![Sudha radhika to upload in slide share [compatibility mode]](https://image.slidesharecdn.com/sudharadhikatouploadinslidesharecompatibilitymode-100902014520-phpapp01/85/Sudha-radhika-to-upload-in-slide-share-compatibility-mode-31-320.jpg)
![Sudha radhika to upload in slide share [compatibility mode]](https://image.slidesharecdn.com/sudharadhikatouploadinslidesharecompatibilitymode-100902014520-phpapp01/85/Sudha-radhika-to-upload-in-slide-share-compatibility-mode-39-320.jpg)
Sudha radhika to upload in slide share [compatibility mode]
- 1. Using Wavelets as an Effective Alternative Tool for Wind Disaster Detection from Satellite Images Sudha Radhika, Yukio Tamura, Masahiro Matsui Contact: radhikasabareesh@arch.t-kougei.ac.jp radhikasabareesh@arch.t- Wind Engineering Research Center Tokyo Polytechnic University Japan
- 2. OBJECTIVE Automated Identification of Wind Damaged Building Structures from the Pre- and Post- Satellite Images Pre- Post- using Wavelet as an effective tool
- 3. PAST RESEARCHES • Major contribution in disaster detection in earthquake using aerial images by Hasegawa et al 2000, Mitomi et al 2001, Sumer et al 2004 and Ozisik 2004 • Major contribution using satellite imagery is done by Matsuoka et al 2000, Vu et al 2005, in earthquake disaster.
- 4. PAST RESEARCHES • Researches were also done by computational identification using low resolution satellite images in other natural disasters like (1) Wild fire by Ambrosia et al 1998, (2) Flood by Groeve et al 2009, (3) Landslides by Danneels et al 2008 and so on.
- 5. PAST RESEARCHES • In wind disaster major contribution is done by Womble et al 2007 and Womble 2005, using the ordinary 2005, statistical analysis of the histogram of the high resolution satellite image pixel radiance value. • Introduction of RS-Scale (Remote RS- Sensing Scale) table rating building damage scale by Womble 2005.
- 6. PAST RESEARCHES More accurate and faster the damage identification save more lives and more building structures can be restored faster. Wavelets + ANN + High resolution satellite imagery
- 7. RS SCALE Remote Sensing Ground Truth Data Visual Inspection of Satellite Images Scale •No significant change in texture, color, or edges. • Edges are well-defined and linear. RS1 No apparent damage • Roof texture is uniform. • Larger area of roof may be visible. •Nonlinear, internal edges appear Shingles/tiles removed, • Newly visible material gives strong spectral RS2 leaving decking exposed return. • Original outside roof edges are still intact. Decking removed, •Nonlinear, internal edges appear • Holes in roof may not give strong spectral RS3 leaving roof structure return. exposed • Original outside edges usually intact. Roof structure •Original roof edges are not intact. collapsed or removed. RS4 Walls may have • Texture and uniformity may or may not experience significant changes. collapsed.
- 8. METHODOLOGY DATA ACQUISITION BUILDING INSPECTION FOR DAMAGE RECOGNITION TRAINING EXTRACTION OF PIXEL RADIANCE SATELLITE BUILDING DATA FOR IMAGES STRUCTURES TESTING VISUAL FEATURE RECOGNITION EXTRACTION GROUND TRUTH ARTIFICIAL NEURAL DATA VALIDATION NETWORK TRAINING CLASSIFICATION OUTPUT ACTUAL SCALE (RS1, RS2, RS3, RS4) OF TRAINED THE WIND ARTIFICIAL NEURAL DAMAGED NETWORK BUILDING STRUCTURES
