Frontiers in Earth Science | |
Towards fine-grained object-level damage assessment during disasters | |
Earth Science | |
Katelyn Keegan1  Steve Peterson1  Aya El-Sakka2  Ferda Ofli2  Muhammad Imran2  Zainab Akhtar2  Rizwan Sadiq2  | |
[1] MD Community Emergency Response Team (CERT), Gaithersburg, MD, United States;Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar; | |
关键词: object detection; instance segmentation; disaster management; social media; deep learning; disaster object taxonomy; | |
DOI : 10.3389/feart.2023.990930 | |
received in 2022-07-11, accepted in 2023-03-08, 发布年份 2023 | |
来源: Frontiers | |
【 摘 要 】
Social media can play an important role in current-day disaster management. Images shared from the disaster areas may include objects relevant to operations. If these objects are identified correctly, they can offer a preliminary damage assessment report and situational awareness for response and recovery. This research is carried out in collaboration with a Community Emergency Response Team (CERT) to understand the state-of-the-art object detection model’s capability to detect objects in multi-hazard disaster scenes posted on social media. Specifically, 946 images were collected from social media during major earthquake and hurricane disasters. All the images were inspected by trained volunteers from CERT and, 4,843 objects were analyzed for applicability to specific functions in disaster operations. The feedback provided by the volunteers helped determine the existing model’s key strengths and weaknesses and led to the development of a disaster object taxonomy relevant to specific disaster support functions. Lastly, using a subset of classes from the taxonomy, an instance segmentation dataset is developed to fine-tune state-of-the-art models for damage object detection. Empirical analysis demonstrates promising applications of transfer learning for disaster object detection.
【 授权许可】
Unknown
Copyright © 2023 Sadiq, Akhtar, Peterson, Keegan, El-Sakka, Imran and Ofli.
【 预 览 】
Files | Size | Format | View |
---|---|---|---|
RO202310107792901ZK.pdf | 18969KB | download |