| Summary: | Over the past decades, climate change has driven an increase in the frequency and intensity of natural disasters. In an effort to increase the situational awareness and timely support for search and rescue missions in the aftermath of a disaster, the United States Civil Air Patrol (CAP) gathers aerial imagery of the impacted areas. However, these high resolution and timely images are seldom used for quantitative assessment of damage. This thesis focuses on the following question: How can we use modern computer vision techniques to utilize CAP imagery for post-disaster needs assessment, specifically for the purpose of damage estimation and localization? This question is important because the data gathered by CAP has significant potential to expedite response operations and help reduce significant societal costs. The key technical challenge to address is problem arises from the fact that CAP-gathered aerial images are spatially sparse and oblique, and well-calibrated object detection datasets are not available for damage-prone situations.
To address the aforementioned challenge, we develop an approach to simultaneously detect and localize damage within images using ideas from weakly-supervised object localization and structure from motion. Firstly, we refine a well-known proposed technique called class activation mapping to detect the extent of damage within an image solely relying on image-level labels. Secondly, we utilize structure from motion to georeference batches of CAP images from an area of interest. The main advantage of our approach is that the outputs of these two techniques can be easily combined to assign real-world coordinates to damage hotspots in the aftermath of a natural disaster. Finally, we evaluate its potential using data from the 2016 Louisiana floods and provide estimates of flood-related damage.
Our approach achieves a precision of 88% when compared against official flooding estimates. Practical deployment of this approach depends on how the current practices and technologies used by CAP are tailored to improve damage detection and localization. To this end, we propose the following technical and policy recommendations: 1) Implement best practices that allow for a high-quality image sequences that can be labeled and georeferenced using modern computer vision techniques; 2) Incorporate other sensing modalities such as satellite imagery into CAP imagery analysis for quantitative damage assessment over large spatial regions; and 3) Invest in low altitude imaging technologies and benchmark dataset development.
|
|---|