Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Study of soft classification approaches for identification of earthquake-induced liquefied soil

Study of soft classification approaches for identification of earthquake-induced liquefied soil Geomatics, Natural Hazards and Risk, 2014 Vol. 5, No. 4, 334–352, http://dx.doi.org/10.1080/19475705.2013.811444 Study of soft classification approaches for identification of earthquake-induced liquefied soil SANDEEP SINGH SENGAR*y, ANIL KUMARz, HANS RAJ WASONy, SANJAY KUMAR GHOSHx, Y.V.N. KRISHNA MURTHYz and P.L.N. RAJUz Earthquake Engineering Department, Indian Institute of Technology Roorkee, Roorkee 247667, India Indian Institute of Remote Sensing, Indian Space Research Organization, Dehradun 248001, India Civil Engineering Department, Indian Institute of Technology Roorkee, Roorkee 247667, India (Received 14 December 2012; in final form 29 May 2013) The existence of mixed pixels led to the development of several approaches for soft (or fuzzy) classification in which each pixel is allocated to all classes in varying proportions. However, while the proportions of each land cover within each pixel may be predicted, the spatial location of each land cover within each pixel is not. There exist many different potential techniques for sub-pixel mapping from remotely sensed imagery to identify specific class. The fuzzy-based possibilistic c-means (PCM), noise cluster (NC) and noise cluster with entropy (NCE) classifiers were applied to identify the Bhuj, India (2001), earthquake induced soil liquefaction and compared as soft computing approaches via supervised classifica- tion. The soil liquefaction identification was empirically investigated and com- pared with class-based sensor-independent (CBSI) spectral band ratio using Landsat-7 temporal images. It has been found that CBSI-based temporal indices yield the better results for identification of liquefied soil areas while it was easily separated with pre-earthquake existing water body in that area. The NCE classi- fier performed better for conventional temporal indices, while NC classifier per- formed better for soil liquefaction and PCM classifier performed better for water body identification with CBSI temporal indices. 1. Introduction The last two decades have witnessed the increasing use of remote sensing for under- standing the geophysical phenomena underlying natural hazards. By acquiring temporal data, i.e. both pre- and post-event satellite images, it is possible to identify earthquake-induced soil liquefaction with relative ease and effectiveness by incorpo- rating temporal spectral changes. Gupta et al. (1994, 1995) successfully demon- strated the application of remote sensing techniques to delineate zones of seismically induced soil liquefaction in the Ganga plains. Ramakrishnan et al. (2006) mapped the earthquake-induced soil liquefaction around Bhuj by calculating the absorption of energy in the near-infrared (NIR) and short wave infrared region of electromag- netic spectrum. Champati Ray et al. (2001) and Rao et al. (2001) demonstrated *Corresponding author. Email: talktosengar@yahoo.co.in 2013 Taylor & Francis Soft classification approaches for identification of earthquake-induced liquefied soil 335 applications of principal component analysis and unsupervised image classification techniques, respectively, to change detection induced by the 2001 Bhuj earthquake. Singh et al. (2001) represents the surface manifestations due to Bhuj earthquake by band rationing IRS-1D data. However, the most promising area seems to be the application of soft classification fuzzy approach, not widely used in the field of such events. For this reason, it has been proposed that fuzziness should be accommodated in the classification proce- dure so that pixels may have multiple or partial class membership and handle mixed pixel (Foody et al. 1997). The work has been designed to identify soil liquefaction without merging it with existing water body using coarse-resolution remotely sensed data. To achieve these goals, three classifiers (PCM, NC and NCE) were tested with class-based sensor-independent (CBSI) technique to improve the classifiers output for specific class identification. This paper is structured with introductory part provides a review of the recent studies on the sub-pixel classification of remote sensing data using spectral indices. Then, the 2001 Bhuj earthquake induced soil liquefaction case studies were discussed. After that the proposed methodology will be discussed to identify soil liquefaction at sub-pixel level. Then, results and details on image interpretation and analysis are described. From the results, it has also been tried to explain that the liquefied area clearly discriminated with pre-earthquake existing water body. At the end, conclud- ing benefits were presented about the proposed methodology. 2. Sub-pixel classification To study various aspects related to global change studies and environmental applica- tions, land cover information is one of the crucial data components (Sellers et al. 1995). Multi-spectral-based classification is one of the major approaches for extract- ing land cover information from remotely sensed images. While extracting land cover from remote sensing images, each pixel in the image is allocated to one of the possible classes. In reality, different land covers within a pixel can be found due to continuum of variation in landscape and intrinsic mixed nature of most classes (Ju et al. 2003). Mixed pixels may not be appropriately processed by hard classifiers, which assume that pixels are pure. Mixed pixels may be problem in mapping land cover and it may be more severe in heterogeneous classes from coarse spatial resolution images. Thus, it is important to implement different soft classifiers for handling mixed pixels in an image. The objective of classification is to classify each pixel into one and only one land cover class (i.e. hard classification) or to estimate the partial membership of the clas- ses in a pixel (i.e. soft classification). In fuzzy set theory, an unsupervised approach, fuzzy c-means (FCM) clustering (Bezdek 1981), has been used for classifying remote sensing data. In FCM, the summation of class memberships of a pixel is equal to 1, which is based on a probabilistic constraint. The major drawback of this constraint is that classes present in a pixel are inter-dependent, which may not be the case in reality. In possibilistic c-means (PCM) clustering (Krishnapuram & Keller 1993), probabilistic constraint has been relaxed to produce absolute class memberships, which may indicate the class proportions in a pixel. Most clustering methods are plagued with the problem of noisy data, i.e. characterization of good clusters amongst noisy data. The noise that is just due to the statistical distribution of the 336 S.S. Sengar et al. measuring instruments is usually of no concern. On the other hand, the completely arbitrary noise points that just do not belong to the pattern or class being searched for are of real concern. Dave (1991) introduced concept of ‘noise cluster’ (NC) with ability to detect ‘good’ clusters amongst noisy data. Another approach adopted in this study was the use of spectral indices to generate temporal spectral index data for classification. The temporal index database incorpo- rates temporal change information of a class in classification procesure. It has been found that the spectral response varies for different objects at distinctive electromag- netic regions. This inherent phenomenon can be exploited by performing certain well-defined arithmetical operations to generate new digital data set to highlight the specific phenomenon, such as presence of vegetation or moisture. These procedures are known as spectral indices. Several researchers have evaluated the use of spectral indices in various applications ranging from vegetation, environmental change detec- tion and water resource analysis to mineral, snow and ice studies, post-fire burn detection (Jordan 1969; Rouse et al. 1973; Tucker 1979; Gao 1996; Lyon et al. 1997; Fung & Siu 2000; Wang et al. 2001; Baugh & Groeneveld 2006; Chen et al. 2011). Literature reveals that to identify liquefied area, spectral indices with PCM, NC and noise cluster with entropy (NCE) classifiers have not been used earlier. One of the advantages of using such classifiers is that the specific class of interest (in this study liquefied area) can be identified without merging it with other classes at sub- pixel level. These classifiers require minimal reference data as training sample and less information about image characteristic for identification of disaster damage using coarse-resolution remotely sensed data. 3. Factors responsible for soil liquefaction in study area and data used Soil liquefaction occurs when the soil is saturated, i.e. the space between individual par- ticles is completely filled with water, thus causing a pressure on the soil particles. Prior to an earthquake, the water pressure is relatively low. However, an earthquake may cause the water pressure to increase to such a level that