Wildfires pose an increasing threat to ecosystems, human life, and infrastructure, particularly in light of climate change and the growing frequency of extreme fire events. Rapid and accurate detection of fire outbreaks is critical for minimizing damage and improving the response time. This research focuses on developing a machine learning model for early and reliable wildfire detection using geostationary satellite data. The main emphasis is on the use of multispectral imagery from the GOES satellite series, which provide images every five minutes. In addition, high-resolution data from polar-orbiting satellites, specifically the VIIRS sensor, are used for training and validating the algorithm.
Background and Motivation: Remote sensing systems currently serve as an important source for wildfire detection and monitoring. Geostationary satellites such as GOES-16 and GOES-18 offer high temporal resolution (every 5–10 minutes), but their main limitation is the relatively coarse spatial resolution of 2 km per pixel. In contrast, polar-orbiting sensors like VIIRS provide global coverage and finer spatial resolution (375 meters). But they have much lower temporal resolution (two passes per day for each region). This gap creates an opportunity to use VIIRS products as a reference for ML fire detection algorithms based on geostationary imagery.
Research Objective: To develop machine learning–based algorithm for fire detection using GOES imagery, and VIIRS fire detections as labels for training.
Methodology:
- Data Collection: GOES-16 and VIIRS imagery were downloaded for extensive regions across the United States.
- Preprocessing: Data were converted to uniform grid, rasterized, and resampled to a consistent spatial resolution.
- Machine Learning: Datasets were split into training, validation, and test sets. Several models were evaluated, including XGBoost, Random Forest, and CatBoost.
- Comparative Evaluation: Model outputs were compared with VIIRS data and the operational GOES Fire Detection and Characterization (FDC) product.
Key Findings:
- The machine learning model outperformed the operational GOES fire detection algorithm in all evaluation metrics (precision, recall, and F1).
- The model detected fire events earlier than the GOES FDC product.
- The machine learning model demonstrated higher detection probability for small fires compared to GOES.
Conclusions and Main Contributions: This research highlights the strong potential of GEO-based ML fire detection algorithm.
Future Directions:
- Expanding the approach to additional geographic regions (for example, Israel) using data from European MTG satellites.
- Developing a global wildfire detection model that integrates all existing geostationary satellite systems.
