0
Urban Building Damage Detection From Very High Resolution Imagery By One-Class SVM and Shadow Information Peijun Li, Benqi...
Outline <ul><li>Introduction </li></ul><ul><li>Methods </li></ul><ul><li>Results and Discussion </li></ul><ul><li>Conclusi...
Introduction Prompt and accurate detection of damage to urban infrastructure caused by disasters (e.g. earthquakes) Very h...
Methods <ul><li>Image segmentation </li></ul><ul><li>Initial building damage detection and shadow change detection </li></...
Flowchart of method Bitemporal images Bitemporal image segmentation Initial building damage detection: OCSVM Shadow and it...
Image segmentation <ul><li>Image segmentation on bitemporal images, in order to keep consistent object boundary  </li></ul...
Multilevel segmentation method (Multichannel watershed transformation + dynamics of contours) Li, P., Guo, J., Song, B.  a...
Initial building damage detection using OCSVM  Building damage (‘building to non-building’): target class Multi-date compo...
One-class Support Vector Machine (OCSVM) <ul><li>Only samples of target class (e.g. building damage) required in training ...
Shadow change detection <ul><li>1, Shadow detection from bi-temporal images </li></ul><ul><ul><li>A histogram thresholding...
Result refinement using shadow change information <ul><li>If a building collapsed, the shadow will disappear. </li></ul><u...
Study area: Dujianyan, China Datasets: Quickbird images (2005, 2008) Results
Initial building damage detection result Spectral data only
Shadow change information Black : shadow change White : no shadow change
Result comparison Building damage detection results by different methods (all in %)   Damaged Undamaged OA Kappa   PA UA P...
Result comparison Spectral only Proposed method
Before After Spectral only Proposed method No damage Damage Result comparison
No damage Damage Before After Spectral only Proposed method Result comparison
Conclusion <ul><li>Combination of spectral information and shadow change information produced significantly higher accurac...
Thank you  for your attention!
Upcoming SlideShare
Loading in...5
×

unrban-building-damage-detection-by-PJLi.ppt

431

Published on

Published in: Technology, Business
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
431
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
20
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Transcript of "unrban-building-damage-detection-by-PJLi.ppt"

  1. 1. Urban Building Damage Detection From Very High Resolution Imagery By One-Class SVM and Shadow Information Peijun Li, Benqin Song and Haiqing Xu Peking University, P. R. China Email: pjli@pku.edu.cn
  2. 2. Outline <ul><li>Introduction </li></ul><ul><li>Methods </li></ul><ul><li>Results and Discussion </li></ul><ul><li>Conclusion </li></ul>
  3. 3. Introduction Prompt and accurate detection of damage to urban infrastructure caused by disasters (e.g. earthquakes) Very high resolution satellite (VHR) images Automated detection and assessment methods: urgently required Fusion of different sensor data, use of single source data Existing methods (VHR optical data): mostly spectral data only, Objective : use of shadow change information to refine results
  4. 4. Methods <ul><li>Image segmentation </li></ul><ul><li>Initial building damage detection and shadow change detection </li></ul><ul><li>Result refinement using shadow information </li></ul>
  5. 5. Flowchart of method Bitemporal images Bitemporal image segmentation Initial building damage detection: OCSVM Shadow and its change detection Result refinement Final result Accuracy assessment
  6. 6. Image segmentation <ul><li>Image segmentation on bitemporal images, in order to keep consistent object boundary </li></ul><ul><li>A multilevel hierarchical segmentation method required: </li></ul><ul><li>initial building damage detection, shadow identification and change detection: different segmentation levels </li></ul>Multitemporal segmentation
  7. 7. Multilevel segmentation method (Multichannel watershed transformation + dynamics of contours) Li, P., Guo, J., Song, B. and Xiao, X., 2011, A multilevel hierarchical image segmentation method for urban impervious surface mapping using very high resolution imagery. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 4(1), 103-116.
  8. 8. Initial building damage detection using OCSVM Building damage (‘building to non-building’): target class Multi-date composite classification: One-class Support Vector Machine (OCSVM) – one-class classifier
  9. 9. One-class Support Vector Machine (OCSVM) <ul><li>Only samples of target class (e.g. building damage) required in training process </li></ul><ul><li>find the maximal margin hyperplane, which best separates the training data from the origin: more training samples, less outliers </li></ul>+1: target class -1: outlier Hyperplane of separation Target samples classified as outliers +1 -1
  10. 10. Shadow change detection <ul><li>1, Shadow detection from bi-temporal images </li></ul><ul><ul><li>A histogram thresholding method for shadow detection </li></ul></ul><ul><ul><li>Based on intensity difference of shadow and non shadow areas </li></ul></ul><ul><ul><li>Bimodal histogram: shadows occupying the lower end of the histogram </li></ul></ul><ul><li>2, Shadow change detection: comparison of shadows detected from two-date images </li></ul>
  11. 11. Result refinement using shadow change information <ul><li>If a building collapsed, the shadow will disappear. </li></ul><ul><li>After building collapse and shadow change were detected, a simple conditional statements to refine the result: </li></ul><ul><ul><li>For each building collapse area detected, if it is adjacent to an area with shadow change, then it will be remained. Otherwise, it will be considered as non building damage area and will be removed. </li></ul></ul><ul><li>The detected patches less than the size of the average buildings in the scene were removed by thresholding. </li></ul>
  12. 12. Study area: Dujianyan, China Datasets: Quickbird images (2005, 2008) Results
  13. 13. Initial building damage detection result Spectral data only
  14. 14. Shadow change information Black : shadow change White : no shadow change
  15. 15. Result comparison Building damage detection results by different methods (all in %)   Damaged Undamaged OA Kappa   PA UA PA UA     Spectral only 69.63 66.41 84.82 86.63 80.25 53.71 Proposed method 63.73 84.75 95.06 84.44 85.88 63.25
  16. 16. Result comparison Spectral only Proposed method
  17. 17. Before After Spectral only Proposed method No damage Damage Result comparison
  18. 18. No damage Damage Before After Spectral only Proposed method Result comparison
  19. 19. Conclusion <ul><li>Combination of spectral information and shadow change information produced significantly higher accuracy than the use of spectral information alone. </li></ul><ul><li>Further investigation: </li></ul><ul><ul><li>* how to extract shadow more accurately, </li></ul></ul><ul><ul><li>* dealing with partly damaged buildings (some walls still intact), </li></ul></ul><ul><ul><li>* more datasets to evaluate,.. </li></ul></ul>
  20. 20. Thank you for your attention!
  1. A particular slide catching your eye?

    Clipping is a handy way to collect important slides you want to go back to later.

×