NCDP Perspectives

The aftermath of a major wildfire presents a grim landscape: communities devastated, homes reduced to ash, and recovery teams racing against time to assess the extent of the damage.  In the U.S., a preliminary damage assessment supports a request for federal assistance through a Stafford Act Presidential Disaster Declaration. This often involves labor-intensive, on-the-ground fieldwork, requiring significant time and resources.      

Recently, the exploration of artificial intelligence (AI) offers possibilities for enhancing the efficiency and speed of damage assessments, affording a shift toward more technologically integrated approaches in disaster management. In exploring the potential of AI, particularly object detection algorithms, this article seeks to provide a balanced perspective on how these technologies are being tested and applied in the field. By examining both the technological challenges and capabilities, we aim to foster a better understanding of AI’s role in disaster response.  

Challenges of the Current Preliminary Damage Assessment Process

After a major disaster in the U.S., the preliminary damage assessment (PDA) process is a crucial first step in determining the extent of destruction and justifying the need for federal assistance. This collaborative effort involves teams from State, Local, Tribal and Territorial (SLTT) government, along with federal officials, documenting the total number of residences impacted and further categorizing these residences as having been destroyed, experiencing major or minor damage, or merely being affected. If specific thresholds are met, FEMA’s Individual Assistance program will be provided. Similarly, a total cost estimate is needed for clearing debris, repairing roads, bridges, recreation areas, and buildings, and restoring utilities and water systems as part of FEMA’s Public Assistance program. The goal of such efforts is to quickly collect data that will support the decision to declare a federal disaster area, to unlock federal funding for these two FEMA programs.          

During a PDA, teams are dispatched to the disaster-stricken areas to conduct on-the-ground assessments. These teams meticulously document the physical damage by taking photographs, noting the extent of the damage on electronic survey forms, and using GPS devices to log specific locations. This manual process requires assessors to physically visit each site, which can be time-consuming, dangerous, labor-intensive, and subject to human error, particularly in areas with severe destruction or hazardous conditions.

A Paradigm Shift in Damage Assessment

Integrating object detection algorithms into the preliminary damage assessment process has the potential to enhance the speed and accuracy with which various types of damage to the built environment can be assessed following disasters. By incorporating machine learning with unmanned aircraft systems (UAS), this technology has become a key tool in computer vision applications. Its defining feature is its ability to process images in real-time, automatically identifying objects and their locations in a single pass. This capability makes it uniquely suited for disaster management scenarios in which swift and accurate assessments are critical.

Once a major wildfire is extinguished and a preliminary damage assessment begins, UAS with cameras can rapidly survey entire neighborhoods, capturing detailed images of the affected areas within hours. A comprehensive aerial survey has the capability to capture thousands of high-resolution photographs that might be inaccessible to human assessors. These images are then analyzed in real-time using object detection systems such as You Only Look Once (YOLO), which segments the images into grids. Each grid cell is scanned for structural anomalies, environmental disruptions, and evidence of fire-related destruction. This technology simplifies the assessment process, cutting down the time required for traditional door-to-door inspections and enabling faster deployment of resources for recovery and reconstruction.

How Object Recognition Works

Image Division

Object recognition, such as You Only Look Once (YOLO), begins by dividing the input image into an instantaneous grid, with each grid cell acting as a small region of interest. These cells are responsible for identifying whether an object, such as a damaged house, is present within their boundaries. 

For a housing damage assessment, aerial imagery captured by UAS is analyzed in smaller sections. Each grid cell focuses on a specific portion of the landscape to identify features such as roof cracks, collapsed structures, or scattered debris. By breaking down the image into smaller, localized segments, this approach streamlines the detection process, enabling the algorithm to efficiently handle complex and cluttered scenes.

Predictions

Once the grid cells identify potential objects, detailed predictions can be made. For instance, each grid cell estimates:

• Whether the cell contains an object, such as a damaged roof or a collapsed house.
• The exact location and size of the detected object within the image are marked with a rectangular box.
• A probability value indicates how certain the algorithm is about the detection.

