TL;DR
This paper introduces FireCast, a novel deep learning system that leverages AI and GIS to predict wildfire spread with significantly greater accuracy and efficiency than existing models. By analyzing historical fire data, satellite imagery, and weather forecasts, FireCast aims to provide crucial early warnings for firefighters, potentially saving billions in damages and, more importantly, lives.
What Problem Does This Paper Address?
Wildfires are a growing global threat, exacerbated by climate change, leading to immense economic losses and loss of life. Current wildfire spread models, while valuable, often rely on computationally expensive physics-based simulations or mathematical growth predictions. These models can be difficult to implement, require extensive and costly data collection (e.g., drone surveys), and don't always effectively learn from historical fire events. The authors identify a need for a more accessible, efficient, and accurate system that can predict wildfire spread, especially for supporting the critical decision-making of first responders.
Methods
FireCast employs a deep learning approach, specifically a 2D Convolutional Neural Network (CNN) implemented using Keras with a TensorFlow backend. The CNN is designed to learn patterns from various visual inputs, including satellite imagery and elevation data, combined with atmospheric data.
Key aspects of the FireCast algorithm include:
- Supervised Learning: The model is trained on historical wildfire data, learning to predict areas likely to burn in the next 24 hours.
- Geospatial Integration: It combines AI techniques with Geographic Information Systems (GIS) to process and integrate diverse geospatial data.
- CNN Architecture: The CNN utilizes convolutional layers, pooling layers, and dropout layers to extract features from visual inputs. The output of the CNN is concatenated with atmospheric data and fed into a dense layer for final prediction.
- Sliding Window: A sliding window approach is used to analyze the area around sampled pixels of interest (POIs).
- Prediction Output: FireCast outputs a probability map (p ∈ [0,1]) for each pixel, indicating the likelihood of it burning in the next 24 hours. Higher values signify a higher predicted likelihood of burn.
Datasets
- Training Data: Historical wildfire perimeters, Landsat8 satellite imagery, and historical atmospheric data (pressure, temperature, dew point, wind speed/direction, precipitation, humidity) from the Rocky Mountain region of the United States.
- Evaluation Data: The 2016 Beaver Creek, Colorado fire was used as the testing fire.
- Geospatial Data Sources:
- GeoMAC (USGS) for historical fire perimeters.
- GloVis (USGS) for Landsat8 satellite imagery.
- National Oceanic and Atmospheric Administration (NOAA) for atmospheric data.
- USGS National Map for Digital Elevation Models (DEM).
The authors note that their training data was limited and they used data augmentation and weather interpolation to compensate.
Key Insights and Ideas
- AI for Wildfire Prediction is Novel: The paper highlights FireCast as the first application of supervised machine learning for wildfire spread prediction, moving beyond traditional physics-based or purely mathematical models.
- Efficiency and Accessibility: FireCast's reliance on publicly available data and a less computationally intensive AI approach makes it more accessible and faster to implement than established models like Farsite.
- "Hot Spots" as Early Indicators: The emergence of "hot spots" (areas with high predicted burn probability that don't immediately burn) is a key visual insight. While these can negatively impact statistical metrics, they appear to be valuable indicators of high-risk areas that may burn later, potentially weeks in the future, even with data gaps. This suggests a longer-term predictive capability.
- Visual Evaluation is Crucial: The authors emphasize that visual evaluation is as important as statistical metrics for first responders, as it can reveal strategic insights not captured by raw numbers.
Key Results
- Statistical Performance: FireCast outperformed both a random prediction model and Farsite (a commonly used wildfire spread model) in terms of total prediction accuracy, recall, and F-score.
- Total Prediction Accuracy: FireCast achieved an average of 87.7%, compared to 50.4% for random and 67.8%/63.6% for Farsite (wet/dry).
- Recall: FireCast achieved an average of 91.1%, compared to 50.4% for random and 74.8%/81.1% for Farsite (wet/dry).
- F-Score: FireCast achieved an average F-score of 6.4%, significantly higher than random (1.0%) and Farsite (3.5%/2.9%).
- Long-Term Prediction: The model demonstrated predictive power up to two weeks into the future, even across a 14-day data gap in the testing fire, outperforming Farsite in this scenario.
- Efficiency: FireCast requires fewer input variables (4 vs. 7+) and is less computationally expensive than Farsite.
Limitations and Future Work
- Data Availability: The primary limitation is the availability of sufficient and diverse training data. The authors had to rely on data augmentation and interpolation.
- Fire Fighter Activity: The impact of fire fighter intervention on fire spread was not accounted for in the training data.
- Geographic Scope: The current evaluation is limited to the Rocky Mountain region.
- Future Work:
- Gathering more perimeter and historical data from diverse regions worldwide.
- Exploring different input data resolutions.
- Investigating the impact of fire fighter activity.
- Expanding the model's applicability to new regions with varying physical features.
Summary updated