the soil particles may move with respect to each other thus leading to the release of soil moisture. In this process, the soil loses its strength and behaves like a liquid (Liou et al. 1977; Ambraseys 1988). A combination of different geological, seismological and hydrologi- cal factors could be responsible for soil liquefaction induced by the Bhuj earthquake. The study area (figure 1) that was highly affected by soil liquefaction comprises of mainly Mesozoic (sandstones, siltstones, shale and limestone), Tertiary (poorly consoli- dated sandstone, siltstone and clay) and Quaternary (sand and clay) sedimentary sequences (Biswas 1987;Sinha &Mohanty 2012). The Kachchh seismo-tectonic belt extends approximately 250 km (east–west) and 150 km (north–south). It is flanked in the north by Nagar Parkar fault and in the south by Kathiawar fault. The area in between is traversed by several faults/fault systems. Salient among them are Katrol Hill Fault, Kachchh Mainland Fault, Banni Fault, Island Belt Fault and Allah Bund Fault. Rajendran et al. (2001) correlated the cause of soil liquefaction in Kachchh area with the past earthquakes. Thakkar and Goyal (2004) correlated extent of soil lique- faction with function of epicentral distance, magnitude of the earthquake, depth of the hypocentre, availability of the source material and also proximity to the major lineaments due to Bhuj earthquake. Hazarika and Boominathan (2009) observed that 100% saturated soil of Kachchh region is responsible for the soil liquefaction. Soft classification approaches for identification of earthquake-induced liquefied soil 337 Figure 1. Study area Kachchh, Gujarat, India (modified from Biswas 1987). Several soil and water table related factors are also responsible for the site-dependent characteristics of such earthquake-induced soil liquefaction processes, which are as follows (Sarkar & Chander 2003): (a) the nature of the sediments, the grain size and intragranular packing (Seed 1970; Holzer 1989); (b) earthquake hypocentral distance (Papadopoulos & Lefkopoulos 1993); (c) depth of the water table (Seed 1970; Roeloffs 1998); (d) depth of deposit (data available mainly down to a depth of 15 m Krinitzsky et al. 1993); (e) nature and thickness of the cap rock; (f) degree of water saturation of the soil (Seed 1970; Roeloffs 1998); (g) earthquake-induced changes in the pore water pressure (Ishihara et al. 1981); (h) regional topography (Seed 1970). Sarkar and Chander (2003) attributed the reasons of soil liquefaction to (a) shallow water table (less than 15 m) or (b) well-sealed clayey soil overlying thick liquefiable sediments, causing the released water to persist on the earth surface for several days after the main shock. Test data sets for this work have been acquired from ETM sensor of Landsat-7 satellite over two different acquisition dates, i.e. 8 January 2001 and 9 February, 2001. These pre- and post-earthquake remote sensing images of the regions, where such earthquake-induced phenomena occurred, help in systematic monitoring of possible patterns to identify soil liquefaction. 338 S.S. Sengar et al. 4. Methodology adopted There has been requirement for identifying only one class that is, soil liquefaction, in Kachchh area. The adopted methodology for this work has been explained with the help of flow diagram as shown in figure 2. After pre-processing stage, with the help of CBSI (as explained in equation (1)) and conventional index techniques, temporal (two date, i.e. pre- and post-earthquake) index database has been prepared (table 1). Figure 2. Flowchart of methodology adopted. Soft classification approaches for identification of earthquake-induced liquefied soil 339 Table 1. Mean band values from temporal indices data. Conventional indices CBSI indices Soil liquefaction Water body Soil liquefaction Water body Spectral indices Pre-EQ Post-EQ Pre-EQ Post-EQ Pre-EQ Post-EQ Pre-EQ Post-EQ MNDWI 0.388 0.854 0.811 0.901 0.000 0.803 0.674 0.827 (Markham & Barker 1986) NDVI (Rouse 0.454 0.482 0.172 0.188 0.000 0.803 0.674 0.827 et al. 1973) NDWI (Hardisky 0.435 0.874 0.658 0.764 0.000 0.803 0.674 0.827 et al. 1983) SAVI (Huete 0.454 0.482 0.172 0.188 0.000 0.800 0.670 0.823 1988) SR (Birth & 0.149 0.145 0.066 0.062 0.211 0.031 0.035 0.027 McVey 1968) TNDVI (Tucker 0.768 0.772 0.619 0.603 0.000 0.949 0.913 0.956 1979) TVI (Broge & 0.250 0.439 0.317 0.372 0.529 0.623 0.733 0.725 Leblanc 2000) Using these indices, both pre- and post-earthquake index maps were generated. Thus, creating a temporal index map, these data sets have been used as inputs for classification: g ðr; cÞ nir CBSIðr; cÞ¼ ; 8ðr; cÞ2 C ; ð1Þ g ðr; cÞ red where g is the digital number (DN) value of a class, r and c are the row and column of pixels belonging to a class, C is kth class of the pixels and g is the band for which k nir the maximum grey value is found within the pixels belonging to the considered class; analogously, g is the band of the minimum value. red In equation (1), the analyst provides the location of a class in the form of row and column or latitude and longitude. Based on coordinates of a class input, the mini- mum and maximum values of the data are computed. Apart from the location of the class of interest (liquefied area), spectral information as grey value from all the bands will be read. Using minimum and maximum operators, find out which band has min- imum and maximum grey values. The band having the maximum value is designated as NIR and other band having minimum value is designated as RED in different indices as mentioned in table 1. If any index does not have RED and NIR bands (e.g. modified normalized difference water index (MNDWI) and normalized differ- ence water index (NDWI)) then band values replaced by these maximum and mini- mum DN value to achieve maximum index value. If in any pixel, the value of the indices is negative, then replace it with zero values. This enhances the class of interest and requires only the geolocation of a class without any spectral band information of the sensor. 340 S.S. Sengar et al. The objective functions of fuzzy-based classifiers applied in this work may be for- mulated (Miyamoto et al. 2008; Sengar et al. 2013) to identify soil liquefaction at sub-pixel level from equations (2)–(4): "# c N c N XX X X m m ðiÞ PCM ¼ J ðU; VÞ¼ ðm Þ Dðx ; v Þþ h ð1-m Þ ; ð2Þ pcm k i k;i i k;i i¼1 k¼1 i¼1 k¼1 "# c N N XX X m m ðiiÞ NC ¼ J ðU; VÞ¼ ðm Þ Dðx ; v Þþ ðm Þ d ; ð3Þ nc k;i k i k;cþ1 i¼1 k¼1 k¼1 "# c N N N cþ1 XX X XX ðiiiÞ NCE ¼ J ðU;VÞ¼ ðm ÞDðx ;v Þþ ðm Þd þ n m logðm Þ : nce k i ki k;cþ1 k;i k;i i¼1 k¼1 k¼1 k¼1 i¼1 ð4Þ where m is the class membership values of a pixel k belonging to i, Dðx ;v Þ is the k i k;i squared distance in feature space between x and v , Dðx ;v Þ is the squared distance k i k j in feature space between x and v , x is the vector denoting spectral response of a k j k pixel k, v is the collection of vector of cluster centres of class i, v is the collection of i j vector of cluster centres of class j, U is the N c membership matrix, V is the mean vector for class i, c and N are the total number of clusters and pixels, respectively, m is the weighted constant (1< m < 1), c þ 1 is the extra class in the data after con- sidering total number of classes c, d is the resolution parameter and have any float value greater than zero and n is the regularizing parameter. In digital image classification, reference data play an important role, both at train- ing and testing stage of supervised image classification. The training data should be of value, if the environment from which they were obtained is relatively homogenous. However, the ground features are heterogeneous in nature, and many a time, a pixel covers two or more classes due to which remote sensing images are dominated by mixed pixels. As a result, the land cover classes are generally mixed in nature and inter-grade gradually in area. While preparing land cover mapping at regional and global level, coarse-resolution images are used, thereby the chances of occurrence of mixed pixels are high. Shannon entropy and fuzzy-set-based measures such as an index of fuzziness (Binaghi et al. 1999) may be used to estimate the uncertainty in the classification data. Entropy measures show how the strength of class membership in the classification output is partitioned between the classes for each pixel. The value of these measures is maximized (a high degree of uncertainty) when the class membership is partitioned evenly between all the classes, and minimized (a low degree or uncertainty) when the membership is associated entirely with one class. Shannon entropy, a measure conceptualized in terms of probability theory may be computed from equation (5) (Klir 1990): H ¼ m log ðm Þ: ð5Þ k;i 2 k;i i¼1 Soft classification approaches for identification of earthquake-induced liquefied soil 341 But two dimensions are required to show the accuracy of output result, as it arises on a given set of testing data. One is true positive ratio (TPR) and another one is false alarm ratio (FAR). Low FAR (equation (7)) and high TPR (equation (6)) is desirable for good result (Brier & Allen 1951): Number of target pixels correctly detected TPR ¼ ; ð6Þ Total number of target pixels present in the sample Number of background pixels detected as target FAR ¼ : ð7Þ Total number of background pixels presents in the sample 5. Results and discussion The various indices as listed in table 1 provides spectral information of specific class for those temporal spectral indices. The variation in mean indices with the temporal data (two date data) shows the sensitivity of those indices for the specific class. High variation in mean index value (between pre- and post-event) with temporal data means, the changes in specific class with temporal data clearly identified. Different indices help to distinguish reflectance of soils by its moisture content. If this temporal spectral index variation is more from a specific index, then more clearly soil liquefac- tion can be distinguished with water bodies. As in the case of water, the index value has not shown much variation on pre- and post-images because it was not changed due to the earthquake. As in CBSI technique, the band values of conventional indices are replaced by maximum and minimum grey values for the specific class identification to obtain maximum index value. Therefore, conventional-technique-based normalized differ- ence vegetation index (NDVI), MNDWI and NDWI are behaving same if used as CBSI technique because all these indices are using difference and addition of band values in numerator and denominator, respectively. In PCM classifier, the m value is class dependent and has relationship with class membership (equation (2)); therefore, the outputs generated have relationship with constant value ‘m’ and spectral indices as shown in figure 3. The entropy of any class is directly related with membership of the class (equation (5)); therefore, the ‘m’ value for each spectral index should be selected such that the specific class shows minimum entropy (0.005) as given in table 2. The point selected for ‘m’ values as a corner frequency of membership values between 0.996 and 0.992. The selection of optimal value of ‘m’ ensures that the specific class of interest can eas- ily be identified without being overlapped with other classes or background. Simi- larly, less entropy indicates that there is less uncertainty in the identification of the class of interest. Table 3 and figure 4 have been generated by keeping ‘m’ as mentioned in table 2 to find constant value of ‘d’ for NC classifier (equation (3)) for each spectral index. These constant values (‘m’and ‘d’) have been utilized to generate results for NC classifier. The NCE classifier depends on ‘d’and ‘n’ values as in this case we have taken m ¼ 1 (Miyamoto et al. 2008). The d and n values are not class dependent; therefore, these 342 S.S. Sengar et al. Table 2. Optimum values of ‘m’ for PCM classifier. Conventional spectral indices CBSI spectral indices Spectral indices Soil liquefaction Water body Soil liquefaction Water body MNDWI 2.3 2.2 – – NDVI 2.0 2.2 2.3 2.3 NDWI 2.3 1.8 – – SAVI 2.5 2.3 2.2 2.2 SR 2.4 2.5 2.6 1.9 TNDVI 2.5 2.2 2.8 2.7 TVI 2.0 1.5 2.1 2.3 Table 3. Optimum values of ‘d’ for NC classifier. Conventional spectral indices CBSI spectral indices Spectral indices Soil liquefaction Water body Soil liquefaction Water body 4 4 MNDWI 3  10 5  10 –– 3 4 4 5 NDVI 3  10 3  10 5  10 2  10 4 4 NDWI 2  10 2  10 –– 3 4 4 5 SAVI 2  10 3  10 3  10 2  10 2 3 3 3 SR 6  10 4  10 4  10 4  10 2 3 4 5 TNDVI 5  10 8  10 3  10 2  10 3 3 3 4 TVI 2  10 2  10 3  10 2  10 Figure 3. Plots showing optimum value of ‘m’ for PCM classifier. Soft classification approaches for identification of earthquake-induced liquefied soil 343 Figure 4. Plots showing optimum value of ‘d’ for NC classifier. constants vary as shown in figure 5. The constant values for NCE classifier have been selected (d ¼ 8000 and n ¼ 1400) so that the specific class shows minimum entropy (0.005). The classified outputs have been shown in figures 6 and 7 using conventional index approach and in figures 8 and 9 using CBSI approach, with the aim to obtain a better Figure 5. Plot showing optimum values of ‘n’ and ‘d’ for NCE classifier. 344 S.S. Sengar et al. Figure 6. Identification of soil liquefaction using temporal conventional indices data. localization and discrimination of two types (soil liquefaction and water body). Table 4 shows the minimum membership values (up to which specific class not merged with background) for each index. In these images, both water body and liq- uefied areas have high membership values and are seen as bright areas. However, Soft classification approaches for identification of earthquake-induced liquefied soil 345 Figure 7. Identification of water body using temporal conventional indices data. 346 S.S. Sengar et al. Figure 8. Identification of soil liquefaction using temporal CBSI indices data. Soft classification approaches for identification of earthquake-induced liquefied soil 347 Figure 9. Identification of water body using temporal CBSI indices data. 348 S.S. Sengar et al. there are variations in the water body and liquefied areas on different indices. The results of the same have superim- posed on the corresponding water body and soil liquefaction areas on index images. In these results, the memberships for representative feature points keeps as high as possible, while unrepresentative points should have low membership in all clusters. These figures are showing the same areas for identification of soil lique- faction as well as water body and clearly represent that one class (soil liquefac- tion) not merged with other class (water body). The output result shows that all CBSI and conventional temporal indices gener- ated good classification results for soil liquefaction as well as water body except TVI index. This is due to the one class mixed with other classes in TVI indices, while there is a significant change observed in CBSI-TVI as compared with conventional TVI. The output evaluation of the results generated by all classifiers has been carried out with the help of FAR and TPR values (tables 5–7). It has been observed from conven- tional index results that TNDVI and SAVI generated good output for soil liquefaction identification with low FAR and high TPR using all classifiers. Among all conventional spectral index results, MNDWI and NDWI generate best output with FAR ¼ 0.26 and TPR ¼ 0.84 using NCE classifier, while using CBSI index technique for soil liq- uefaction identification TNDVI (NC) with FAR ¼ 0.10 and TPR 0.93. For water body identification SR (NCE) generates best results with FAR ¼ 0.16 and TPR ¼ 0.87 using conventional spectral indices. The CBSI-based NDVI (PCM) generated best output for water body identification with FAR ¼ 0.10 and TPR ¼ 0.93. The results were improved by using CBSI technique and clearly reflected in figures 6–9 and tables 5–7. Table 4. Minimum membership value for PCM, NC and NCE classifier. PCM classifier NC classifier NCE classifier Conventional indices CBSI indices Conventional indices CBSI indices Conventional indices CBSI indices Spectral Soil Water Soil Water Soil Water Soil Water Soil Water Soil Water index liquefaction body liquefaction body liquefaction body liquefaction body liquefaction body liquefaction body MNDWI 0.882 0.941 – – 0.901 0.941 – – 0.980 0.992 – – NDVI 0.992 0.784 0.882 0.909 0.992 0.803 0.882 0.941 0.980 0.988 0.980 0.980 NDWI 0.882 0.992 – – 0.894 0.992 – – 0.980 0.980 – – SAVI 0.937 0.784 0.882 0.921 0.941 0.784 0.882 0.952 0.972 0.980 0.972 0.980 SR 0.968 0.784 0.686 0.901 0.968 0.843 0.745 0.874 0.984 0.988 0.960 0.988 TNDVI 0.933 0.784 0.901 0.960 0.921 0.764 0.862 0.960 0.980 0.992 0.980 0.980 TVI 0.835 0.996 0.941 0.901 0.803 0.996 0.956 0.933 0.992 0.992 0.980 0.980 Soft classification approaches for identification of earthquake-induced liquefied soil 349 Table 5. FAR and TPR values for PCM classifier outputs. Soil liquefaction Water body Conventional CBSI Conventional CBSI spectral indices spectral indices spectral indices spectral indices Soil liquefaction FAR TPR FAR TPR FAR TPR FAR TPR MNDWI 0.27 0.69 – – 0.17 0.85 – – NDVI 0.20 0.62 0.14 0.87 0.15 0.87 0.10 0.93 NDWI 0.25 0.67 – – 0.56 0.38 – – SAVI 0.35 0.87 0.13 0.88 0.16 0.82 0.11 0.87 SR 0.35 0.87 0.12 0.89 0.14 0.86 0.10 0.76 TNDVI 0.35 0.87 0.11 0.86 0.15 0.85 0.24 0.87 TVI 0.66 0.55 0.85 0.34 0.85 0.34 0.38 0.40 Table 6. FAR and TPR values for NC classifier outputs. Soil liquefaction Water body Conventional CBSI Conventional CBSI spectral indices spectral indices spectral indices spectral indices Spectral indices FAR TPR FAR TPR FAR TPR FAR TPR MNDWI 0.20 0.64 – – 0.16 0.67 – – NDVI 0.45 0.60 0.16 0.89 0.16 0.82 0.12 0.86 NDWI 0.25 0.60 – – 0.57 0.47 – – SAVI 0.30 0.79 0.14 0.89 0.14 0.83 0.12 0.88 SR 0.33 0.84 0.09 0.78 0.15 0.84 0.13 0.64 TNDVI 0.34 0.88 0.10 0.93 0.13 0.83 0.25 0.88 TVI 0.62 0.51 0.54 0.47 0.84 0.38 0.41 0.58 Table 7. FAR and TPR values for NCE classifier outputs. Soil liquefaction Water body Conventional CBSI Conventional CBSI spectral indices spectral indices spectral indices spectral indices Spectral indices FAR TPR FAR TPR FAR TPR FAR TPR MNDWI 