• 沒有找到結果。

Application of Digital Image Correlation Method to the Identification of Landslides

N/A
N/A
Protected

Academic year: 2021

Share "Application of Digital Image Correlation Method to the Identification of Landslides"

Copied!
11
0
0

加載中.... (立即查看全文)

全文

(1)

Abstract

History shows that most earthquakes, which cause serious disaster, are induced by recent movement of faults. For example, the Chichi earthquake, happened on September 21 1999, has a magnitude of 7.3 and leaded to a great loss of lives and properties. The cause of this earthquake is the movement of the Chelungpu fault. The massive earthquake not only destroys the houses and roads, but also induces the landslides. Not only the earthquake but also the typhoon or heavy rainfall can induce the landslides. Because the destruction occurred mostly in remote areas, coupled with the traffic system disruption, the investigation and relief are very difficult.

Digital image correlation (DIC) technique is a non-contact-type optical measurement technique. The progress of digital camera and the fast development of the computer calculation capability cause the digital image correlation technique widely applied to different research fields [1-6]. The landslides will change the appearance of the ground surface and the satellite image is already widely used in the remote sensing technology, therefore we can use the DIC method and the satellite image to monitor the occurrence of landslide.

To determine the occurrence of landslides by comparing satellite images is mainly to find out the position, where correlation between two images is low. Although the algorithm of digital images correlation is to find out the highest correlation, it can be slightly modified to agree with our purpose. In this study, digital image correlation and satellite images are used to identify the occurrence of landslides. A preliminary results of automatic landslides identification are achieved by using DIC. NDVI (Normalized Difference Vegetation Index) and DEM (Digital Elevation Model) are also applied to reduce the misidentification rate.

Keywords: satellite image, digital image correlation, landslide monitoring.

Paper 138

Application of the Digital Image Correlation Method

to the Identification of Landslides

S.H. Tung

1

and M.H. Shih

2

1

Department of Civil and Environmental Engineering

National University of Kaohsiung, Taiwan

2

Department of Civil Engineering

National Chi Nan University, Nantou, Taiwan

©Civil-Comp Press, 2010

Proceedings of the Seventh International Conference on Engineering Computational Technology, B.H.V. Topping, J.M. Adam, F.J. Pallarés, R. Bru and M.L. Romero, (Editors), Civil-Comp Press, Stirlingshire, Scotland

(2)

1 Introduction

In Taiwan, a large amount of hillside developments cause the increase of hillside disasters. The hillside disaster monitoring and prevention are innegligible issues. Landslide is one of the major types of disasters, and may lead to other related disasters.

September 21 1999, the Chi-Chi earthquake of magnitude 7.3 occurred. The movement of the Chelungpu fault causes this earthquake. All of the communication ways for the mountain district are broken. Therefore, the disaster information can not be fully realized. This makes the earthquake relief effort very difficult. At that time, the disaster was interpreted mainly by using the French SPOT satellite images. The satellite images captured on September 27 are compared with the images on 31 December 1998. The results were obtained by comparing the normalized difference vegetation index (NDVI). The resolution of the SPOT satellite image is 20 meters. Therefore, the accuracy of interpretation is not good enough. August 8 2009, Typhoon Morakot brings unexpected rainfall. It caused serious flood in the south and east Taiwan. Serious damages are caused in many mountain districts. Because of the damage of the roads, the rescue persons and equipments can not reach these districts. This situation led to the rescue difficult and slow. Therefore, this study proposed a method to identify the landslide automatically from the satellite image, in which the digital image correlation method is applied to compare the satellite images.

Digital image correlation method has the following two characteristics: 1. The brightness of two images can be different.

2. Deformation of the images is allowed.

Based on the above advantages of digital image correlation method, this study applies the digital image correlation method to carry out the investigation of landslide.

2 Analysis method

2.1 Two-Dimensional Digital Image Correlation Method

In this study, the digital image correlation method and satellite images are applied to identify the occurrence of landslide automatically. The principle of two-dimensional DIC method is to determine the local correlation of two images. The local correlation is used to identify the mapping relationship between the images before and after deformation. The structural speckle will be manufactured on the specimen surface. This makes a different grayscale distribution in the image. This grayscale distribution characteristic is utilized to identify the corresponding position of images before and after deformation.

(3)

Figure 1: Schematic drawing of relative location of sub-images of deformed and undeformed images on surface

As shown in Figure 1, central point prior to deformation is point P, and then changed to point P* after deformation, the correlation between P and P* can be determined by using the correlation coefficient [7,8]:

2 2 ij ij ij ij g g COF g g Σ = Σ ⋅ Σ   (1)

Where, g and ij gij is grayscale of sub-image A on coordinate

( )

,i j and sub-image B on coordinate

( )

i j, respectively. And, coordinate

( )

i j, of sub-image B corresponds to coordinate

( )

,i j of sub-image A. The maximum correlation coefficient is equal to 1. It means that the sub-image B is exactly the image of sub-image A after deformation. Therefore, we are looking for the position which yields the maximum value of the correlation coefficient during the analysis.

