Multi-scale texture and color segmentation of oblique airborne video data for damage classification

Natural disasters are rapid and extreme events within the Earth's system. Since unheralded events are associated with widespread destruction and high mortality, there is a need for rapid, accurate and reliable damage information in the critical post-event hours. In practice, collecting such information is a challenge because of the interruptions in communication systems and access difficulties in affected areas. Oblique airborne video imagery, frequently captured by the media or law enforcement agencies, can provide valuable information in place of satellite images, providing critical information in a timely and low cost manner. However, oblique video imagery currently poses substantial processing, registration and integration challenges due to the nature of video imaging. These constraints cause scale variations of texture and colour information within and between video images.
This work addresses the classification of damaged und undamaged areas using texture and colour in video image segmentation. Specifically, this study investigates the use of multi-scale and multivariate texture based segmentations of oblique video imagery. Theoretically and computationally simple and efficient, we deploy Local Binary Pattern (LBP) and Variance measures to differentiate relative texture patterns of damage and undamaged areas. The approach investigated here was tested on video data acquired of the 1999 Kocaeli, Turkey, earthquake site. The following challenges in the use of video data are identified: (i) differentiating texture patterns of damage classes from undamaged regions using poor quality video imagery, (ii) automating such procedures for entire video sequences, and (iii) geo-referencing the results.