0.26 0.84 – – 0.12 0.80 – – NDVI 0.20 0.67 0.15 0.88 0.17 0.78 0.09 0.85 NDWI 0.26 0.84 – – 0.51 0.79 – – SAVI 0.35 0.82 0.14 0.89 0.13 0.82 0.12 0.77 SR 0.48 0.76 0.38 0.71 0.16 0.87 0.12 0.80 TNDVI 0.32 0.82 0.12 0.85 0.13 0.82 0.45 0.90 TVI 0.64 0.35 0.74 0.45 0.54 0.35 0.47 0.48 350 S.S. Sengar et al. 6. Conclusion This work tries to study importance of soft classification fuzzy techniques such as PCM, NC and NCE for identification of earthquake-induced liquefied soil as specific class; from the results, following conclusions can be drawn. (a) CBSI approach generates better discrimination between liquefied areas and water body as compared with conventional indices. (b) One class of interest (soil liquefaction/water body) is not merged with other classes. (c) CBSI index technique has an edge over conventional index technique. (d) The NC classifier shows better result for soil liquefaction identification while PCM classifier generates better results for water body identification as com- pared with other classifiers. The advantage of using CBSI index technique is that the user does not require to know spectral information of remote sensing sensor. This methodology provides a quick and efficient way to identify soil liquefaction locations and may be useful for management and development plans for decision makers, local administrations and scientists interested in earthquake-induced soil liquefaction. Acknowledgements The authors are thankful to learned reviewers and the chief editor for their critical remarks, valuable guidance and comments for improving the manuscript. References Ambraseys NN. 1988. Engineering seismology. Earthq Eng Struct Dyn. 17:1–105. Baugh WM, Groeneveld DP. 2006. Broadband vegetation index performance evaluated for a low-cover environment. Int J Remote Sens. 27:4715–4730. Binaghi E, Brivio PA, Chessi P, Rampini A. 1999. A fuzzy set based accuracy assessment of soft classification. Pattern Recognit Lett. 20:935–948. Birth GS, McVey G. 1968. Measuring the color of growing turf with a reflectance spectroradi- ometer. Agron J. 60:640–643. Biswas SK. 1987. Regional tectonic framework, structure and evolution of the western mar- ginal basins of India. Tectonophysics. 135:307–327. Bezdek JC. 1981. Pattern recognition with fuzzy objective function algorithm. New York: Plenum. Brier GW, Allen RA. 1951. Verification of weather forecasts. In: Malone T, editor. Compen- dium of meteorology. Boston (MA): American Meteorological Society; p. 841–848. Broge NH, Leblanc E. 2000. Comparing predicting power and stability of broadband and hyperspectral vegetation indices for estimation of green area index and canopy chloro- phyll density. Remote Sens Environ. 76:156–172. Champati Ray PK, Foreste K, Roy PS. 2001. Detection and analysis of surface changes due to the recent earthquake in Gujarat using IRS-PAN and L ISS-III data. Proceedings of Workshop on Recent Earthquakes of Chamoli and Bhuj; Department of Earthquake Engineering, University of Roorkee, p. 225–236. Chen X, Vogelmann JE, Rollins M, Ohlen D, Key CH, Yang L, Huang C, Shi H. 2011. Detecting post-fire burn severity and vegetation recovery using multitemporal remote sensing spectral indices and field-collected composite burn index data in a ponderosa pine forest. Int J Remote Sens. 32:7905–7927. Soft classification approaches for identification of earthquake-induced liquefied soil 351 Dave RN. 1991. Characterization and detection of noise in clustering. Pattern Recognit Lett. 12:657–664. Foody GM, Lucas RM, Curran PJ, Honzak M. 1997. Non-linear mixture modeling without end-members using an artificial neural network. Int J Remote Sens. 18:937–953. Fung T, Siu W. 2000. Environmental quality and its changes, an analysis using NDVI. Int J Remote Sens. 21:1011–1024. Gao BC. 1996. NDWI-a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens Environ. 58:257–266. Gupta RP, Chander R, Tewari AK, Saraf AK. 1995. Remote sensing delineation of zones sus- ceptible to seismically induced liquefaction in the Ganga plains. J Geol Soc India. 46:75–82. Gupta RP, Saraf AK, Saxena P, Chander R. 1994. IRS detection of surface effects of the Uttarkashi earthquake of 20 October 1991, Himalaya. Int J Remote Sens. 15:2153–2156. Hardisky MA, Klemas V, Smart RM. 1983. The influences of soil salinity, growth form, and leaf moisture on the spectral reflectance of spartina alterniflora canopies. Photo- gramm. Eng. Remote Sens. 49:77–83. Hazarika H, Boominathan A. 2009. Liquefaction and ground failures during the 2001 Bhuj earthquake, India. In: Kokusho T, editor. Earthquake geotechnical case histories for performance-based design. London: Taylor & Francis Group; p. 201–225. Holzer TL. 1989. Dynamics of liquefaction during the 1987 Superstition Hills, California, earthquake. Science. 244:56–59. Huete AR. 1988. A soil-adjusted vegetation index (SAVI). Remote Sens Environ. 25:295–309. Ishihara K, Shimuzu K, Yamada Y. 1981. Pore water pressures measured in sand deposits dur- ing an earthquake. Soils Found (Japan). 21:85–100. Jordan CF. 1969. Derivation of leaf area index from quality of light on the forest floor. Ecology. 50:663–666. Ju J, Kolaczyk ED, Gopal S. 2003. Gaussian mixture discriminant analysis and sub-pixel land cover characterization in remote sensing. Remote Sens Environ. 84:550–560. Klir GJ. 1990. A principle of uncertainty and information variance. Int J Gen Syst. 17: 249–275. Krinitzsky EL, Gould JP, Edinger PH. 1993. Fundamental of earthquake resistant construc- tion. New York: Wiley. Krishnapuram R, Keller JM. 1993. A possiblistic approach to clustering. IEEE Trans Fuzzy Syst. 1:98–108. Liou CP, Richart FE, Streeter VL. 1977. Numerical Model for liquefaction. J Geotech Eng ASCE. 103:589–606. Lyon JG, Yaun D, Lunetta RS, Eldridge CD. 1997. A change detection experiment using vege- tation indices. Photogramm Eng Remote Sens. 64:143–150. Markham BL, Barker IJ. 1986. Landsat MSS andTMpost-calibration dynamic ranges, exoat- mospheric reflectances and at-satellite temperatures. EOSAT Tech Notes. 1:3–8. Miyamoto S, Ichihashi H, Honda K. 2008. Algorithms for fuzzy clustering: methods in c-means clustering with applications. Berlin: Springer; p. 46–66. Papadopoulos GA, Lefkopoulos G. 1993. Magnitude distance relations for liquefaction in soil from earthquakes. Bull Seismol Soc Am. 83:925–938. Rajendran K, Rajendran CP, Thakkar M, Tuttle M. 2001. The 2001 Kutch (Bhuj) earthquake: coseismic surface features and their significance. Curr Sci. 80:1397–1405. Ramakrishnan D, Mohanty KK, Nayak SR. 2006. Mapping the liquefaction induced soil moisture changes using remote sensing technique: an attempt to map the earthquake induced liquefaction around Bhuj, Gujarat, India. Geotech Geol Eng. 24:1581–1602. Rao KS, Sakalley J, Rao YS, Kannan J, Pani BS. 2001. Study of Bhuj earthquake – a remote sensing approach. In: Proceedings of the Workshop on Recent Earthquake of Chamoli and Bhuj. Roorkee, India: Indian Society of Earthquake Technology; p. 217–224. 352 S.S. Sengar et al. Roeloffs EA. 1998. Persistent water level changes in a well near Parkfield, California, due to local and distant earthquakes. J Geophys Res. 103:869–889. Rouse JW, Haas RH, Schell JA, Deering DW. 1973. Monitoring vegetation systems in the Great Plains with ERTS. In: Third ERTS Symposium, NASA SP-351-I, p. 309–317. Sarkar I, Chander R. 2003. A simulation of earthquake induced undrained pore pressure changes with bearing on some soil liquefaction observations following the 2001 Bhuj earthquake. Proc Indian Acad Sci (Earth Planet Sci). 112:471–477. Seed HB. 1970. Soil problems and soil behavior. In: Wiegel RL, editor. Earthquake engineer- ing. Englewood Cliffs (NJ): Prentice-Hall; p. 227–251. Sellers PJ, Meeson BW, Hall FG, Asrar, G., Murphy RE, Schiffer RA, Bretherton FP. 1995. Remote sensing of the land surface for studies of global change: models, algorithms- experiments. Remote Sens Environ. 51:3–26. Sengar SS, Kumar A, Ghosh SK, Wason HR. 2013. Earthquake induced built-up damage identification using IRS-P6 data – a comparative study using fuzzy based classifiers. Geocarto Int. [DOI:http://dx.doi.org/10.1080/10106049.2012.756939]. Sinha S, Mohanty S. 2012. Spatial variation of crustal strain in the Kachchh region, India: implication on the Bhuj earthquake of 2001. J Geodyn. 61:1–11. Singh RP, Bhoi S., Sahoo AK, Raj U., Ravindranath S. 2001. Surface manifestations after the Gujarat earthquake. Curr Sci. 81:338–340. Thakkar MG, Goyal B. 2004. On Relation between magnitude and liquefaction dimension at the epicentral zone of 2001 Bhuj earthquake. Curr Sci. 87:811–817. Tucker CJ. 1979. Red and photographic infrared linear combinations for monitoring vegeta- tion. Remote Sens Environ. 8:127–150. Wang J, Price KP, Rich PM. 2001. Spatial patterns of NDVI in response to precipitation and temperature in the Central Great Plains. Int J Remote Sens. 22:3827–3844. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png "Geomatics, Natural Hazards and Risk" Taylor & Francis

Study of soft classification approaches for identification of earthquake-induced liquefied soil

Download PDF
19 pages

Loading next page...
