Preliminary results for a semi-automated quantiﬁcation of

Preliminary results for a semi-automated quantification ofsite effects using geomorphometry and ASTER satellite

data for Mozambique, Pakistan and Turkey

Alan Yong1,∗, Susan E Hough1, Michael J Abrams2 and Christopher J Wills3

1United States Geological Survey, 525 South Wilson Avenue, Pasadena, CA 91106, USA.2Jet Propulsion Laboratory/California Institute of Technology, 4800 Oak Grove Drive, Pasadena, CA 91109, USA.

3California Geological Survey, 801 K Street. MS 12–32, Sacramento, CA 95814, USA.∗e-mail: [email protected]

Estimation of the degree of local seismic wave amplification (site effects) requires precise informa-tion about the local site conditions. In many regions of the world, local geologic information is eithersparse or is not readily available. Because of this, seismic hazard maps for countries such as Mozam-bique, Pakistan and Turkey are developed without consideration of site factors and, therefore, donot provide a complete assessment of future hazards. Where local geologic information is available,details on the traditional maps often lack the precision (better than 1:10,000 scale) or the levelof information required for modern seismic microzonation requirements. We use high-resolution(1:50,000) satellite imagery and newly developed image analysis methods to begin addressing thisproblem. Our imagery, consisting of optical data and digital elevation models (DEMs), is recordedfrom the ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) sensor sys-tem. We apply a semi-automated, object-oriented, multi-resolution feature segmentation methodto identify and extract local terrain features. Then we classify the terrain types into mountain,piedmont and basin units using geomorphometry (topographic slope) as our parameter. Next, onthe basis of the site classification schemes from the Wills and Silva (1998) study and the Willset al (2000) and Wills and Clahan (2006) maps of California, we assign the local terrain units withVs30 (the average seismic shear-wave velocity through the upper 30 m of the subsurface) rangesfor selected regions in Mozambique, Pakistan and Turkey. We find that the applicability of oursite class assignments in each region is a good first-approximation for quantifying local site con-ditions and that additional work, such as the verification of the terrain’s compositional rigidity, isneeded.

1. Introduction

Detailed information for mitigating seismic hazardsis not readily available for all tectonically activeregions of the world. As an example, geotechni-cal site characterizations are often imprecise, ifavailable at all. There are many reasons for thisdeficiency. Although the mapping of tectonic ter-rain by local and international experts is rela-tively inexpensive, the resultant map information

is often not accessible due to geopolitical sensiti-vity. When map data are obtainable, it has beenfrequently observed to be too coarse, inconsistentor thematically unsuitable for detailed microzona-tion (Park and Elrick 1998; Wills and Clahan2006; Yong et al 2006a). Precise geophysical data,such as in situ shear-wave velocity measurements,are often too expensive or prohibitive due to theuse of invasive techniques, such as drilling andactive sources. In economically thriving countries,

Keywords. Shear-wave velocity; site conditions; geomorphometry; terrain modeling; object-oriented; satellite data;Geophysics, Geomorphology; remote sensing.

J. Earth Syst. Sci. 117, S2, November 2008, pp. 797–808© Printed in India.

Alan Yong et al

Figure 1. Regional map showing the locations(stars) of the 1999 M 7.4 Izmit–Turkey, 2005 M 7.6Muzaffarabad–Pakistan, and 2006 M 7.0 Beira–Mozam-bique earthquakes and the locations (green rhombi) of ourstudy areas.

such as the United States, Japan, and Taiwan,efforts to investigate and quantify seismic haz-ards dwarf the efforts in less prosperous coun-tries that often have comparable, if not higher,exposures to the same devastating effects fromstrong ground motions. Although the 2006 M 7.0Beira–Mozambique earthquake was not as dam-aging as the 1999 M 7.4 Izmit–Turkey, and 2005M 7.6 Muzaffarabad–Pakistan earthquakes (fig-ure 1), enormous damage and loss-of-life in theselatter regions underscored the need for a cost-effective and simple approach to mitigate seismichazards in all economically and technologicallypoor regions.