- 9. DATA ACQUISITION 2004/03/23 PUNTA GORDA 2004/08/14 PUNTA GORDA Satellite Ground Data Truth Data Before Hurricane After Hurricane Courtesy : Womble, J.A., 2005. TTU Source: DigitalGlobe Co., Ltd, Hurricane Charley
- 10. SYSTEM RECOGNITION 1. Extraction of Building Structures -- 40 samples (houses) 2. Visual Recognition of Samples Extracted -- Categorized into 4 Four Damage Scales (RS1, RS2, RS3, RS4) -- 10 Samples each -- With the available Ground Truth Information -- 6 Samples each for Training and 4 Samples each for Validation
- 11. SYSTEM RECOGNITION 1. Pixel Radiance Data H -- RGB channels -- HSV layers S V HSV : Layers
- 12. SYSTEM RECOGNITION 2. Feature Extraction --(1) --(1) Deletion of common area --(2) --(2) Conventional Feature Extraction Method Before and After --(3) --(3) Wavelet Extraction Method Disaster FEATURES EXTRACTED FOR BOTH METHODS -- Statistical Features • Standard Deviation • Maximum -- Image Features Hse 1 with Common Area Deleted • Edge detection
- 13. STATISTICAL FEATURE STANDARD DEVIATION 1 Standard Deviation Vision Layer 0.8 Blue Channel Green Channel 0.6 Red Channel 0.4 0.2 0 RS1 RS2 RS3 RS4 Remote Sensing Scale
- 14. STATISTICAL FEATURE 1 MAXIMUM Vision Layer 0.8 Blue Channel Green Channel Maximum 0.6 Red Channel 0.4 0.2 0 RS1 RS2 RS3 RS4 Remote Sensing Scale
- 15. EDGE DETECTION • An edge in an image is a contour across which the brightness of the image changes suddenly a b f (i, j ) w * h w(m, n)h(i m, j n) m a nb Where f(i,j) output image pixel h(i,j) input image pixel w(m,n) convolution kernel or a filter mask of size (2a+1) (2b+1). (2a+1) (2b+1).
- 16. EDGE DETECTION • Prewitt Operator :Finds edges using the :Finds Prewitt approximation. approximation. • It measures two components. 1. vertical edge component with kernel wx 2. horizontal edge component with kernel wy. wy. wx = wy =
- 17. Image origin Mask y Image w (−1, −1) w (−1, 0) w (−1, 1) f (x, y) w (0, −1) w (0, 0) w (0, 1) f (x − 1, y − 1) f (x − 1, y ) w (1, −1) w (1, 0) w (1, 1) f (x , y − 1) f (x , y ) f (x , y + 1) Mask x coefficient showing coordinate Eg: Eg: Hse 1 RS3 RS3 arrangements f (x + 1, y − 1) f (x + 1, y ) f (x + 1, y + 1) Pixel of image section under mask
- 18. DISTRIBUTION OF THE DETECTED EDGE PIXEL VALUE RS4 RS2 RS3 RS1
- 20. ANN CLASSIFICATION PROCEDURE Feed Forward – Each input neuron receives input signal and broad casts to hidden layer and pass it to each output unit. Back Propagation of error- Net output is error- compared with the target value. Appropriate error is calculated and it is distributed back to the hidden layer Weights adjusted - accordingly
- 21. WAVELETS Feature Extracted by Wavelet Feature Extraction 2 Dimensional discrete Wavelets are used Family of discrete Wavelets : -Daubechies, Biorthogonal, Coiflets, Symlets, Discrete Meyer Best wavelet -- the larger % Margin of separation between the two least different RS scale (RS1 and RS2) (RS1 RS2)
- 22. WAVELETS Biorthogonal Wavelets --distribution of --distribution the damaged area i.e. Std Dev and damaged edge detection Daubechies -- maximum value of the damaged area RS1 % Marginof separation(RS1& RS2) 1 100 RS2 where σ(RS1) = Average standard deviation of all the sample images at RS1.