2.2 Normalized Difference Vegetation Index (NDVI)

The objects on the ground surface have different reflection characteristics of the electromagnetic wave due to the structure and composition differences. Therefore, the remote sensing techniques use this feature to record the reflected signal of the objects on the ground surface. The intensity difference of the reflected signal can be applied to determine the object type and its distribution approximately. Because the reflection characteristics of the plants, water and soil to the waves of different wavelengths are different, Rouse [9] defines the normalized difference vegetation index (NDVI) as follows:

(NIR-RED) NDVI=

(4)

Where, NIR and RED represent the reflection intensities of near-infrared and red light waves respectively. The NDVI value lies between -1 and +1.

The chlorophyll of green leaves will strongly reflect the near-infrared light (NIR) and strongly absorb the red light (RED). Therefore, the plants have high NDVI values. The reflection intensities of near-infrared light and red light for soil are close. Hence the NDVI values of soil are close to 0. The red light reflection intensity of water is greater than that of near-infrared light. This property results in a NDVI value less than 0. Because the vegetation and non vegetation (for example: barren land, cloud) have different NDVI values. Therefore, this research uses this index to filter the analysis results.

2.3 Digital Terrain Model

Digital terrain model includes digital elevation model (DEM) and digital surface model (DSM). Their contents are outlined as follows (Figure 2):

z Digital Elevation Model (DEM): The three-dimensional topography is digitized. The vegetation and artificial structures are not included in this model.

z Digital Surface Model (DSM): The three-dimensional ground surface (including buildings, vegetation, etc.) is digitized in this model.

Figure 2: Schematic drawing of the digital terrain model.

Because the DEM is closer to the original ground surface, in this study DEM is used to calculate the slope of different position. The slope is applied to filter the results of DIC analysis. The area with low possibility of landslide will be removed in order to reduce the occurrence of misidentification.

3 Experiment Results and Discussions

3.1 Experiment

The experiment steps are as follows:

(5)

in this study. Before analysis, the RGB information is retrieved from the satellite images and transferred into grayscale images. The grayscale value is calculated as follows:

0.299 0.587 0.114

gray= ×red green+ × +blue× (3) The formula of grayscale value is developed according to the human visual

characteristics. The most sensitive color is green, blue is the least. Therefore, three coefficients are used to change the weightings of different colors in equation (3).

2. Image division: The satellite image could cover a very large area. If the covered area is too large, the satellite images will be divided into smaller blocks before the analysis. The analysis results will be combined together at last. This can reduce the calculating time and the calculation system requirements.

3. Set the grid size: DIC are applied to interpret the occurrence of landslide. A grid size is selected. Then all grids on the satellite images will be analyzed. Finally, the correlation coefficient value at each position is applied to determine whether the landslide occurs.

4. Combine the results and filter: The results of every block will be combined together to form the results of the complete analysis region. Then the NDVI and DEM filter will be applied in order to reduce the misidentification rate.

3.2 Analysis Results

In this study, the multi-spectral satellite images captured by the Formosa-II in the area of Kaohsiung Polai are used. The resolution of the satellite image is eight meters. Two images were taken on August 5 2007 (Figure. 3) and October 26 2007 (Figure. 4). They are captured before and after Typhoon Sepat. Therefore, there exist significant landslides in the figures. The images are analyzed by using DIC. The landslide areas caused by Typhoon Sepat are shown in Figure 5.

Figure 3: Satellite image of Kaohsiung

(6)

Figure 5: Landslide areas caused by Typhoon Sepat.(yellow mark)

3.2.1 Analysis results of DIC

In this study, three kinds of grid size, 3×3, 5×5, 7×7 pixels are tested. The correlation coefficients from 0.990 to 0.999 are used as thresholds. The relationship between grid size, correlation coefficient and accuracy is investigated.

Experimental results show that as the grid size equals 5×5 pixels and the correlation coefficient is equal to 0.997 or 0.998, most of the landslide areas can be determined. However, when the correlation coefficient reaches 0.999, the identification error will significantly increase, and this reduces the practice value of this method. In order to compare the experiment results, the accuracy rate is defined as follows:

Identified landslide areas consistent with that identified manually Accuracy rate =

Landslide areas identified manually (4) The accuracy rates are 82.03% and 90.60% as the correlation coefficients are 0.997 and 0.998 respectively. The results are shown in Figure 6 and 7.