To provide a first-approximation of local geo-technical site conditions, we develop an approachto characterize potential ground motions on thebasis of known correlations between variations inshear-wave velocity and topographically distinc-tive landforms (Bard and Bouchon 1980; Bard andTucker 1985; Bard and Gabriel 1986; Sanchez-Sesma and Campillo 1993; Wills et al 2007; Yonget al 2008). The need for consistent and objectivemap information is fulfilled by satellite-based digi-tal elevation models (DEMs) with global coveragesuch as those captured by the Advanced Space-borne Thermal Emission and Reflection Radiome-ter (ASTER)(Hirano et al 2003). For correlatingshear-wave velocity to the DEM, we apply semi-automated imaging analysis methods based only ongeomorphometry (Yong et al 2005, 2006c, 2008).

Using tectonically active regions as our studyareas and no a priori information about localsite conditions, we develop methods using

Table 1. Characteristics of the three ASTER sensor sub-systems (Abrams et al 2002).

Band Spectral range SpatialSubsystem no. (µm) resolution (m)

VNIR

1 0.52–0.60

152 0.63–0.69

3N 0.78–0.86

3B 0.78–0.86

SWIR

4 1.60–1.70

30

5 2.145–2.185

6 2.185–2.225

7 2.235–2.285

8 2.295–2.365

9 2.360–2.430

10 8.123–8.475

11 8.475–8.825

TIR 12 8.925–9.275 90

13 10.25–10.95

14 10.95–11.65

geomorphometry and digital imaging analysis onASTER DEMs for selected regions in Turkey,Pakistan and Mozambique. These areas havegeomorphologically distinctive terrains that seteach depositional environment apart. Our results,based on the classification of terrain units inour Pakistan study area (Yong et al 2008),are surprisingly good for a first-approximationof local site effects. The results for our studyareas in Mozambique and Turkey are good first-approximations, but further work is required toverify the inferred, rigidity and shear-wave veloci-ties of the terrain units.

2. Data

ASTER is the ‘zoom lens’ instrument onboardthe National Aeronautics and Space Administra-tion (NASA) satellite Terra. Launched in 1999and currently in service, Terra’s payload includesfour other Earth Observation System (EOS)instruments that observe and record local- andregional-scale processes occurring on the Earth’sland surface and in the biosphere, atmosphere andoceans (Yamaguchi et al 1998; Abrams 2000).

ASTER was specifically designed to conduct geo-logical investigations (Abrams and Hook 1995).Terra’s sun-synchronous orbit is at an altitude of705 km and it returns to view the same locationevery 16 days. Four optical telescopes collectspectra in three separate recording subsystems(Abrams et al 2002) at spatial resolutions from15–90 m with spectral wavelength ranges from0.52–11.65 µm (table 1). The visible near-infrared(VNIR) system records four discrete bandpasses

Quantification of site effects using geomorphometry and ASTER satellite data

Figure 2. Geometry of the along-track recording mode forthe ASTER VNIR nadir (3N) and backward-viewing sensors(Kaab 2002).

(channels 1, 2, 3N, and 3B) that are collected fromtwo separate telescopes: a nadir-viewing telescopethat records channels 1, 2, and 3N; and a backward-viewing telescope that records channel 3B. Bothtelescopes can be rotated (+/− 24◦) as a unitand have spatial resolutions of 15 m. The short-wave infrared (SWIR) system records six discretebandpasses (channels 4–9) that are collected froma fixed viewing telescope that uses an adjustable(+/− 8.54◦ from nadir) scanning mirror for rota-tion and has a spatial resolution of 30 m. Also usingan adjustable scanning mirror with the same rota-tional range, the thermal infrared (TIR) channelsystem records five discrete bands (channels 10–14)and has a spatial resolution of 90 m.

ASTER imagery of the study areas in theMozambique, Pakistan and Turkey regions (fig-ure 1) were acquired from the EOS Data Gateway(EDG). The imagery consists of two types of data:the first type is the ASTER Level-1B (AST−L1B)product consisting of 14 discrete spectral bandsranging from the VNIR-TIR (0.52–11.65µm)region (table 1); and the second type is the ASTERLevel-3 (AST14DEM) single-band product consis-ting of a digital elevation model (DEMs) (table 1).

Using photogrammetric principles, the ASTERDEM is generated from the stereocorrelation(Hirano et al 2003) of spectral bands (VNIRregion) recorded by the nadir-viewing (3N) andbackward-viewing (3B) sensors (figure 2). At analtitude of 705 km, the ASTER input spatial res-olution is 15 m, but the output resolution inthe DEMs is a reduced 30 m posting of 3-Dcoordinates. The stereo-pair has a base-to-heightratio of 0.6 and an intersection angle of approxi-mately 27.6◦ (figure 2), which is close to idealfor a variety of terrain conditions (Hirano et al

Figure 3. Compilation map of the Mozambique study area.The top image layer (intense color code) is the rDEM(AST14DEM) and the bottom image layer (muted color) isa VNIR composite (AST−L1B).

