FireWatch — AI-Powered Wildfire Risk Prediction and Fire/Smoke Detection

Traditional wildfire monitoring often depends on lookout towers, manual patrols, fixed cameras, and reports from local communities. These methods can detect fires late, require continuous human attention, and usually do not combine weather-based risk information with visual fire or smoke evidence. Prediction systems alone cannot confirm whether a fire is already visible, while camera systems alone do not provide the daily weather-risk context needed for preparedness. Operators therefore need an integrated platform that combines wildfire risk estimation, visual detection, alert history, system status, and monitoring tools in one workflow.

FireWatch solves this problem through two coordinated AI pipelines. The first pipeline predicts daily wildfire risk for Karabük using weather data and engineered temporal features. A HistGradientBoostingRegressor estimates the continuous Fire Weather Index, while a RandomForestClassifier evaluates high-risk conditions and reduces the probability of missing dangerous days. The second pipeline uses a custom YOLOv8 object-detection model to detect fire and smoke from webcam, camera, or drone-ready video sources. Detected events are stored as persistent alerts with labels, confidence values, timestamps, sources, snapshots, and read/unread status. A FastAPI backend connects prediction, monitoring, scheduling, drone-ready services, and SQLite persistence. A React and TypeScript dashboard presents risk decisions, monitoring feeds, alerts, maps, model information, and system status to the operator.

The FireWatch architecture consists of six main layers: 1. Weather and data layer Open-Meteo daily weather data and historical seasonal records from Karabük are processed into a locked 34-feature input schema. 2. Wildfire risk prediction layer A HistGradientBoostingRegressor predicts continuous FWI values. A RandomForestClassifier then provides a high-risk safety decision using predicted FWI, humidity, wind speed, and fuel-drying information. 3. Visual detection layer A custom YOLOv8 model detects fire and smoke from local camera, webcam, or drone-ready video streams. 4. Backend and API layer FastAPI coordinates prediction, monitoring, alerts, scheduler operations, model information, and drone-ready controls. 5. Persistence layer SQLite stores prediction history, system state, and persistent fire/smoke detection alerts. 6. Frontend dashboard React and TypeScript provide operator-facing views for live monitoring, risk decisions, detection alerts, system and model information, system flow, and map context.