Figure 6: The result as grid size is 5×5 pixels and correlation coefficient is 0.997.

Figure 7: The result as grid size is 5×5 pixels and correlation coefficient is 0.998.

(7)

3.2.2 Filter with NDVI

Figure 6 and 7 show that the amount of the identified landslide areas far exceeds that shown in Figure 5. Due to the existed barren land before the occurrence of landslide and the recovery of vegetation of the existed barren land, there exist many wrong identified landslide areas. Because the NDVI can be used to identify the barren land, the efficiency of using NDVI to filter the misidentified landslide area is studied. The results show that the accuracy rates are 67.26% and 74.38% after filtration corresponding to the correlation coefficient 0.997 and 0.998 respectively. The results are shown in Figure 8 and 9. It shows that the wrong identified situation has been greatly reduced.

Figure 8: The result filtered with NDVI corresponding to the correlation

coefficient 0.997.

Figure 9: The result filtered with NDVI corresponding to the correlation

coefficient 0.998.

3.2.3 Filter with DEM

Because the slope of the hillside can be derived from the digital elevation model, this study also tries to use the DEM as the filter. The landslide occurs seldom in the region with lower slope, therefore the area with slope lower than 15% is removed to reduce the misidentification. The slope of the whole region is shown in Figure 10. After filtration, the accuracy rates are 68.44% and 75.22% as the correlation coefficients are 0.997 and 0.998 respectively. The results are shown in Figure 11 and 12. It shows that the misidentification situation is indeed significantly reduced in both cases.

(8)

Figure 10: The slope distribution.

Figure 11: The result filtered with DEM corresponding to the correlation

coefficient 0.997.

Figure 12: The result filtered with DEM corresponding to the correlation

coefficient 0.998.

3.2.4 Filter with NDVI and DEM

NDVI and DEM are also applied simultaneously to filter the area of misidentification. The results are shown in Figure 13 and 14. The accuracy rates are 55.33% and 60.54% respectively. The misidentification situation is reduced substantially, but the accuracy rate is also strong affected.

(9)

Figure 13: The result filtered with NDVI and DEM simultaneously corresponding

to the correlation coefficient 0.997.

Figure 14: The result filtered with NDVI and DEM simultaneously corresponding

to the correlation coefficient 0.998.

3.3 Discussions

The results show that the accuracy rate will increase as the grid size decreases and the correlation coefficient threshold increases, but the misidentification rate also increases accordingly. That is to say, there is no really landslide in many red areas of the analysis results. There are many reasons to cause the misidentification. The main reasons are the remainder errors after orthorectification, existence of clouds and the change of vegetation. In this study, the best grid size is 5×5 pixels. The accuracy rates are 82.03% and 90.60% corresponding to the correlation coefficient threshold of 0.997 and 0.998 respectively. But the misidentification rate is also relative high.

To reduce the misidentification rate, we try to use the NDVI and DEM as filters. The filtration results show that either NDVI or DEM alone or a combination of the both can significantly reduce the misidentification. In other words, the filtration approaches used in this study are feasible. But the accuracy rate is also affected. There are two possible reasons to explain this situation. Firstly, the shooting angles of the satellite images are different. Because of the difference of elevation the image will be still little distorted after the orthorectification. This could make the positions of the same point in two images inconsistent. It leads to the misidentification and the error of filtration. Secondly, the areas with lower slope will be filter out by using the DEM filtering. However, debris flows may flow through some of the flat regions. These regions will not be excluded as the identification process carries out manually. Therefore, some landslide areas are filter out and this induces the decline of the identification accuracy. And the 40-meter data spacing of the DEM also affects the filtration effect.

In addition, the magnitude of the slope has large influence on the occurrence of the landslide. In this study, the areas of slope smaller than 15% are removed in order to improve the accuracy rate of the identification. Because the selected slope threshold has influence on the filtration result, how to select a proper threshold of slope will be studied in the future.

(10)

4 Conclusions

In this study, the digital image correlation method is applied to identify the areas of landslide. The feasibility and accuracy are studied. The following conclusions can be drawn according to the analysis results:

1.

The amount of the identified landslide areas will increase as the grid size decreases. For the same size of grid, more identified landslide areas will be determined by using a larger correlation coefficient threshold. The best result of this experiment is obtained as the grid size is 5×5 pixels, correlation coefficient threshold equals 0.998, and the corresponding accuracy rate is 90.60%.

2.

NDVI can be used to distinguish the existence of vegetation. This property can be applied to remove the misidentified landslide areas. But as the misidentification rate reduces the accuracy rate is also negative influenced. Experimental results show that the highest accuracy rate is 74.38%. To remove the areas with lower slope using DEM can also reduce the DIC misidentification rate. The corresponding accuracy rate is 75.22%. To filter the DIC results with NDVI and DEM simultaneously can filter most of the misidentification, but the accuracy rate reduces to 60.54%.