2003). As opposed to cross-track modes of dataacquisition, where the nadir and backward imagesforming the stereo-pairs are acquired on differentdates, the along-track mode of the ASTER systemproduces an approximately one-minute lag timebetween the acquisition of the stereo-pairs (fig-ure 2) such that data are acquired in very similarenvironmental conditions, resulting in very consis-tent image quality (Kaab et al 2002; Hirano et al2003). Because the AST14DEM product is a rel-ative digital elevation model (rDEM), no groundcontrol points (GCPs) are used to tie in corre-lated features on the surface; instead of GCPs, anephemeris onboard the satellite is used.

The study areas in each region are determined bythe nominal size of the AST14DEM data becausethe coverage of the rDEM footprint is slightlysmaller than that of the AST−L1B scenes (fig-ure 3). This is a result of overlapping ASTER 3Nand 3B imagery during stereo-pairing to create the3-D coordinates. Thus, the extent of our study areais slightly less than 60 × 60 km.

3. Approach

3.1 Geomorphometry

To automatically identify the landform types thataffect site conditions, we first group these relief ele-ments into terrain features. Terrain features canbe described and categorized into simple topo-graphic relief elements or units by parameteri-zing DEMs (Bolongaro–Crevenna et al 2005).Here, parameterization is made possible because byapplying geomorphometry (quantitative descrip-tion of landforms) on DEMs, we can use semantic

Alan Yong et al

modeling (Dehn et al 1999) to represent land-forms. For example, on the basis of the Hammond(1964) landform classification scheme, Dikau et al(1991) were able to automatically classify land-form units in New Mexico using 96 different sub-classes and a 200 m window on DEM coverage.In addition, Yong et al (2008) have proposedthat using geomorphometric parameters (slope,aspect, azimuth, concavity/convexity, etc.) to clas-sify elemental terrain features into units (moun-tains, piedmonts, and basins) where such featuresare already known makes it possible to reapplythe same parameters to other regions where similargeomorphology exists. Geomorphometry also offersa variety of additional approaches for defining ter-rain units, such as the classification of terrainfeature parameters, filtering techniques, clusteranalysis, and multivariate statistics (Pike 2002;Bolongaro–Crevenna et al 2005).

For this study, we apply elements of the methoddescribed in Yong et al (2008) to characterize ter-rain features. On the basis of these characteriza-tions, seismic site conditions are then assigned asa first-approximation approach. In most cases, theterrain’s topographic relief is an adequate proxyto infer compositional rigidity of the site (Yonget al 2008). Relatively steep terrains are typicallycomposed of older and well-consolidated materi-als. In contrast, relatively flat terrains are com-monly younger and formed with less-consolidatedmaterials (Yong et al 2008). But in complex caseswhere tectonics are co-mingled with unexpectedgeomorphology (see discussion for examples), sitegeology must be considered because the localtopography may no longer reflect the expectedcompositional rigidity. Although this additionalapproach is beyond the scope of this study, the veri-fication of compositional rigidity can be performedthrough the identification of lithology by applyingpixel-based spectral analysis methods described byYong et al (2008).

3.2 Semi-automated imaginganalysis approach

We apply an object-oriented multi-resolution seg-mentation algorithm (Definiens 2006) to systemati-cally partition and identify the terrain classes ineach rDEM. The object-oriented paradigm is usedto segment the ASTER rDEM because traditionalper-pixel-based approaches tend to diffuse coherentfeatures and tend to produce a salt-and-peppereffect in the resultant image (figure 4).

The object-oriented approach treats groups ofpixels as object primitives. This is because, for high(30-m) spatial resolution data such as the ASTERrDEM, meso-scale terrain units are invariably rep-resented by multiple contiguous pixels that are

Figure 4. Diffused salt-and-pepper effect in the Vs30 clas-sification map of the Islamabad study are based on ASTERrDEM (Yong et al 2006b).