 
/lp/taylor-francis/study-of-soft-classification-approaches-for-identification-of-v6meUx1l5Y

References (47)

Publisher
Taylor & Francis
Copyright
© 2013 Taylor & Francis
ISSN
1947-5713
eISSN
1947-5705
DOI
10.1080/19475705.2013.811444
Publisher site
See Article on Publisher Site

Abstract

Geomatics, Natural Hazards and Risk, 2014 Vol. 5, No. 4, 334–352, http://dx.doi.org/10.1080/19475705.2013.811444 Study of soft classification approaches for identification of earthquake-induced liquefied soil SANDEEP SINGH SENGAR*y, ANIL KUMARz, HANS RAJ WASONy, SANJAY KUMAR GHOSHx, Y.V.N. KRISHNA MURTHYz and P.L.N. RAJUz Earthquake Engineering Department, Indian Institute of Technology Roorkee, Roorkee 247667, India Indian Institute of Remote Sensing, Indian Space Research Organization, Dehradun 248001, India Civil Engineering Department, Indian Institute of Technology Roorkee, Roorkee 247667, India (Received 14 December 2012; in final form 29 May 2013) The existence of mixed pixels led to the development of several approaches for soft (or fuzzy) classification in which each pixel is allocated to all classes in varying proportions. However, while the proportions of each land cover within each pixel may be predicted, the spatial location of each land cover within each pixel is not. There exist many different potential techniques for sub-pixel mapping from remotely sensed imagery to identify specific class. The fuzzy-based possibilistic c-means (PCM), noise cluster (NC) and noise cluster with entropy (NCE) classifiers were applied to identify the Bhuj, India (2001), earthquake induced soil liquefaction and compared as soft computing approaches via supervised classifica- tion. The soil liquefaction identification was empirically investigated and com- pared with class-based sensor-independent (CBSI) spectral band ratio using Landsat-7 temporal images. It has been found that CBSI-based temporal indices yield the better results for identification of liquefied soil areas while it was easily separated with pre-earthquake existing water body in that area. The NCE classi- fier performed better for conventional temporal indices, while NC classifier per- formed better for soil liquefaction and PCM classifier performed better for water body identification with CBSI temporal indices. 1. Introduction The last two decades have witnessed the increasing use of remote sensing for under- standing the geophysical phenomena underlying natural hazards. By acquiring temporal data, i.e. both pre- and post-event satellite images, it is possible to identify earthquake-induced soil liquefaction with relative ease and effectiveness by incorpo- rating temporal spectral changes. Gupta et al. (1994, 1995) successfully demon- strated the application of remote sensing techniques to delineate zones of seismically induced soil liquefaction in the Ganga plains. Ramakrishnan et al. (2006) mapped the earthquake-induced soil liquefaction around Bhuj by calculating the absorption of energy in the near-infrared (NIR) and short wave infrared region of electromag- netic spectrum. Champati Ray et al. (2001) and Rao et al. (2001) demonstrated *Corresponding author. Email: talktosengar@yahoo.co.in 2013 Taylor & Francis Soft classification approaches for identification of earthquake-induced liquefied soil 335 applications of principal component analysis and unsupervised image classification techniques, respectively, to change detection induced by the 2001 Bhuj earthquake. Singh et al. (2001) represents the surface manifestations due to Bhuj earthquake by band rationing IRS-1D data. However, the most promising area seems to be the application of soft classification fuzzy approach, not widely used in the field of such events. For this reason, it has been proposed that fuzziness should be accommodated in the classification proce- dure so that pixels may have multiple or partial class membership and handle mixed pixel (Foody et al. 1997). The work has been designed to identify soil liquefaction without merging it with existing water body using coarse-resolution remotely sensed data. To achieve these goals, three classifiers (PCM, NC and NCE) were tested with class-based sensor-independent (CBSI) technique to improve the classifiers output for specific class identification. This paper is structured with introductory part provides a review of the recent studies on the sub-pixel classification of remote sensing data using spectral indices. Then, the 2001 Bhuj earthquake induced soil liquefaction case studies were discussed. After that the proposed methodology will be discussed to identify soil liquefaction at sub-pixel level. Then, results and details on image interpretation and analysis are described. From the results, it has also been tried to explain that the liquefied area clearly discriminated with pre-earthquake existing water body. At the end, conclud- ing benefits were presented about the proposed methodology. 2. Sub-pixel classification To study various aspects related to global change studies and environmental applica- tions, land cover information is one of the crucial data components (Sellers et al. 1995). Multi-spectral-based classification is one of the major approaches for extract- ing land cover information from remotely sensed images. While extracting land cover from remote sensing images, each pixel in the image is allocated to one of the possible classes. In reality, different land covers within a pixel can be found due to continuum of variation in landscape and intrinsic mixed nature of most classes (Ju et al. 2003). Mixed pixels may not be appropriately processed by hard classifiers, which assume that pixels are pure. Mixed pixels may be problem in mapping land cover and it may be more severe in heterogeneous classes from coarse spatial resolution images. Thus, it is important to implement different soft classifiers for handling mixed pixels in an image. The objective of classification is to classify each pixel into one and only one land cover class (i.e. hard classification) or to estimate the partial membership of the clas- ses in a pixel (i.e. soft classification). In fuzzy set theory, an unsupervised approach, fuzzy c-means (FCM) clustering (Bezdek 1981), has been used for classifying remote sensing data. In FCM, the summation of class memberships of a pixel is equal to 1, which is based on a probabilistic constraint. The major drawback of this constraint is that classes present in a pixel are inter-dependent, which may not be the case in reality. In possibilistic c-means (PCM) clustering (Krishnapuram & Keller 1993), probabilistic constraint has been relaxed to produce absolute class memberships, which may indicate the class proportions in a pixel. Most clustering methods are plagued with the problem of noisy data, i.e. characterization of good clusters amongst noisy data. The noise that is just due to the statistical distribution of the 336 S.S. Sengar et al. measuring instruments is usually of no concern. On the other hand, the completely arbitrary noise points that just do not belong to the pattern or class being searched for are of real concern. Dave (1991) introduced concept of ‘noise cluster’ (NC) with ability to detect ‘good’ clusters amongst noisy data. Another approach adopted in this study was the use of spectral indices to generate temporal spectral index data for classification. The temporal index database incorpo- rates temporal change information of a class in classification procesure. It has been found that the spectral response varies for different objects at distinctive electromag- netic regions. This inherent phenomenon can be exploited by performing certain well-defined arithmetical operations to generate new digital data set to highlight the specific phenomenon, such as presence of vegetation or moisture. These procedures are known as spectral indices. Several researchers have evaluated the use of spectral indices in various applications ranging from vegetation, environmental change detec- tion and water resource analysis to mineral, snow and ice studies, post-fire burn detection (Jordan 1969; Rouse et al. 1973; Tucker 1979; Gao 1996; Lyon et al. 1997; Fung & Siu 2000; Wang et al. 2001; Baugh & Groeneveld 2006; Chen et al. 2011). Literature reveals that to identify liquefied area, spectral indices with PCM, NC and noise cluster with entropy (NCE) classifiers have not been used earlier. One of the advantages of using such classifiers is that the specific class of interest (in this study liquefied area) can be identified without merging it with other classes at sub- pixel level. These classifiers require minimal reference data as training sample and less information about image characteristic for identification of disaster damage using coarse-resolution remotely sensed data. 3. Factors responsible for soil liquefaction in study area and data used Soil liquefaction occurs when the soil is saturated, i.e. the space between individual par- ticles is completely filled with water, thus causing a pressure on the soil particles. Prior to an earthquake, the water pressure is relatively low. However, an earthquake