SEMANTIC SEGMENTATION OF AERIAL IMGES FOR TSUNAMI DEBRIS DETECTION

DOI Web Site 8 References

Bibliographic Information

Other Title
  • セマンティックセグメンテーションによる航空写真の津波瓦礫判別

Abstract

<p> In this study, semantic segmentation based on deep convolutional neural networks was applied to aerial images taken after the 2011 tsunami event for detecting tsunami debris. With both of SegNet and U-Net, tsunami debris area has been successfully extracted from the images with 75% of IoU at maximum. In the conditions tested in this study, SegNet shows better accuracy than U-Net, and the recall and IoU get higher with larger input image sizes while the precision doesn’t show clear trends with image sizes. Although the total accuracy can be higher with larger image sizes, the reproducibility of the boundaries of debris area seems better with smaller image sizes. The accuracy of debris detection with both of SegNet and U-Net is higher than object-based analysis and also simple CNN based segmentation. For further discussions of optimal learning conditions and architectures, more sample images should be included for the learning and testing processes.</p>

Journal

References(8)*help

See more

Keywords

Details 詳細情報について

Report a problem

Back to top