Figure 5. Diagram showing the hierarchical levels of objectprimitives (Definiens 2006; Yong et al 2008).

clustered collectively in a shape of finite spatialextent. To effectively segment image objects rep-resented by groups of pixels in the image space ordomain, shape, texture, color, context, and othermorphological characteristics must first be con-sidered (Navulur 2006). Then, through iterations,a multi-resolution segmentation algorithm refinesthe distinctive features of each terrain unit (moun-tain, piedmont and basin). The refinement processis performed at multiple levels and is based onthe hierarchical relationship between each objectprimitive (figure 5). Once the image objects inthe image space are analyzed, parameterized,extracted and segmented, the feature objects (ter-rain types) are ready for classification into theappropriate terrain unit.

4. Methods and results

To demonstrate how we extract mountain, pied-mont, and basin units from the AST14DEM of


Figure 6. Work flow diagram for generating preliminarysite characterization map.

Mozambique, Pakistan and Turkey, we use therDEM (figure 3) of Mozambique to illustrate ourapproach. Yong et al (2008) applied a similarapproach, but they used both rDEM and TIRspectral signatures for identifying rock types inPakistan. Here, we only apply the geomorphome-tric parameters and the object-oriented imaginganalysis methods (figure 6) developed by Yong et al(2008) to the rDEMs for each study area. Theresults for each study area (Mozambique, Pakistanand Turkey) are shown in figures 13, 14 and 15(respectively).

4.1 Preprocessing data

First, the AST−L1B data are deconvolved into14 discrete bands as 15 separate layers. Thenadir- (3N) and backward-viewing (3B) bands arecounted as one band because they both collectspectra in the same range, 0.78–0.86µm, but havedifferent viewing angles and so are automaticallydeconvolved into separate layers (table 1). It ispossible to stereographically pair the 3N and 3Bimages to generate our own rDEM, but to do so,atmospheric corrections, GCPs and other informa-tion, such as the satellite’s ephemeris, must be

used. For simplification, we use the ASTER rDEMproduct (AST14DEM).

Next, the rDEM is filtered to exclude pixels inthe data that are not related to the scene, suchas the black color pixels of the image frame (fig-ure 3). The minimum pixel value in each rDEMis −9999. Pixels with −9999 values are commonlyused to represent no data and are masked out priorto any imaging analysis work.

Then, the layer stacking of Bands (layers) 1, 2,and 3N are performed to generate the VNIR com-posite scenes (figure 7). The creation of the com-posites is intended for use as a visual guide toassist in the parameterization phases during theiterations of segmentation and classification workthat follow for each study area. No other informa-tion about the character of the local terrains wasused.

4.2 Object-oriented multi-resolutionsegmentation

We use Definiens Professional 5.0 (eCognition)software to segment groups of contiguous pixels,the so-called image objects. The eCognitionalgorithms take into account important contex-tual information such as the allowable degreeof heterogeneity; the importance of color; theimportance of shape; the importance of com-pactness; and the importance of smoothness(Definiens 2006). On the basis of these parame-ters, hierarchical levels of object primitives arecreated (figure 5) (Definiens 2006). Starting withthe pair-wise clustering of pixels (figure 5A) andfollowed by a bottom-up region-merging technique(figure 5B–D), the segmentation process takes intoconsideration semantic relationships between itsneighbor-objects (figure 5C), its sub-objects (fig-ure 5B), and its super-objects (figure 5D), creatingcontextual information (based on parent-child rela-tions) on multiple scales (Definiens 2006).

Four object primitive levels of segmentation werefound to be necessary. To determine the four effec-tive sets of contextual parameters that producedthe fourth and final object primitive level, multipleiterations of manual selection and deletion, priorto establishing the parameters, were required. Thecontextual parameters defined for all the objectprimitive levels are given in table 2.

Using the Mozambique study area to illustrateour approach, the first object primitive level (fig-ure 8) produces a very dense scene filled with smallobject primitives that outlined the steep reliefterrains (dark areas) that dominate the northerntwo-thirds of the study area. The second objectprimitive level (figure 9) is derived from the first(previous) object primitive level with modifiedparameters as given in table 2. The second object

Alan Yong et al

Figure 7. ASTER VNIR (AST−L1B) composites of studyareas in Mozambique (A), Pakistan (B) and Turkey (C).Bands 3N, 2 and 1 are assigned to red (R), green (G) andblue (B) colors, such that vegetation is reflecting the colorred.

primitive level scene maintains the outline of thesteep terrains as characterized by the dark fea-tures in the first object primitive level. Othersteep terrains, previously not apparent, start totake shape (figure 9). The third object primitivelevel (figure 10) is based on the second (previous)object primitive level (figure 9) and the modifiedparameters given in table 2. This level reveals amore coherent segmentation result, in terms of

distinguishable terrain features. The steep reliefterrains, trending northwest, near the northeastquadrant of the study area, implied in the firstobject primitive level have now taken form (fig-ure 10). In the fourth and final object primitivelevel (figure 11) the object primitives are basedon the third (previous) object primitive level (fig-ure 10) and the modified parameters given intable 2. Here, on the basis of visual comparisonswith the VNIR composite (figure 3), the mountain,piedmont and basin units, represented by the finalobject primitives, have now taken what we regardas acceptable form (figure 11).