may cause the water pressure to increase to such a level that the soil particles may move with respect to each other thus leading to the release of soil moisture. In this process, the soil loses its strength and behaves like a liquid (Liou et al. 1977; Ambraseys 1988). A combination of different geological, seismological and hydrologi- cal factors could be responsible for soil liquefaction induced by the Bhuj earthquake. The study area (figure 1) that was highly affected by soil liquefaction comprises of mainly Mesozoic (sandstones, siltstones, shale and limestone), Tertiary (poorly consoli- dated sandstone, siltstone and clay) and Quaternary (sand and clay) sedimentary sequences (Biswas 1987;Sinha &Mohanty 2012). The Kachchh seismo-tectonic belt extends approximately 250 km (east–west) and 150 km (north–south). It is flanked in the north by Nagar Parkar fault and in the south by Kathiawar fault. The area in between is traversed by several faults/fault systems. Salient among them are Katrol Hill Fault, Kachchh Mainland Fault, Banni Fault, Island Belt Fault and Allah Bund Fault. Rajendran et al. (2001) correlated the cause of soil liquefaction in Kachchh area with the past earthquakes. Thakkar and Goyal (2004) correlated extent of soil lique- faction with function of epicentral distance, magnitude of the earthquake, depth of the hypocentre, availability of the source material and also proximity to the major lineaments due to Bhuj earthquake. Hazarika and Boominathan (2009) observed that 100% saturated soil of Kachchh region is responsible for the soil liquefaction. Soft classification approaches for identification of earthquake-induced liquefied soil 337 Figure 1. Study area Kachchh, Gujarat, India (modified from Biswas 1987). Several soil and water table related factors are also responsible for the site-dependent characteristics of such earthquake-induced soil liquefaction processes, which are as follows (Sarkar & Chander 2003): (a) the nature of the sediments, the grain size and intragranular packing (Seed 1970; Holzer 1989); (b) earthquake hypocentral distance (Papadopoulos & Lefkopoulos 1993); (c) depth of the water table (Seed 1970; Roeloffs 1998); (d) depth of deposit (data available mainly down to a depth of 15 m Krinitzsky et al. 1993); (e) nature and thickness of the cap rock; (f) degree of water saturation of the soil (Seed 1970; Roeloffs 1998); (g) earthquake-induced changes in the pore water pressure (Ishihara et al. 1981); (h) regional topography (Seed 1970). Sarkar and Chander (2003) attributed the reasons of soil liquefaction to (a) shallow water table (less than 15 m) or (b) well-sealed clayey soil overlying thick liquefiable sediments, causing the released water to persist on the earth surface for several days after the main shock. Test data sets for this work have been acquired from ETM sensor of Landsat-7 satellite over two different acquisition dates, i.e. 8 January 2001 and 9 February, 2001. These pre- and post-earthquake remote sensing images of the regions, where such earthquake-induced phenomena occurred, help in systematic monitoring of possible patterns to identify soil liquefaction. 338 S.S. Sengar et al. 4. Methodology adopted There has been requirement for identifying only one class that is, soil liquefaction, in Kachchh area. The adopted methodology for this work has been explained with the help of flow diagram as shown in figure 2. After pre-processing stage, with the help of CBSI (as explained in equation (1)) and conventional index techniques, temporal (two date, i.e. pre- and post-earthquake) index database has been prepared (table 1). Figure 2. Flowchart of methodology adopted. Soft classification approaches for identification of earthquake-induced liquefied soil 339 Table 1. Mean band values from temporal indices data. Conventional indices CBSI indices Soil liquefaction Water body Soil liquefaction Water body Spectral indices Pre-EQ Post-EQ Pre-EQ Post-EQ Pre-EQ Post-EQ Pre-EQ Post-EQ MNDWI 0.388 0.854 0.811 0.901 0.000 0.803 0.674 0.827 (Markham & Barker 1986) NDVI (Rouse 0.454 0.482 0.172 0.188 0.000 0.803 0.674 0.827 et al. 1973) NDWI (Hardisky 0.435 0.874 0.658 0.764 0.000 0.803 0.674 0.827 et al. 1983) SAVI (Huete 0.454 0.482 0.172 0.188 0.000 0.800 0.670 0.823 1988) SR (Birth & 0.149 0.145 0.066 0.062 0.211 0.031 0.035 0.027 McVey 1968) TNDVI (Tucker 0.768 0.772 0.619 0.603 0.000 0.949 0.913 0.956 1979) TVI (Broge & 0.250 0.439 0.317 0.372 0.529 0.623 0.733 0.725 Leblanc 2000) Using these indices, both pre- and post-earthquake index maps were generated. Thus, creating a temporal index map, these data sets have been used as inputs for classification: g ðr; cÞ nir CBSIðr; cÞ¼ ; 8ðr; cÞ2 C ; ð1Þ g ðr; cÞ red where g is the digital number (DN) value of a class, r and c are the row and column of pixels belonging to a class, C is kth class of the pixels and g is the band for which k nir the maximum grey value is found within the pixels belonging to the considered class; analogously, g is the band of the minimum value. red In equation (1), the analyst provides the location of a class in the form of row and column or latitude and longitude. Based on coordinates of a class input, the mini- mum and maximum values of the data are computed. Apart from the location of the class of interest (liquefied area), spectral information as grey value from all the bands will be read. Using minimum and maximum operators, find out which band has min- imum and maximum grey values. The band having the maximum value is designated as NIR and other band having minimum value is designated as RED in different indices as mentioned in table 1. If any index does not have RED and NIR bands (e.g. modified normalized difference water index (MNDWI) and normalized differ- ence water index (NDWI)) then band values replaced by these maximum and mini- mum DN value to achieve maximum index value. If in any pixel, the value of the indices is negative, then replace it with zero values. This enhances the class of interest and requires only the geolocation of a class without any spectral band information of the sensor. 340 S.S. Sengar et al. The objective functions of fuzzy-based classifiers applied in this work may be for- mulated (Miyamoto et al. 2008; Sengar et al. 2013) to identify soil liquefaction at sub-pixel level from equations (2)–(4): "# c N c N XX X X m m ðiÞ PCM ¼ J ðU; VÞ¼ ðm Þ Dðx ; v Þþ h ð1-m Þ ; ð2Þ pcm k i k;i i k;i i¼1 k¼1 i¼1 k¼1 "# c N N XX X m m ðiiÞ NC ¼ J ðU; VÞ¼ ðm Þ Dðx ; v Þþ ðm Þ d ; ð3Þ nc k;i k i k;cþ1 i¼1 k¼1 k¼1 "# c N N N cþ1 XX X XX ðiiiÞ NCE ¼ J ðU;VÞ¼ ðm ÞDðx ;v Þþ ðm Þd þ n m logðm Þ : nce k i ki k;cþ1 k;i k;i i¼1 k¼1 k¼1 k¼1 i¼1 ð4Þ where m is the class membership values of a pixel k belonging to i, Dðx ;v Þ is the k i k;i squared distance in feature space between x and v , Dðx ;v Þ is the squared distance k i k j in feature space between x and v , x is the vector denoting spectral response of a k j k pixel k, v is the collection of vector of cluster centres of class i, v is the collection of i j vector of cluster centres of class j, U is the N c membership matrix, V is the mean vector for class i, c and N are the total number of clusters and pixels, respectively, m is the weighted constant (1< m < 1), c þ 1 is the extra class in the data after con- sidering total number of classes c, d is the resolution parameter and have any float value greater than zero and n is the regularizing parameter. In digital image classification, reference data play an important role, both at train- ing and testing stage of supervised image classification. The training data should be of value, if the environment from which they were obtained is relatively homogenous. However, the ground features are heterogeneous in nature, and many a time, a pixel covers two or more classes due to which remote sensing images are dominated by mixed pixels. As a result, the land cover classes are generally mixed in nature and inter-grade gradually in area. While preparing land cover mapping at regional and global level, coarse-resolution images are used, thereby the chances of occurrence of mixed pixels are high. Shannon entropy and fuzzy-set-based measures such as an index of fuzziness (Binaghi et al. 1999) may be used to estimate the uncertainty in the classification data. Entropy measures show how the strength of class membership in the classification output is partitioned between the classes for each pixel. The value of these measures is maximized (a high degree of uncertainty) when the class membership is partitioned evenly between all the classes, and minimized (a low degree or uncertainty) when the membership is associated entirely with one class. Shannon entropy, a measure conceptualized in terms of probability theory may be computed from equation (5) (Klir 1990): H ¼ m log ðm Þ: ð5Þ k;i 2 k;i i¼1 Soft classification approaches for identification of earthquake-induced liquefied soil 341 But two dimensions are required to show the accuracy of output result, as it arises on a given set of testing data. One is true positive ratio (TPR) and another one is false alarm ratio (FAR). Low