FireSight-IR: A Satellite Infrared Wildfire Detection System Driven by Physics-Informed Neural Networks

FireSight-IR is a satellite infrared wildfire detection system developed by Ibekwemmanuel7 and open-sourced on GitHub (Release date: 2026-05-28, Link: https://github.com/Ibekwemmanuel7/firesight-IR). The core innovation of this system lies in combining physics-informed neural networks (PINNs) with satellite infrared remote sensing technology. It aims to solve problems in traditional wildfire detection methods such as high false alarm rates, poor environmental adaptability, and difficulty in detecting small targets. By integrating physical laws, it improves detection accuracy and effectively suppresses false alarms, providing an efficient solution for wildfire monitoring.

Limitations of Traditional Wildfire Detection

Traditional methods rely on satellite thermal infrared remote sensing and other means, but have the following problems:

• High False Alarm Rate: Easily misjudges high-temperature industrial facilities, solar reflections, cloud edges, etc., as fire points, especially in urban heat islands, deserts, and other areas;
• Insufficient Environmental Adaptability: Fixed thresholds struggle to cope with different climates, surface types, and seasonal changes;
• Real-time Performance and Small Target Detection Issues: Large-scale data processing delays, and initial small fires are easily missed.

Advantages of PINNs

Physics-informed neural networks (PINNs) integrate physical laws (such as heat conduction, energy conservation) as soft constraints into the model. In wildfire detection, they can:

• Use thermal radiation models to distinguish real fire points from other high-temperature objects;
• Verify the temporal continuity of fire points through energy conservation;
• Correct observations by combining atmospheric transmission models;
• Integrate combustion laws to identify typical fire point characteristics.

FireSight-IR adopts a phased processing flow:

1. Data Preprocessing: Receive multi-satellite infrared data (MIR, TIR channels), perform radiometric calibration, geometric correction, atmospheric correction, and adapt to data sources such as MODIS and VIIRS;
2. Candidate Fire Point Screening: Use a dynamic threshold strategy (adjusted based on regional background, season, and time period) to initially screen thermal anomalies;
3. PINN Verification (Core Stage): Evaluate candidate points with physical constraints, including thermal radiation model comparison, energy conservation verification, background temperature gradient analysis, and atmospheric impact correction;
4. False Alarm Suppression: Use classifiers combined with spatial context (population density, industrial distribution, etc.) to filter false alarms such as industrial heat sources, natural heat sources, and reflection interference;
5. Result Output: Output parameters such as fire point location, area, and intensity, supporting emergency system interfaces.

Composite Loss Function

The system uses a loss function that balances data fitting and physical constraints: L_total = L_data + λ1 * L_physics + λ2 * L_boundary

• L_data: Measures the fitting degree between predictions and observed data (mean square error/Huber loss);
• L_physics: Residuals of physical equations such as heat conduction and radiation transfer (calculated via automatic differentiation);
• L_boundary: Ensures boundary conditions (surface changes, coastlines) meet physical expectations; λ1 and λ2 are hyperparameters tuned via the validation set.

Network Architecture

Uses an encoder-decoder convolutional neural network:

• Encoder: Residual connections + attention mechanism to extract multi-scale features (input includes multi-band infrared, NDVI, DEM, and temporal information);
• Physical Constraint Layer: Embeds physical processing modules in the middle layer, integrating physical residuals into backpropagation;
• Decoder: Generates pixel-level fire point probability maps, outputs temperature inversion results and uncertainty estimates.

Training Strategy

• Multi-task Learning: Simultaneously learn fire point detection, temperature inversion, and burn area estimation;
• Domain Adaptation: Adversarial domain adaptation solves data distribution differences across different sensors/regions;
• Hard Example Mining: Focus on training easily misjudged samples such as small fires and edge fires.

Evaluation Metrics

Multi-dimensional evaluation system:

• Detection Accuracy: Precision, recall, F1 score;
• False Alarm Control: False alarm rate, false positive rate;
• Spatial Accuracy: Position error, area estimation error;
• Time Performance: Processing delay, throughput.

Typical Applications

• Early Fire Monitoring: Sensitivity to initial fires with area <100㎡ is better than traditional algorithms (verified in Australia and California tests);
• Large-scale Situational Awareness: Fuses multi-satellite data to provide global wildfire situation maps, supporting risk early warning and resource scheduling;
• Post-disaster Assessment: Precisely delineates the boundaries of burned areas, providing data for ecological restoration and insurance claims;
• False Alarm Analysis: Interpretable design helps identify the source of systematic false alarms.

FireSight-IR will expand in the following directions in the future:

1. Multi-source Data Fusion: Integrate visible light, shortwave infrared, and SAR data (to achieve all-weather monitoring) and connect to ground sensor networks;
2. Fire Spread Prediction Model: Combine meteorological (wind speed, humidity), terrain, and vegetation data to simulate fire spread, supporting emergency decision-making;
3. Edge Computing Deployment: Model quantization and pruning, deploy edge nodes at ground stations/regional centers to achieve near-real-time early warning;
4. Global Wildfire Database: Accumulate historical data to support climate change research and provide open API interfaces.

FireSight-IR is open-sourced on GitHub, providing complete documentation, sample data, and pre-trained models. Community contribution methods include:

• Submit annotated datasets from different regions;
• Improve the parameterization scheme of physical models;
• Develop visualization tools and data interfaces;
• Report issues and suggestions.

Project maintainers regularly hold online seminars to share technical progress and application cases, promoting academic exchanges and industrial cooperation.