4.3 Classification

We use the Nearest Neighbor (NN) classifier tocategorize the final object primitives into classeson the basis of the pre-defined sample object prim-itives. The NN classifier is a commonly used clas-sification method developed by Clark and Evans(1954).

First, we use the NN classifier in our supervisedtraining approach to sample and class our terrainunits. On the basis of the sampled representativevalues and a distance metric (neighborhood),the pixels, or in this case, object primitives areassigned to the associated class. The samples areacquired through several iterations of manuallyselecting the optimal object primitives in thefourth and final object primitive level (figure 11).The optimal object primitives are defined as thebest representation of the terrain units and arebased on the correlation of the mean relative ele-vation values for each terrain unit to the visuallyinterpretable terrain features in the VNIR com-posites. Next, on the basis of the selected sam-ples and the average distance between the centersof the object primitives, the NN classifier assignsthe remaining object primitives into the definedclasses. Then, we assign the shear-wave velocityranges to the three types of terrain units.

4.4 Assignment of shear-wave velocity ranges

Classifications of site conditions are commonlymade on the basis of Vs30 values, the average seis-mic shear-wave velocity through the upper 30 m ofthe subsurface (Borcherdt 1994; I.C.B.O. 1997). Indeveloping the California statewide site-conditionsmap, Wills and Silva (1998), Wills et al (2000)and Wills and Clahan (2006) used known Vs30-lithology correlations (Borcherdt 1970; Andersonet al 1996; Boore and Joyner 1997) to character-ize estimated shear-wave velocities in the near sur-face. On the basis of the Vs30-lithology correlationsand that lithology typically controls the nature


Table 2. Parameterization of the four object primitive levels (Yong et al 2008).

Object Parameterizationprimitivelevel Scale Shape Color Compactness Smoothness

1 10 0.1 0.9 0.5 0.5

2 50 0.5 0.5 0.5 0.5

3 100 0.7 0.3 0.5 0.5

4 (final) 250 0.3 0.7 0.5 0.5

Figure 8. Composite of the first object primitive level andthe relative DEM image (figure 3). The segmentation resultsproduced a very dense scene filled with small object primi-tives that outlined the steep relief terrains (dark areas).

of topography (thus, defining physiographic land-form types), we extend and generalize the resultsof the preliminary site-conditions map (Wills andClahan 2006) to a Vs30-terrain correlation. As afirst-approximation approach, we apply the Vs30-terrain correlation and assume that our terrainunits have typical velocities for mountains (hardrock), piedmonts (intermediate hard to soft rock)and basins (soft rock).

In addition to applying typical velocities to eachterrain unit, we apply overlapping velocity rangesto the three terrain units. For hard rock sites, weassigned the predicted Vs30 values to be greaterthan 500 m/s, which is observed mostly in rocks,such as granites, typical of mountain sites in south-ern California, as well as in the more consolidatedsedimentary rocks, such as carbonates (Wills andSilva 1998). Because most mountain features arecomposed of harder rock and have characteris-tically steep topographic profile, we assign themountain units the high portion of the velocityrange. For the soft rock sites, generally correlatedto the relatively flat relief of the terrain, the lowestvelocity values (< 300m/s) are assigned. For theintermediate terrain, we assign a Vs30 range that

Figure 9. Composite of the second object primitive leveland the relative DEM image (figure 3). The segmentationresults maintain the outline of the steep terrains as char-acterized by the dark features in the first object primitivelevel. Other steep terrains, previously not apparent, are nowtaking shape.

spans the difference between the end members ofthe mountain and basin units and overlaps eachunit by 100 m/s. These intermediate, moderatelysloped piedmont units have Vs30 values that rangefrom 200–600 m/s.