From Chaos to Action

The following description outlines why this AI technology represents a transformative advancement in disaster damage assessment:

• Efficiency:  AI object detection algorithms can process images and video streams in real time, enabling rapid assessment of large areas. This efficiency is critical when coordinating an emergency response.
• Accuracy: With proper training, this technology can reliably distinguish between different types and degrees of damage, providing granular data for decision-making.
• Scalability: The adaptable nature of this technology allows it to be deployed on UAS, satellites, and ground-level cameras, making it a versatile tool for various operational scales. In the private sector, insurance companies such as Allstate and State Farm leverage AI-powered object detection to streamline post-disaster claims processing and enhance the accuracy of damage assessments. 

Barriers and Ethical Considerations When Incorporating Object Recognition

Despite its promise, deploying this technology as part of a wildfire damage assessment is not without challenges. One of the primary difficulties is training the algorithm to perform effectively in disaster scenarios. Wildfires often leave behind chaotic landscapes with complex, irregular patterns of damage. Smoke, debris, and varying lighting conditions can reduce the clarity of images and hinder accurate detection. Additionally, the speed-accuracy trade-off can result in missed smaller damages, requiring careful optimization to balance performance.

Integrating this technology into disaster response workflows also requires coordination between multiple stakeholders, including SLTT government agencies, federal partners, and the private sector, including insurance companies. UAS equipped with object recognition systems also needs to operate seamlessly with other disaster management tools, such as geographic information system (GIS) platforms and communication networks, to ensure data can be effectively utilized. Additionally, if launched too soon, deploying UAS in wildfire zones could impede the response by entering restricted airspace. This occurred on January 9, 2025, when a civilian drone collided with a firefighting airplane over the Palisades Fire in Los Angeles County, further reducing the window of time needed to save lives and mitigate property damage. These operational constraints must be addressed to fully realize this technology’s potential.

Ethical considerations also play a crucial role. For instance, how can we ensure that the data collected is used responsibly and respects the privacy of individuals? UAS surveillance, even in disaster zones, can inadvertently capture sensitive information about affected communities, leading to concerns about data misuse, overreach, or government surveillance. To address these factors, policymakers and developers must collaborate to create clear guidelines that safeguard privacy and ensure the technology benefits all communities, especially those typically underserved in disaster response efforts.

Conclusion

The integration of artificial intelligence, particularly through object detection algorithms, has the potential to speed up and improve the accuracy of wildfire damage assessment and disaster recovery. By leveraging UAS and real-time image analysis, these technologies enable faster, more accurate damage assessments, reducing the time required for traditional fieldwork and allowing for quicker resource deployment. However, while the potential of this technology in disaster response is vast, several challenges remain. The chaotic nature of wildfire environments, such as the potential for interfering in the ongoing aerial response, must be addressed for optimal performance. Additionally, ethical concerns around privacy, data usage, and UAS operations in restricted airspaces require careful consideration. As this technology continues to evolve, it will be crucial for policymakers, developers, and emergency management professionals to work together to create frameworks that maximize the benefits while safeguarding privacy and ensuring the effective use of resources in disaster recovery. With thoughtful implementation, object recognition algorithms have the potential to significantly improve the resilience and efficiency of disaster recovery efforts, ultimately fast-tracking debris removal and the rebuilding process.

__

References

Durve, Mihir & Bonaccorso, Fabio & Montessori, Andrea & Lauricella, Marco & Tiribocchi, Adriano & Succi, Sauro. (2021). Tracking droplets in soft granular flows with deep learning techniques. The European Physical Journal Plus. 136. 10.1140/epjp/s13360-021-01849-3. 

Jing, Y., Ren, Y., Liu, Y., Wang, D., & Yu, L. (2022). Automatic Extraction of Damaged Houses by Earthquake Based on Improved YOLOv5: A Case Study in Yangbi. Remote Sensing, 14(2), 382. https://doi.org/10.3390/rs14020382