Harnessing Self-Supervised Deep Learning and Geostationary Remote Sensing for Advancing Wildfire and Associated Air Quality Monitoring: Improved Smoke and Fire Front Masking using GOES and TEMPO Radiance Data

Technical Summary: Harnessing Self-Supervised Deep Learning and Geostationary Remote Sensing for Wildfire and Air Quality Monitoring

Problem Statement
Wildfires in the western United States have increased in frequency and severity since 2000, creating significant challenges for monitoring and predicting wildfire-induced air pollutants. A primary bottleneck is the difficulty in distinguishing wildfire smoke plumes from clouds in satellite imagery. Both exhibit similar characteristics, such as high reflectivity and opacity, leading operational algorithms to frequently misclassify thick smoke as clouds. This misclassification results in the filtering out of critical atmospheric composition data (e.g., NO2, aerosols) in downstream products.

While the Tropospheric Emissions: Monitoring of Pollution (TEMPO) mission offers unprecedented hourly monitoring of trace gases across North America, its current data processing relies on cloud fraction thresholds (e.g., filtering retrievals where cloud fraction > 0.2). Consequently, valuable data within smoke plumes is often discarded. Furthermore, existing ground-based networks have spatial gaps, and Low Earth Orbit (LEO) satellites like TROPOMI provide only daily snapshots, missing the rapid chemical transformations that occur in the first few hours of wildfire emissions.

Methodology
To address these challenges, the authors propose a "Wildfire Mapping System" utilizing an innovative self-supervised deep learning framework called SIT-FUSE (Segmentation, Instance Tracking, and data Fusion Using multi-SEnsor imagery).

• Self-Supervised Learning: The system leverages self-supervised representation learning to operate in domains with limited or no labeled datasets. It employs Deep Belief Networks (DBNs) with 2–3 layers, selected for their compact parameter space (~2 million parameters) and proven representational capabilities compared to larger models.
• Hierarchical Deep Clustering: The framework utilizes Information Invariant Clustering (IIC) arranged in a hierarchical tree structure. Output heads are trained such that child nodes only process samples assigned to their parent label sets. This allows for scene segmentation at varying levels of specificity without requiring exhaustive labeling of every pixel.
• Expert-in-the-Loop Labeling: Instead of requiring 100% labeled training data (as in supervised learning), subject matter experts label only "high-certainty" areas of smoke, fire, and backgrounds. Uncertain boundary areas remain unlabeled. This approach reduces pre-analysis labor while maintaining scientific rigor.
• Multi-Sensor Fusion: The system processes data from GOES-18 (geostationary, high temporal resolution) and TEMPO (trace gas monitoring). These separate pipelines are designed to be aggregated to create sub-hourly progression maps.
• Validation Metrics: Due to the lack of direct ground truth and the inherent uncertainty in boundary cases (e.g., fire fronts), traditional precision/recall metrics are deemed too harsh. Instead, the authors use the Structural Similarity Index (SSIM) to evaluate performance against operational products and hand-labeled high-certainty subsets.

Key Contributions

1. Smoke/Cloud Discrimination: The development of a method to successfully distinguish smoke plumes from clouds using GOES-18 and TEMPO radiance data, addressing a critical gap in current operational products.
2. Data Restoration: The generation of smoke masks that allow for the restoration of missing atmospheric data (such as NO2 tropospheric vertical columns) in wildfire smoke plumes, which would otherwise be filtered out as clouds.
3. Efficient Deep Learning Framework: The demonstration of a hierarchical, self-supervised deep learning system (SIT-FUSE) that achieves robust segmentation with significantly fewer labeled samples than traditional supervised or semi-supervised approaches.
4. Multi-Modal Agreement: The creation of smoke and fire front masks from different sensing modalities (GOES and TEMPO) that show strong agreement, enabling the fusion of data streams for more granular monitoring.

Results
The system was tested using the Park Fire in California on July 26, 2024, as a case study.

• Visual Validation: The system successfully generated smoke masks (blue) and fire front masks (red) that aligned with the visual evidence in GOES-18 RGB imagery, effectively identifying the smoke plume that was previously causing data filtering in TEMPO NO2 products.
• Quantitative Performance: On a validation set of over 7 million pixels from five separate days not seen during training:
  • Compared to operational VIIRS and GOES products (in cases where those products performed well), the system achieved an SSIM of 0.86 for smoke and 0.71 for fire.
  • Compared against hand-labeled high-certainty pixels, the system achieved an SSIM of 0.83 for smoke and 0.70 for fire.
• Operational Improvement: The results demonstrate significant improvement over existing operational products, specifically in scenarios where thick smoke was misclassified as clouds, thereby enabling the retrieval of atmospheric composition data in previously excluded areas.

Significance and Claims
The paper claims that this work represents a step toward transforming wildfire and air quality management in the western U.S. by leveraging the hourly data capabilities of TEMPO and advances in self-supervised deep learning.

• Disaster Response: The products are positioned to enhance disaster response by enabling agencies like FEMA to issue timely and accurate wildfire watches and air quality warnings, potentially protecting lives and property in vulnerable communities.
• Data Integration: The generated masks and progression maps are designed to be integrated into Earth System Digital Twins and modeling platforms, such as the Jet Propulsion Laboratory's Fire Alarm and Pyrecast.
• Future Utility: The authors note that these granular datasets can be leveraged for short-term fire and smoke forecasting, as well as for impact assessments and future risk evaluations regarding fire, smoke damage, and air quality.
• Scientific Exploration: The intermediate segmentation subsets allow subject matter experts to co-explore parameterizations and features correlated with specific subclasses, offering a richer analysis domain for future research.

The authors maintain a modest stance, acknowledging that while the system improves upon operational products, it operates in a domain without direct ground truth, and further work is required to optimize encoder complexity for global operational production.