AI-Powered Fire Probability Prediction and Spread Simulation
Team: The Minions
Problem Statement: Simulation/Modelling of Forest Fire Spread using AI/ML techniques
Technology Stack: Geospatial Analysis + Deep Learning + Cellular Automata
- Primary Repository: Forest Fire Spread System
- Model Development: ML Implementation
- Raw Dataset: Private dataset on Kaggle
- Processed Dataset: Private dataset on Kaggle
- Design & Architecture: System Wireframes (Figma)
Our system, dual-phase forest fire prediction and simulation platform, combines:
- ResUNet-A Deep Learning Model - Predicts fire probability from multi-spectral satellite imagery, stacked into a 10-band raster image
- Cellular Automata Engine - Simulates realistic fire spread dynamics with physics-based rules (realistic part, to be added in the future)
- Interactive Web Interface - Allows real-time scenario testing, visualization and sending & receiving of alerts
- Comprehensive Data Pipeline - Processes Landsat 8, MODIS, SRTM, Sentinel-2 and auxiliary geospatial datasets
Satellite Data → ML Prediction → Fire Spread Simulation → Interactive Visualization
(10-band imagery) → (Probability maps) → (Cellular Automata) → (Web Interface)
- Forest fires have increased by 75% globally in the last decade
- Our target region (Uttarakhand) faces 2000+ fire incidents annually=
- Protecting critical ecosystems and wildlife habitats
- Multi-disciplinary Problem: Combines computer vision, geospatial analysis, and physics simulation
- ISRO BAH Hackathon Opportunity: Platform to contribute to national disaster management
- Cutting-edge Technology Application: Implementing state-of-the-art ResUNet-A architecture
- Scalable Solution Development: Building system for real-world deployment
- Current systems lack spatial precision (30m resolution), and are reactive, instead of being proactive
- Missing Dynamic Simulation, static risk maps don't show fire spread patterns
- Poor Integration: Disconnected tools for prediction vs. simulation
- Complex systems not user-friendly for field personnel
# Multi-source data integration
- Landsat 8 OLI/TIRS (optical + thermal bands)
- GHSL Urban Infrastructure Data
- SRTM Digital Elevation Models
- VIIRS-SNPP Active Fire Products (ground truth)
Custom temporal alignment and spatial registrationKey Innovation: Fire-focused patch sampling strategy (80% fire-prone areas) dramatically improved model training efficiency.
class ResUNetA:
# Enhanced U-Net with Res (residual connections) + A (Atrous Convolution)
- Encoder: 4-stage residual blocks with max pooling
- Bridge: Atrous Spatial Pyramid Pooling (ASPP)
- Decoder: Bilinear upsampling with skip connections
- Output: Sigmoid activation for probability mappingKey Innovation: Focal loss implementation (γ=2.0, α=0.25) to handle severe class imbalance (fire pixels <1% of total).
class FireSpreadEngine:
# Physics-based spread rules
- Moore neighborhood (8-cell connectivity)
- Wind-driven directional bias, Terrain slope effects
- Infrastructure barriers (roads, water, urban)
- Hourly time step simulationKey Innovation: Integration of ML probability maps as base probability conditions for CA engine, creating seamless prediction-to-simulation pipeline.
# Flask REST API + React Frontend
- Real-time simulation execution
- Interactive ignition point selection
- Multiple scenario comparison
- Animation generation and playback- Geospatial Data Complexity: Managing multi-temporal, multi-sensor datasets requires sophisticated preprocessing pipelines (plans automate the pipeline, from data collection to stacking)
- Focal loss + fire-focused sampling reduced false negatives by 60%
- Sliding window prediction with 64-pixel overlap enables processing of large geographical areas efficiently
- Cellular automata simulation must balance accuracy with computational speed for web deployment
- Wind speed/direction critically affects spread patterns (2x faster in wind direction) (to be implemented in the future)
- Satellite Remote Sensing: NIR (Near Infrared) and SWIR (Short-wave Infrared) bands (absored by burning and burned areas) most informative for fire detection
- Fire activity shows strong diurnal patterns requiring time-aware modeling
- Geospatial Analysis: 30m resolution optimal balance between detail and computational feasibility
- Dual-AI Architecture: First system combining ResUNet-A + Probabilistic Cellular Automata for forest fires
- High Spatial Resolution: 30m pixel accuracy vs. 1km standard in existing systems
- Multi-scenario Analysis: Compare different ignition points and weather conditions simultaneously
- Complete Pipeline: From raw satellite data to interactive web visualization
- Production Architecture: Flask API + React frontend ready for cloud deployment
- Validation Metrics: IoU=0.821, Dice=0.857 on validation set demonstrate strong performance
- Temporal Splitting: Training on April data, testing on May prevents data leakage
- Physics-Based Simulation: CA rules incorporate actual fire behavior principles
- Uncertainty Quantification: Confidence zones (high/medium/low/no fire) for decision support
- Interactive Interface: Point-and-click ignition selection on map
- Real-time Feedback: Immediate simulation results with progress indicators
- Visualization Quality: Custom fire-themed colormaps and smooth animations
- Multi-format Output: GeoTIFF, JSON, and PNG outputs for different use cases
- 50 Training Epochs with early stopping and learning rate scheduling
- 10-Band Multi-spectral satellite imagery processing
- 6-Hour Fire Spread simulation with realistic physics
- 5 REST API Endpoints for complete system integration
- 3-Tier Confidence mapping (high/medium/low fire risk)
Our system addresses the critical gap between static fire risk assessment and dynamic spread prediction. Traditional systems show where fires might start, but not how they will spread. We provide:
- Precise Fire Probability Maps (30m resolution)
- Dynamic Spread Simulation (hour-by-hour progression)
- Interactive Scenario Testing (what-if analysis)
- Decision Support Tools (confidence zones, multiple outputs)
- Real-time Deployment Capability (web-based interface)
This enables forest departments, disaster management agencies, and researchers to make data-driven decisions for:
- Resource Allocation (where to position firefighting teams)
- Evacuation Planning (which areas to evacuate first)
- Prevention Strategies (where to create firebreaks)
- Risk Communication (clear visualizations for public awareness)
-
Diakogiannis, F. I., Waldner, F., Caccetta, P., & Wu, C. (2020). ResUNet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing, 162, 94-114. https://doi.org/10.1016/j.isprsjprs.2020.01.013
-
Huot, F., Hu, R. L., Goyal, N., Sankar, T., Ihme, M., & Chen, Y. F. (2022). Next day wildfire spread: A machine learning dataset to predict wildfire spreading from remote-sensing data. IEEE Transactions on Geoscience and Remote Sensing, 60, 1-13. https://doi.org/10.1109/TGRS.2022.3192974
-
Karafyllidis, I., & Thanailakis, A. (1997). A model for predicting forest fire spreading using cellular automata. Ecological Modelling, 99(1), 87-97. https://doi.org/10.1016/S0304-3800(96)01942-4
-
United Nations, Department of Economic and Social Affairs - Sustainable Development. (2015). Transforming our world: The 2030 Agenda for Sustainable Development (A/RES/70/1). https://sdgs.un.org/2030agenda
-
Forest Survey of India. (2023). India State of Forest Report 2023. Ministry of Environment, Forest and Climate Change, Government of India. https://fsi.nic.in/forest-report-2023