TLDR: A new self-supervised deep learning system, SIT-FUSE, combined with NASA’s TEMPO satellite data, significantly improves the monitoring of wildfires and associated air quality. It accurately distinguishes smoke plumes from clouds, enabling near real-time, hourly mapping of wildfire fronts and smoke, which was previously challenging. This advancement provides more complete data for forecasting and enhances disaster response and air quality warnings in wildfire-prone areas.
Wildfires in the western United States have become increasingly frequent and widespread since 2000, posing significant challenges for monitoring and predicting the air pollutants they release. Smoke plumes from these fires undergo rapid and complex chemical transformations, making it difficult to track their evolution and impact on air quality.
A new research paper, available at this link, introduces an innovative approach to address these challenges. The study, conducted by Nicholas LaHaye, Thilanka Munashinge, Hugo Lee, Xiaohua Pan, Gonzalo Gonzalez Abad, Hazem Mahmoud, and Jennifer Wei, demonstrates how self-supervised deep learning combined with data from NASA’s TEMPO satellite mission can significantly improve wildfire and air quality management.
The TEMPO Breakthrough
The Tropospheric Emissions: Monitoring of Pollution (TEMPO) mission represents a major advancement in atmospheric monitoring. Launched recently, TEMPO is the first space-based instrument to provide hourly observations of air pollution across Greater North America. Unlike previous satellites that could only observe a specific location once a day, TEMPO tracks rapid changes in atmospheric chemical compounds released from wildfires, such as nitrogen dioxide (NO2) and volatile organic compounds (VOCs), almost in real-time.
This hourly data, combined with its higher spatial resolution compared to other instruments, allows TEMPO to fill critical gaps in ground-based monitoring networks. However, a challenge with current TEMPO data is that thick smoke plumes are often misclassified as clouds, leading to the filtering out of valuable atmospheric composition retrievals downwind of the smoke.
SIT-FUSE: A Deep Learning Solution
To overcome this limitation, the researchers developed an innovative self-supervised deep learning system called Segmentation, Instance Tracking, and data Fusion Using multi-SEnsor imagery (SIT-FUSE). This system is designed to accurately distinguish smoke plumes from clouds and map the near real-time hourly spread of wildfire fronts and smoke plumes using data from both GOES-18 and TEMPO satellites.
SIT-FUSE utilizes self-supervised representation learning and segmentation, which is particularly effective for complex domains where labeled datasets are limited. The framework is hierarchical, allowing for different levels of detail in its analysis. It employs Deep Belief Networks (DBNs) for efficient processing, offering a compact model that still provides robust representational capabilities.
A key aspect of SIT-FUSE is its ability to generate smoke plume masks for TEMPO data, restoring missing information in wildfire smoke plumes that were previously misclassified. This means that crucial data on atmospheric compositions like NO2 can now be accurately retrieved within smoke plumes, providing a more complete picture of air quality impacts.
Also Read:
- MSCloudCAM: A New Approach to Cloud Detection in Multispectral Satellite Data
- Advanced Hyperspectral Vision for Unmanned Aerial Vehicles
Demonstrated Success and Future Impact
The efficacy of SIT-FUSE was demonstrated using the Park Fire in California as a case study. The system successfully distinguished between clouds and smoke and accurately identified wildfire fronts. When compared to operational products, the new system showed strong agreement and significant improvement, especially in cases where operational products misclassified thick smoke as clouds.
The validation set, consisting of over 7 million pixels from multiple days, showed high structural similarity index (SSIM) scores for both smoke (0.83) and fire (0.7) when compared against hand-labeled high-certainty pixels. This indicates the system’s reliability in real-world scenarios.
The smoke and fire masks generated by SIT-FUSE from different satellite data streams can be combined to create a comprehensive sub-hourly mapping dataset of wildfire progression. These datasets can be integrated into Earth System Digital Twins and modeling platforms, such as the Jet Propulsion Laboratory’s Fire Alarm and Pyrecast. This capability will not only improve the study of wildfire progression and downstream retrieval coverage but also provide more granular data for short-term fire and smoke forecasting.
Ultimately, this advanced data and forecasting capability will enhance disaster response efforts, enabling agencies like FEMA to issue more timely and accurate wildfire watches and air quality warnings. This proactive approach is vital for protecting lives and property in wildfire-prone regions of North America, particularly in vulnerable communities.