Although it is sufficient to assume that the velo-city range in each unit typifies the known cor-relations between Vs30 and the terrain types, weapply overlapping velocities to each unit’s rangeto account for the variability of observed velo-city (Wills and Clahan 2006) in the transitionzones (alluvial fans) located between the respec-tive units. Furthermore, these overlapping rangesaddress complexities (discussed in the next section)associated with the continuity of landforms and thevariability of slope in each landform type.

4.5 Resultant Vs30 maps

The geomorphology of the study areas inMozambique (figure 12A), Pakistan (figure 12B)and Turkey (figure 12C) are very different in natureand the resultant Vs30 maps (figures 13, 14 and

Alan Yong et al

Figure 10. Composite of the third object primitive leveland the relative DEM image (figure 3). The segmentationresults delineate the steep relief terrains, previously impliedin the first object primitive level, have now taken form.

Figure 11. Composite of the fourth and final objectprimitive level and the relative DEM image (figure 3). Atthis segmentation stage, based on visual comparisons withthe relative DEM image, the mountain, piedmont and basinunits, represented by the final object primitives, have nowfinally taken acceptable form.

15, respectively) reflect this disparity. Below, wedescribe selected features of the resultant Vs30maps and their relationship to the terrain featuresin each geomorphologically distinct region.

Knobby plateaus, steep gorges and ruggedmountains are dominant landform types in ourMozambique study area, near the Lago de CaboraBassa and Zambezi River (figure 12A). Albeit, inrelatively small numbers, topographically low-lyingbasins are adjacent to the southern and south-eastern portions of the Cabora Bassa shoreline andthe Serra Mepataluanda mountain range. A fewother low basin features intersperse the regions

Figure 12. Northerly perspective views of study areas inMozambique (A), Pakistan (B) and Turkey (C). Eachmodel is based on their respective VNIR composites drapedon to the rDEMs.

Figure 13. Map of predicted Vs30 for the Mozambiquestudy area.

north of Lago de Cabora Bassa and the ZambeziRiver. Here, we add a pseudo terrain unit (with-out shear-wave velocity assignment) to accountfor the body of water occupying Lago de CaboraBassa and the Zambezi River (figure 13). Themountain units, assigned the highest Vs30 range(> 500m/s), occupy the areas where we expectto find the region’s northwest trending mountains.The piedmont units, assigned the intermediate


Figure 14. Map of predicted Vs30 (Yong et al 2008) for thePakistan study area with verification of limestone geologybased on Williams et al (1999).

Figure 15. Map of predicted Vs30 for the Turkey studyarea.

range of Vs30 values (200–600 m/s), occupy most ofthe plateau terrain in the study area as expected.Also as expected, the basin units, assigned to thelowest Vs30 range (< 300m/s), occupy the low-lying basins, adjacent to the southern and south-eastern portions of the Cabora Bassa shoreline andthe Serra Mepataluanda mountain range.

The topographic variations in our study areanear Islamabad, Pakistan (figure 12B), consist ofsteep mountains and gentle valleys, which contrastthe plateau and gorge features in our Mozambiquestudy area (figure 12A). The Margala Hills, anortheast-trending mountain range, dominate thetopographic expression of the Pakistan study area.Similarly trending linear ridges occupy the south-western part of the study area. At the foot ofthe Margala Hills mountain range, an intermediate

(piedmont) terrain feature extends from the moun-tain front until it transitions into the Soan RiverValley basin. Here, the capital city of Islamabadand the old city of Rawalpindi occupy this pied-mont bench (Williams et al 1999). The remain-ing major basin in the study area is the Peshawarbasin, just northwest of Islamabad. The mountainand basin units are clearly identified and corre-late well (figure 14) to the results of recent studies(Williams et al 1999; Nawaz et al 2004; Munir andButt 2007). In contrast, on the basis of the studyby Williams et al (1999), the piedmont units arenot as well correlated to their expected locations.A discussion about the factors potentially con-tributing to the misclassification of the piedmontterrain features can be found at the end of thissection. The mountain units, with the highest Vs30range, are assigned to the limestone Margala Hillsand the northeast-trending linear ridges (Williamset al 1999; Nawaz et al 2004; Munir and Butt 2007;Yong et al 2008). The piedmont units, assigned tothe intermediate and largest Vs30 range, occupythe Peshawar basin and Soan River Valley basin.The basin units appear to occupy their expectedlocations (Williams et al 1999) in the Peshawarand Soan River Valley basins. These units areassigned to the lowest Vs30 range (< 300m/s).