- 23. A 2-D WAVELET ANALYSIS 2-
- 24. COMPARISON– COMPARISON– Statistical Features MAXIMUM VALUE – RED BAND 80 % Margin of Separation RED Without Wavelet With Wavelet 60 40 20 0 RS1&RS2 RS2&RS3 RS3&RS4 Remote Sensing Scale
- 25. COMPARISON– COMPARISON– Statistical Features MAXIMUM VALUE – GREEN BAND 60 % Margin of Separation Without Wavelet GREEN 50 With Wavelet 40 30 20 10 0 RS1&RS2 RS2&RS3 RS3&RS4 Remote Sensing Scale
- 26. COMPARISON– COMPARISON– Statistical Features MAXIMUM VALUE – BLUE BAND 80 % Margin of Separation Without Wavelet BLUE With Wavelet 60 40 20 0 RS1&RS2 RS2&RS3 RS3&RS4 Remote Sensing Scale
- 27. COMPARISON– COMPARISON– Statistical Features MAXIMUM VALUE – VISION LAYER 80 % Margin of Separation Without Wavelet VISION With Wavelet 60 40 20 0 RS1&RS2 RS2&RS3 RS3&RS4 Remote Sensing Scale
- 28. Statistical Features – Standard Deviation Margin of Without With Wavelet Separation Wavelet Red Band RS1 and RS2 36 % 57 % RS2 and RS3 53 % 56 % RS3 and RS4 23 % 27 % Green Band RS1 and RS2 26 % 34 % RS2 and RS3 43 % 53 % RS3 and RS4 47 % 56 % Blue Band RS1 and RS2 25 % 32 % RS2 and RS3 39 % 43 % RS3 and RS4 54 % 69 %
- 29. Statistical Features – Standard Deviation Vision RS1 and RS2 37 % 57% RS2 and RS3 53 % 56 % RS3 and RS4 24 % 28% Higher the Margin of Separation More Accurate will be the classification
- 30. IMAGE FEATURES – with and without wavelet Eg: Hse 1 WITH WITHOUT RS3 WAVELET WAVELET When edge detection done with wavelet, the wavelet, detection of non-damaged edges as damaged non- edges are reduced rather than edge detection using Prewitt operator i.e. error reduced
- 32. House Visual System Recognition Ground Number Recognit Without With Truth TRAINING ion Wavelet Wavelet Data SAMPLES Hse 6 RS1 RS1 RS1 NA Hse 8 RS1 RS1 RS1 NA Hse 3 RS2 RS2 RS2 RS2 Hse 4 RS2 RS2 RS2 NA Hse 1 RS3 RS3 RS3 RS3 Hse 2 RS3 RS3 RS3 RS3 Hse 5 RS4 RS4 RS4 NA Hse 7 RS4 RS4 RS4 RS4
- 33. CLASSIFICATION RESULT House Visual System Ground Number Recogni Recognition Truth TESTING tion Without With Data SAMPLES Wavelet Wavelet Hse 9 RS1 RS1 RS1 NA Hse 11 RS2 RS1 RS2 RS2 Hse 10 RS3 RS2 RS4 NA Hse 12 RS4 RS4 RS4 NA
- 34. CLASSIFICATION RESULT WITHOUT WITH BUILDING RS SCALES WAVELET WAVEL CONDITION % ET % RS1 100 100 No obvious damage RS2 60 90 Roof shingles removed and Deck exposed RS3 50 50 Deck removed and Roof structure exposed RS4 90 100 Completely collapsed %Accuracy of Identification of Samples
- 35. CLASSIFICATION RESULT RS1 RS3 RS2 Actual Results from Actual RS3 Damage Wavelet Error RS4 RS3 RS3 Extraction RS4 RS1 RS2 RS3 RS4 RS1 RS2 RS1 RS4 RS4 RS2 Classification Result with Wavelet Extracted Features
- 36. CONCLUSION 1. Wind Damage Building Structures can be successfully identified from the statistical and image features extracted from the Pre- Pre- and the Post- Satellite Images. Post- 2. Classification of the identified Building Structures into different Scales in the Remote Sensing Perspective (RS Scale- Scale- RS1,RS2, RS3 and RS4) is successfully obtained. 3. The % Margin of separation between different RS Scale is obtained and it is observed that features extracted by wavelets have got a larger Margin of separation. separation.
- 37. CONCLUSION Cont……… 4. Larger the Margin of Separation More Accurate will be the classification. classification. 5. Thus an Accurate Identification is done by using Wavelet Feature Extraction. Extraction. 6. Pattern for the following Damaged Portions are recognized successfully as : (a) The distribution of the disaster area Biorthogonal Wavelet pattern (b) The peak value of the disaster area Daubechies Wavelet pattern (c) Faulty or Broken edges Biorthogonal Wavelet pattern
- 38. ACKNOWLEDGEMENTS This study was funded by MEXT, Japan, MEXT, through the Global Center of Excellence Program, 2008-2012, which is gratefully 2008- acknowledged. The authors would also like to thank Dr. Arn J Womble, of Texas Tech., Prof. Ahasan Womble, Kareem of University of Notre Dame & Dr. Masashi Matsuoka of Earthquake Disaster Mitigation Research Center, RIKEN, Japan, for their valuable suggestions and advice.