FAR (equation (7)) and high TPR (equation (6)) is desirable for good result (Brier & Allen 1951): Number of target pixels correctly detected TPR ¼ ; ð6Þ Total number of target pixels present in the sample Number of background pixels detected as target FAR ¼ : ð7Þ Total number of background pixels presents in the sample 5. Results and discussion The various indices as listed in table 1 provides spectral information of specific class for those temporal spectral indices. The variation in mean indices with the temporal data (two date data) shows the sensitivity of those indices for the specific class. High variation in mean index value (between pre- and post-event) with temporal data means, the changes in specific class with temporal data clearly identified. Different indices help to distinguish reflectance of soils by its moisture content. If this temporal spectral index variation is more from a specific index, then more clearly soil liquefac- tion can be distinguished with water bodies. As in the case of water, the index value has not shown much variation on pre- and post-images because it was not changed due to the earthquake. As in CBSI technique, the band values of conventional indices are replaced by maximum and minimum grey values for the specific class identification to obtain maximum index value. Therefore, conventional-technique-based normalized differ- ence vegetation index (NDVI), MNDWI and NDWI are behaving same if used as CBSI technique because all these indices are using difference and addition of band values in numerator and denominator, respectively. In PCM classifier, the m value is class dependent and has relationship with class membership (equation (2)); therefore, the outputs generated have relationship with constant value ‘m’ and spectral indices as shown in figure 3. The entropy of any class is directly related with membership of the class (equation (5)); therefore, the ‘m’ value for each spectral index should be selected such that the specific class shows minimum entropy (0.005) as given in table 2. The point selected for ‘m’ values as a corner frequency of membership values between 0.996 and 0.992. The selection of optimal value of ‘m’ ensures that the specific class of interest can eas- ily be identified without being overlapped with other classes or background. Simi- larly, less entropy indicates that there is less uncertainty in the identification of the class of interest. Table 3 and figure 4 have been generated by keeping ‘m’ as mentioned in table 2 to find constant value of ‘d’ for NC classifier (equation (3)) for each spectral index. These constant values (‘m’and ‘d’) have been utilized to generate results for NC classifier. The NCE classifier depends on ‘d’and ‘n’ values as in this case we have taken m ¼ 1 (Miyamoto et al. 2008). The d and n values are not class dependent; therefore, these 342 S.S. Sengar et al. Table 2. Optimum values of ‘m’ for PCM classifier. Conventional spectral indices CBSI spectral indices Spectral indices Soil liquefaction Water body Soil liquefaction Water body MNDWI 2.3 2.2 – – NDVI 2.0 2.2 2.3 2.3 NDWI 2.3 1.8 – – SAVI 2.5 2.3 2.2 2.2 SR 2.4 2.5 2.6 1.9 TNDVI 2.5 2.2 2.8 2.7 TVI 2.0 1.5 2.1 2.3 Table 3. Optimum values of ‘d’ for NC classifier. Conventional spectral indices CBSI spectral indices Spectral indices Soil liquefaction Water body Soil liquefaction Water body 4 4 MNDWI 3  10 5  10 –– 3 4 4 5 NDVI 3  10 3  10 5  10 2  10 4 4 NDWI 2  10 2  10 –– 3 4 4 5 SAVI 2  10 3  10 3  10 2  10 2 3 3 3 SR 6  10 4  10 4  10 4  10 2 3 4 5 TNDVI 5  10 8  10 3  10 2  10 3 3 3 4 TVI 2  10 2  10 3  10 2  10 Figure 3. Plots showing optimum value of ‘m’ for PCM classifier. Soft classification approaches for identification of earthquake-induced liquefied soil 343 Figure 4. Plots showing optimum value of ‘d’ for NC classifier. constants vary as shown in figure 5. The constant values for NCE classifier have been selected (d ¼ 8000 and n ¼ 1400) so that the specific class shows minimum entropy (0.005). The classified outputs have been shown in figures 6 and 7 using conventional index approach and in figures 8 and 9 using CBSI approach, with the aim to obtain a better Figure 5. Plot showing optimum values of ‘n’ and ‘d’ for NCE classifier. 344 S.S. Sengar et al. Figure 6. Identification of soil liquefaction using temporal conventional indices data. localization and discrimination of two types (soil liquefaction and water body). Table 4 shows the minimum membership values (up to which specific class not merged with background) for each index. In these images, both water body and liq- uefied areas have high membership values and are seen as bright areas. However, Soft classification approaches for identification of earthquake-induced liquefied soil 345 Figure 7. Identification of water body using temporal conventional indices data. 346 S.S. Sengar et al. Figure 8. Identification of soil liquefaction using temporal CBSI indices data. Soft classification approaches for identification of earthquake-induced liquefied soil 347 Figure 9. Identification of water body using temporal CBSI indices data. 348 S.S. Sengar et al. there are variations in the water body and liquefied areas on different indices. The results of the same have superim- posed on the corresponding water body and soil liquefaction areas on index images. In these results, the memberships for representative feature points keeps as high as possible, while unrepresentative points should have low membership in all clusters. These figures are showing the same areas for identification of soil lique- faction as well as water body and clearly represent that one class (soil liquefac- tion) not merged with other class (water body). The output result shows that all CBSI and conventional temporal indices gener- ated good classification results for soil liquefaction as well as water body except TVI index. This is due to the one class mixed with other classes in TVI indices, while there is a significant change observed in CBSI-TVI as compared with conventional TVI. The output evaluation of the results generated by all classifiers has been carried out with the help of FAR and TPR values (tables 5–7). It has been observed from conven- tional index results that TNDVI and SAVI generated good output for soil liquefaction identification with low FAR and high TPR using all classifiers. Among all conventional spectral index results, MNDWI and NDWI generate best output with FAR ¼ 0.26 and TPR ¼ 0.84 using NCE classifier, while using CBSI index technique for soil liq- uefaction identification TNDVI (NC) with FAR ¼ 0.10 and TPR 0.93. For water body identification SR (NCE) generates best results with FAR ¼ 0.16 and TPR ¼ 0.87 using conventional spectral indices. The CBSI-based NDVI (PCM) generated best output for water body identification with FAR ¼ 0.10 and TPR ¼ 0.93. The results were improved by using CBSI technique and clearly reflected in figures 6–9 and tables 5–7. Table 4. Minimum membership value for PCM, NC and NCE classifier. PCM classifier NC classifier NCE classifier Conventional indices CBSI indices Conventional indices CBSI indices Conventional indices CBSI indices Spectral Soil Water Soil Water Soil Water Soil Water Soil Water Soil Water index liquefaction body liquefaction body liquefaction body liquefaction body liquefaction body liquefaction body MNDWI 0.882 0.941 – – 0.901 0.941 – – 0.980 0.992 – – NDVI 0.992 0.784 0.882 0.909 0.992 0.803 0.882 0.941 0.980 0.988 0.980 0.980 NDWI 0.882 0.992 – – 0.894 0.992 – – 0.980 0.980 – – SAVI 0.937 0.784 0.882 0.921 0.941 0.784 0.882 0.952 0.972 0.980 0.972 0.980 SR 0.968 0.784 0.686 0.901 0.968 0.843 0.745 0.874 0.984 0.988 0.960 0.988 TNDVI 0.933 0.784 0.901 0.960 0.921 0.764 0.862 0.960 0.980 0.992 0.980 0.980 TVI 0.835 0.996 0.941 0.901 0.803 0.996 0.956 0.933 0.992 0.992 0.980 0.980 Soft classification approaches for identification of earthquake-induced liquefied soil 349 Table 5. FAR and TPR values for PCM classifier outputs. Soil liquefaction Water body Conventional CBSI Conventional CBSI spectral indices spectral indices spectral indices spectral indices Soil liquefaction FAR TPR FAR TPR FAR TPR FAR TPR MNDWI 0.27 0.69 – – 0.17 0.85 – – NDVI 0.20 0.62 0.14 0.87 0.15 0.87 0.10 0.93 NDWI 0.25 0.67 – – 0.56 0.38 – – SAVI 0.35 0.87 0.13 0.88 0.16 0.82 0.11 0.87 SR 0.35 0.87 0.12 0.89 0.14 0.86 0.10 0.76 TNDVI 0.35 0.87 0.11 0.86 0.15 0.85 0.24 0.87 TVI 0.66 0.55 0.85 0.34 0.85 0.34 0.38 0.40 Table 6. FAR and TPR values for NC classifier outputs. Soil liquefaction Water body Conventional CBSI Conventional CBSI spectral indices spectral indices spectral indices spectral indices Spectral indices FAR TPR FAR TPR FAR TPR FAR TPR MNDWI 0.20 0.64 – – 0.16 0.67 – – NDVI 0.45 0.60 0.16 0.89 0.16 0.82 0.12 0.86 NDWI 0.25 0.60 – – 0.57 0.47 – – SAVI 0.30 0.79 0.14 0.89 0.14 0.83 0.12 0.88 SR 0.33 0.84 0.09 0.78 0.15 0.84 0.13 0.64 TNDVI 0.34 0.88 0.10 0.93 0.13 0.83 0.25 0.88 TVI 0.62 0.51 0.54 0.47 0.84 0.38 0.41 0.58 Table 7. FAR and TPR values for NCE classifier outputs. Soil liquefaction Water body Conventional CBSI Conventional CBSI spectral indices