3.

The results show that using DIC to identify the region of landslide is feasible. This method can save a lot of time and cost compared with the manual method. But the satellite images are shot in different directions. This makes that the images after orthorectification are still lightly different. Therefore, the analysis results will be influenced. In the future, the influence of the orthorectification difference on the identification results can be further studied and discussed.

Acknowledgement

The authors would like to acknowledge the support of Taiwan National Science Council through grant No. NSC-98-2625-M-390-001.

References

[1] D. Raffard, P. Ienny and J.-P. Henry, “Displacement and Strain Fields at a Stone/Mortar Interface by Digital Image Processing”, Journal of Testing and Evaluation, Vol. 29 (2), 115-122, 2001.

[2] M. Dost, D. Vogel, T. Winkler, J. Vogel, R. Erb, E. Kieselstein,“How to detect Edgar Allan Poe’s ‘purloined letter’ - or: Cross correlation algorithms in digitised video images for object identification, movement evaluation and deformation analysis”, Nondestructive Detection and Measurement for Homeland Security, Proceedings of SPIE Vol. 5048, 2003.

[3] M. Dost, N.Rümmler, E. Kieselstein, R. Erb, V. Hillmann, V. Großer, “Correlation Analysis at Grey Scale Patterns in an in-situ Measuring Module for Microsystem Technology”, Materials Mechanics – Fracture Mechanics – Micromechanics, Eds. T. Winkler, A. Schubert, pp. 259-266, Berlin/ Chemnitz, 1999.

(11)

[4] J.C. Kuo, S. Zaefferer, Z. Zhao, M. Winning and D. Raabe, “Deformation Behavior of Aluminum Bicrystals”, Advanced Engineering Materials, 5, 563-566, 2003.

[5] S. Zaefferer, J.C. Kuo, Z. Zhao, M. Winning and D. Raabe, “On the influence of the grain boundary misorientation on the plastic deformation of aluminum bicrystals”, Acta Materialia, 51, 4719-4735, 2003.

[6] M.H. Shih, S.H. Tung, J.C. Kuo, W.P. Sung, "The Application of a Digital Image Correlation Method for Crack Observation", in B.H.V. Topping, G. Montero, R. Montenegro, (Editors), "Proceedings of the Eighth International Conference on Computational Structures Technology", Civil-Comp Press, Stirlingshire, UK, Paper 86, 2006. doi:10.4203/ccp.83.86.

[7] W.H. Peters and W.F. Ranson, “Digital Imaging Techniques in Experimental Stress Analysis”, Optical Engineering, Vol. 21 (3), 427-432, 1982.

[8] T.C. Chu, W.F. Ranson, M.A. Sutton and W.H. Peters, “Application of Digital-Image-Correlation Techniques to Experimental Mechanics”, Experimental Mechanics, 25(3), 232-244, 1985.

[9] J. W. Rouse, R. H. Haas, J. A. Schell and D. W. Deering, “Monitoring vegetation systems in the Great Plains with ERTS”, Third ERTS Symposium, NASA SP-351 I, 309-317, 1973.

數據

Figure 1: Schematic drawing of relative location of sub-images of deformed and  undeformed images on surface
Figure 2: Schematic drawing of the digital terrain model.
Figure 3: Satellite image of Kaohsiung
Figure 6: The result as grid size is 5×5  pixels and correlation coefficient is 0.997
+4

參考文獻

相關文件

volume suppressed mass: (TeV) 2 /M P ∼ 10 −4 eV → mm range can be experimentally tested for any number of extra dimensions - Light U(1) gauge bosons: no derivative couplings. =>

Define instead the imaginary.. potential, magnetic field, lattice…) Dirac-BdG Hamiltonian:. with small, and matrix

incapable to extract any quantities from QCD, nor to tackle the most interesting physics, namely, the spontaneously chiral symmetry breaking and the color confinement.. 

• Formation of massive primordial stars as origin of objects in the early universe. • Supernova explosions might be visible to the most

(Another example of close harmony is the four-bar unaccompanied vocal introduction to “Paperback Writer”, a somewhat later Beatles song.) Overall, Lennon’s and McCartney’s

Microphone and 600 ohm line conduits shall be mechanically and electrically connected to receptacle boxes and electrically grounded to the audio system ground point.. Lines in

™ ™ When ready to eat a bite of your bread, place the spoon on the When ready to eat a bite of your bread, place the spoon on the under plate, then use the same hand to take the

/** Class invariant: A Person always has a date of birth, and if the Person has a date of death, then the date of death is equal to or later than the date of birth. To be