Wildfire Decision Support System

An AI-powered wildfire situational awareness platform built for emergency managers. The system combines satellite hotspot data, machine learning fire spread prediction, and real-time crowd intelligence to generate structured situational briefings and support evacuation decision-making.

Built for ENGO 651 — Advanced Geospatial Topics, University of Calgary, 2025.

Live demo: https://wildfire-ai.com/demo

---

Screenshots

Main Map View

Real-time fire perimeter, ML-predicted risk zones, hotspots, evacuation road status, and situation dashboard.

Crowd Field Report

Public field reports submitted during the event, with comments and Google Maps integration.

AI Situational Awareness Report — Overview

Structured executive briefing generated by 4 specialist AI agents: risk level, stat tiles, situation / key risks / immediate actions.

AI Report — Crowd Intelligence Tab

AI-synthesized crowd intelligence: urgent help requests, fire observations, and situational patterns extracted from field reports.

---

Features

Fire Spread Prediction

• Logistic Regression ML model trained on VIIRS FIRMS hotspot + ERA5 weather data
• Predicts fire perimeter at +3h / +6h / +12h horizons
• Wind-driven analytical model as an alternative projection
• Crowd-augmented prediction: field reports inject additional hotspot anchors

AI Agent Pipeline

Four specialist agents each return structured JSON:

Agent	Output
Risk Agent	Fire behaviour, growth trajectory, weather drivers, risk factors
Impact Agent	Population at risk, affected communities, worsening factors
Evacuation Agent	Primary + alternative routes with waypoints, road warnings
Crowd Agent	Urgent help requests, fire observations, situational patterns
Summary Agent	Risk level, key points, situation / key risks / immediate actions

Reports are cached in AI_report/ per timestep and invalidated when crowd data changes.

Crowd Intelligence

• Citizens submit field reports (fire sighting, road info, help requests)
• AI field report simulator generates GIS-informed synthetic reports for testing
• Reports are anchored to real perimeter, road, and landmark coordinates
• Crowd hotspots are injected into the ML prediction pipeline

Replay Timeline

• Step through historical wildfire timesteps (Fort McMurray 2016)
• Virtual clock synchronises crowd report timestamps with replay position
• Satellite imagery (Sentinel-2) and weather data update per timestep

Streaming Chat

• Stateless AI assistant with full event context
• Suggested questions generated from structured report fields

---

Architecture

Frontend (Vanilla JS + Leaflet)
    │
    ▼
Flask REST API  ──────────────────────────────────────────────────────────────
    │                                                                         │
    ├── /api/events          Fire event list + replay clock                   │
    ├── /api/events/:id/timesteps/:ts_id                                      │
    │       ├── /perimeter   ?crowd=true → crowd-augmented                    │
    │       ├── /hotspots                                                     │
    │       ├── /risk-zones                                                   │
    │       ├── /roads       Evacuation status overlay                        │
    │       ├── /population  At-risk population counts                        │
    │       ├── /report      AI analysis (cached in AI_report/)               │
    │       └── /chat        Streaming Claude/Gemini response                 │
    │                                                                         │
    └── /api/events/:id/field-reports   Crowd intelligence CRUD               │
                                                                              │
Pipeline (background thread)                                                  │
    ├── ERA5 weather download → forecast.json + wind_field.json               │
    ├── FIRMS hotspot fetch + crowd hotspot augmentation                      │
    ├── ML inference (Logistic Regression) → perimeter + risk zones (GeoJSON)             │
    └── Spatial analysis → road status + population counts                    │
                                                                              │
AI Agents (on-demand)                                                         │
    Risk → Impact → Evacuation → [Crowd] → Summary → AI_report/*.json ────────

---

Tech Stack

Layer	Technology
Backend	Python 3.11, Flask, SQLAlchemy, PostgreSQL + PostGIS
ML / Spatial	Logistic Regression, scikit-learn, GeoPandas, Rasterio, Shapely
Weather	ERA5 via CDS API, VIIRS FIRMS
AI	Anthropic Claude API (configurable to Gemini)
Frontend	Vanilla JS, Leaflet.js, CSS custom properties
Auth	JWT (PyJWT), bcrypt

---

Getting Started

Prerequisites

• Python 3.11+
• PostgreSQL with PostGIS extension
• Anthropic API key (or Google Gemini API key)
• ERA5 CDS API credentials

Installation

git clone https://github.com/geo-raypan/wildfire-decision-support.git
cd wildfire-decision-support

pip install -r requirements.txt

Configuration

Copy .env.example to .env and fill in:

DATABASE_URL=postgresql://user:pass@localhost/wildfire
ANTHROPIC_API_KEY=sk-ant-...
SECRET_KEY=your-jwt-secret

Run

cd backend
python main.py

The server starts on http://localhost:5000. Default admin credentials: admin / admin.

---

Data

The system uses the Fort McMurray 2016 wildfire as the primary demonstration event.

Pre-processed data (ERA5, FIRMS, fuel type rasters, ML model weights) is stored under data/events/2016_0001/ and is not included in this repository due to size. The pipeline will attempt to download and process it on first run if CDS API credentials are configured.

---

Project Structure

See PROJECT_STRUCTURE.md for a detailed breakdown of every module.

---

API Reference

Full OpenAPI 3.0 specification: docs/api.yaml

Key endpoint groups:

• POST /api/auth/login — JWT authentication
• GET /api/events — list fire events
• GET /api/events/:id/timesteps/:ts_id/report — generate AI situational report
• POST /api/events/:id/field-reports — submit crowd field report
• POST /api/events/:id/field-reports/simulate — AI-generate synthetic field reports
• GET /api/events/:id/timesteps/:ts_id/chat (SSE) — streaming AI chat