spectral indices spectral indices spectral indices Spectral indices FAR TPR FAR TPR FAR TPR FAR TPR MNDWI 0.26 0.84 – – 0.12 0.80 – – NDVI 0.20 0.67 0.15 0.88 0.17 0.78 0.09 0.85 NDWI 0.26 0.84 – – 0.51 0.79 – – SAVI 0.35 0.82 0.14 0.89 0.13 0.82 0.12 0.77 SR 0.48 0.76 0.38 0.71 0.16 0.87 0.12 0.80 TNDVI 0.32 0.82 0.12 0.85 0.13 0.82 0.45 0.90 TVI 0.64 0.35 0.74 0.45 0.54 0.35 0.47 0.48 350 S.S. Sengar et al. 6. Conclusion This work tries to study importance of soft classification fuzzy techniques such as PCM, NC and NCE for identification of earthquake-induced liquefied soil as specific class; from the results, following conclusions can be drawn. (a) CBSI approach generates better discrimination between liquefied areas and water body as compared with conventional indices. (b) One class of interest (soil liquefaction/water body) is not merged with other classes. (c) CBSI index technique has an edge over conventional index technique. (d) The NC classifier shows better result for soil liquefaction identification while PCM classifier generates better results for water body identification as com- pared with other classifiers. The advantage of using CBSI index technique is that the user does not require to know spectral information of remote sensing sensor. This methodology provides a quick and efficient way to identify soil liquefaction locations and may be useful for management and development plans for decision makers, local administrations and scientists interested in earthquake-induced soil liquefaction. Acknowledgements The authors are thankful to learned reviewers and the chief editor for their critical remarks, valuable guidance and comments for improving the manuscript. References Ambraseys NN. 1988. Engineering seismology. Earthq Eng Struct Dyn. 17:1–105. Baugh WM, Groeneveld DP. 2006. Broadband vegetation index performance evaluated for a low-cover environment. Int J Remote Sens. 27:4715–4730. Binaghi E, Brivio PA, Chessi P, Rampini A. 1999. A fuzzy set based accuracy assessment of soft classification. Pattern Recognit Lett. 20:935–948. Birth GS, McVey G. 1968. Measuring the color of growing turf with a reflectance spectroradi- ometer. Agron J. 60:640–643. Biswas SK. 1987. Regional tectonic framework, structure and evolution of the western mar- ginal basins of India. Tectonophysics. 135:307–327. Bezdek JC. 1981. Pattern recognition with fuzzy objective function algorithm. New York: Plenum. Brier GW, Allen RA. 1951. Verification of weather forecasts. In: Malone T, editor. Compen- dium of meteorology. Boston (MA): American Meteorological Society; p. 841–848. Broge NH, Leblanc E. 2000. Comparing predicting power and stability of broadband and hyperspectral vegetation indices for estimation of green area index and canopy chloro- phyll density. Remote Sens Environ. 76:156–172. Champati Ray PK, Foreste K, Roy PS. 2001. Detection and analysis of surface changes due to the recent earthquake in Gujarat using IRS-PAN and L ISS-III data. Proceedings of Workshop on Recent Earthquakes of Chamoli and Bhuj; Department of Earthquake Engineering, University of Roorkee, p. 225–236. Chen X, Vogelmann JE, Rollins M, Ohlen D, Key CH, Yang L, Huang C, Shi H. 2011. Detecting post-fire burn severity and vegetation recovery using multitemporal remote sensing spectral indices and field-collected composite burn index data in a ponderosa pine forest. Int J Remote Sens. 32:7905–7927. Soft classification approaches for identification of earthquake-induced liquefied soil 351 Dave RN. 1991. Characterization and detection of noise in clustering. Pattern Recognit Lett. 12:657–664. Foody GM, Lucas RM, Curran PJ, Honzak M. 1997. Non-linear mixture modeling without end-members using an artificial neural network. Int J Remote Sens. 18:937–953. Fung T, Siu W. 2000. Environmental quality and its changes, an analysis using NDVI. Int J Remote Sens. 21:1011–1024. Gao BC. 1996. NDWI-a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens Environ. 58:257–266. Gupta RP, Chander R, Tewari AK, Saraf AK. 1995. Remote sensing delineation of zones sus- ceptible to seismically induced liquefaction in the Ganga plains. J Geol Soc India. 46:75–82. Gupta RP, Saraf AK, Saxena P, Chander R. 1994. IRS detection of surface effects of the Uttarkashi earthquake of 20 October 1991, Himalaya. Int J Remote Sens. 15:2153–2156. Hardisky MA, Klemas V, Smart RM. 1983. The influences of soil salinity, growth form, and leaf moisture on the spectral reflectance of spartina alterniflora canopies. Photo- gramm. Eng. Remote Sens. 49:77–83. Hazarika H, Boominathan A. 2009. Liquefaction and ground failures during the 2001 Bhuj earthquake, India. In: Kokusho T, editor. Earthquake geotechnical case histories for performance-based design. London: Taylor & Francis Group; p. 201–225. Holzer TL. 1989. Dynamics of liquefaction during the 1987 Superstition Hills, California, earthquake. Science. 244:56–59. Huete AR. 1988. A soil-adjusted vegetation index (SAVI). Remote Sens Environ. 25:295–309. Ishihara K, Shimuzu K, Yamada Y. 1981. Pore water pressures measured in sand deposits dur- ing an earthquake. Soils Found (Japan). 21:85–100. Jordan CF. 1969. Derivation of leaf area index from quality of light on the forest floor. Ecology. 50:663–666. Ju J, Kolaczyk ED, Gopal S. 2003. Gaussian mixture discriminant analysis and sub-pixel land cover characterization in remote sensing. Remote Sens Environ. 84:550–560. Klir GJ. 1990. A principle of uncertainty and information variance. Int J Gen Syst. 17: 249–275. Krinitzsky EL, Gould JP, Edinger PH. 1993. Fundamental of earthquake resistant construc- tion. New York: Wiley. Krishnapuram R, Keller JM. 1993. A possiblistic approach to clustering. IEEE Trans Fuzzy Syst. 1:98–108. Liou CP, Richart FE, Streeter VL. 1977. Numerical Model for liquefaction. J Geotech Eng ASCE. 103:589–606. Lyon JG, Yaun D, Lunetta RS, Eldridge CD. 1997. A change detection experiment using vege- tation indices. Photogramm Eng Remote Sens. 64:143–150. Markham BL, Barker IJ. 1986. Landsat MSS andTMpost-calibration dynamic ranges, exoat- mospheric reflectances and at-satellite temperatures. EOSAT Tech Notes. 1:3–8. Miyamoto S, Ichihashi H, Honda K. 2008. Algorithms for fuzzy clustering: methods in c-means clustering with applications. Berlin: Springer; p. 46–66. Papadopoulos GA, Lefkopoulos G. 1993. Magnitude distance relations for liquefaction in soil from earthquakes. Bull Seismol Soc Am. 83:925–938. Rajendran K, Rajendran CP, Thakkar M, Tuttle M. 2001. The 2001 Kutch (Bhuj) earthquake: coseismic surface features and their significance. Curr Sci. 80:1397–1405. Ramakrishnan D, Mohanty KK, Nayak SR. 2006. Mapping the liquefaction induced soil moisture changes using remote sensing technique: an attempt to map the earthquake induced liquefaction around Bhuj, Gujarat, India. Geotech Geol Eng. 24:1581–1602. Rao KS, Sakalley J, Rao YS, Kannan J, Pani BS. 2001. Study of Bhuj earthquake – a remote sensing approach. In: Proceedings of the Workshop on Recent Earthquake of Chamoli and Bhuj. Roorkee, India: Indian Society of Earthquake Technology; p. 217–224. 352 S.S. Sengar et al. Roeloffs EA. 1998. Persistent water level changes in a well near Parkfield, California, due to local and distant earthquakes. J Geophys Res. 103:869–889. Rouse JW, Haas RH, Schell JA, Deering DW. 1973. Monitoring vegetation systems in the Great Plains with ERTS. In: Third ERTS Symposium, NASA SP-351-I, p. 309–317. Sarkar I, Chander R. 2003. A simulation of earthquake induced undrained pore pressure changes with bearing on some soil liquefaction observations following the 2001 Bhuj earthquake. Proc Indian Acad Sci (Earth Planet Sci). 112:471–477. Seed HB. 1970. Soil problems and soil behavior. In: Wiegel RL, editor. Earthquake engineer- ing. Englewood Cliffs (NJ): Prentice-Hall; p. 227–251. Sellers PJ, Meeson BW, Hall FG, Asrar, G., Murphy RE, Schiffer RA, Bretherton FP. 1995. Remote sensing of the land surface for studies of global change: models, algorithms- experiments. Remote Sens Environ. 51:3–26. Sengar SS, Kumar A, Ghosh SK, Wason HR. 2013. Earthquake induced built-up damage identification using IRS-P6 data – a comparative study using fuzzy based classifiers. Geocarto Int. [DOI:http://dx.doi.org/10.1080/10106049.2012.756939]. Sinha S, Mohanty S. 2012. Spatial variation of crustal strain in the Kachchh region, India: implication on the Bhuj earthquake of 2001. J Geodyn. 61:1–11. Singh RP, Bhoi S., Sahoo AK, Raj U., Ravindranath S. 2001. Surface manifestations after the Gujarat earthquake. Curr Sci. 81:338–340. Thakkar MG, Goyal B. 2004. On Relation between magnitude and liquefaction dimension at the epicentral zone of 2001 Bhuj earthquake. Curr Sci. 87:811–817. Tucker CJ. 1979. Red and photographic infrared linear combinations for monitoring vegeta- tion. Remote Sens Environ. 8:127–150. Wang J, Price KP, Rich PM. 2001. Spatial patterns of NDVI in response to precipitation and temperature in the Central Great Plains. Int J Remote Sens. 22:3827–3844.

Journal

"Geomatics, Natural Hazards and Risk" – Taylor & Francis

Published: Oct 2, 2014

There are no references for this article.