Published in 16th European Conference on Earthquake Engineering Proceedings, June 18, 2018.
A critical task after a significant earthquake is determining the extent of damage to infrastructure networks. The decision-making process to dispatch emergency, repair, and in-field reconnaissance teams depends on whether road/railways and bridges are passible. Another concern is the rapid identification and resolution of physical disruptions to large-volume gas and water pipeline systems. After large seismic events, citizens, amateur photographers, and journalists now post thousands of photographs to formal/social media platforms. In the past, these images would have had to be reviewed by trained volunteers or expert engineers to evaluate whether: road/rail ways were significantly impacted by ground fracture, heaving, slope failure, or rock slides; bridges experienced severe damage or partial-to-complete collapse; and pipeline systems were interrupted by differential ground movement or liquefaction. The manual review of large imagesets for assessing damage has shown to be inefficient and, in cases, error-prone. This paper presents an automatic and rapid approach, based on computer vision techniques, to assessing damage to above-ground infrastructure networks via images uploaded in the immediate aftermath of the earthquake. The authors developed an algorithm based on deep learning (DL) that automatically tags images. Progress to date shows the algorithm correctly assigns individual tags to 92% of roadway images exhibiting cracking (of varying directionality and severities) and 80% of railways affected by horizontal offset (lateral translation). These results show promise and future research efforts entail tagging both of the aforementioned damage types in a single image.
Number of Pages