—helping public agencies plan for wildfires more proactively. • Team growth and resilience All three of us were brand new to hackathons and just completed our first year of study. We overcame steep learning curves, tight timelines, and a few Git disasters to deliver a project we’re genuinely proud of.

What we learned: • Data wrangling in Python

We learned how to use pandas to clean up raw CSV files and process them into a consistent format ready for training an ML model. • ML model experimentation We explored different approaches to regression, feature engineering, and aggregation to improve accuracy. • Deployment and teamwork We got hands-on experience connecting a backend FastAPI service to a React frontend and collaborating smoothly under time pressure. • The importance of focusing on user value over technical perfection when time is short.

What’s next for Fire Sight: • Granular predictions: Incorporate station-level historical fire counts to pinpoint high-risk regions more precisely.

•    Enhanced weather features: Add air pressure, wind, and lightning strike data for richer modeling.
•    Proactive alerts: Offer AI-generated recommendations and automated notifications when predicted risk exceeds thresholds.
•    Integrations: Work with government agencies to embed Fire Sight predictions into their resource planning workflows.