The steep ridge and valley landform features inour Turkey study area (figure 12C), north of thecapital of Ankara, near the city of Cubuk, also con-trast the features of our study areas in Pakistanand Mozambique. Other than the basins occupiedby the cities of Cubuk and Akyurt, the regionis dominated by multiple mountain ranges. Themountain units, near the Mir Dagi mountain range,dominate the northwestern portion of the Vs30 mapand are assigned shear-wave velocities greater than500 m/s (figure 15). Other mountain units, in thesoutheastern and northeastern parts of the studyarea, near the Eldivan Dagi and the Idris Dagimountain ranges (respectively), are assigned Vs30values greater than 500 m/s. Cubuk and Akyurtare assigned to the basin unit as expected andhave Vs30 values less than 300 m/s because thesecities occupy the low-lying basins in the study area.The piedmont units are assigned with Vs30 val-ues between 200 and 600 m/s. Without additionalinformation, such as surface geology, the qualityof the shear-wave velocity assignment to the pied-mont unit cannot be determined with certainty.

The geomorphology of the study areas inMozambique, Pakistan and Turkey are very differ-ent in nature. These differences can be problematicwhen using generalized parameters to account forthe regional variations in topographic trends. Forexample, the parameters set for the intermediatepiedmont terrain unit in one region might notproperly identify the geomorphic expression of

Alan Yong et al

a similar unit in another region: Williams et al(1999) indicated that Islamabad and Rawalpindiare on a piedmont bench at the foot of the MargalaHills mountain range, but the resultant Vs30 map(figure 14) did not identify the intermediate unitin its expected location. The contrast in the resultsfor the extraction of the mountain and basinunits, in comparison to the piedmont unit, maybe attributable to the lack of a standard defini-tion for determining the boundary conditions atthe termini of each terrain unit. On the basis ofour assignment scheme, our semi-automated imag-ing analysis approach, based on geomorphometryand terrain units, appears to provide a good first-approximation of seismic site conditions.

5. Discussions and conclusions

Our motivation for developing an approach using asemi-automated imaging analysis method on satel-lite data is to objectively and systematically char-acterize geotechnical site conditions in tectonicallyactive regions of the world. Traditional mappingmethods have been found to be unsuitable for thelevel of details (better than 1:10,000 scale) requiredin modern seismic microzonation purposes becausethey are either too inconsistent (Yong et al 2006b),too coarse (Park and Elrick 1998) or thematicallyinappropriate (Wills and Clahan 2006). Traditionalgeophysical site characterization approaches areoften prohibitive because they are environmentallyinvasive, too costly, and/or the sites of interestare inaccessible. To meet the need for more pre-cise geotechnical site information, we use high-resolution (1:50,000 scale) satellite data to beginaddressing the issues found in traditional mappingtechniques. Other studies, such as Romero andRix (2001) and Wald and Allen (2007), have beenreported on the use of satellite data to account forsite effects.

Our approach has both advantages and dis-advantages when compared to these studies. Byusing the combination of high spatial (15 m post-ings) and multi-spectral data, our approach takesadvantage of terrain topography (rDEM) and com-position information (VNIR imagery) at a mapscale that is comparable to the 1:50,000 level of pre-cision for mapping purposes (Hirano et al 2003).In contrast, Romero and Rix (2001) used Landsat7 ETM+ (Enhanced Thematic Mapper Plus) satel-lite data with lower spatial (30 m postings) reso-lutions for characterizing age deposits to predictsite response in the Mississippi embayment. Theirpixel resolution was at the 1:100,000 scale, whichcan resolve only half of the details in an ASTERimage. An additional advantage to our approach isthat our methods are extensible, in that, these can

be applied to higher resolution data from satellites,such as Quickbird, IKONOS and Cartosat-2. Fur-thermore, the data acquired by the Landsat mis-sion have recently discovered to be plagued withdata gaps. These artifacts are currently beingfilled by applying interpolation methods (Kurtzet al 2006). In another study, Wald and Allen(2007) used DEMs from the Shuttle Radar TerrainMission (SRTM) project to predict site responsebased on their correlation of topographic slopeto observed shear-wave velocity measurements.The DEMs used were at 30-arc-second resolution(∼ 900m pixel resolution) and were generatedusing Synthetic Aperture Radar (SAR) technology.When compared to the Romero and Rix (2001)study and our study, the effective resolution of theDEM used by Wald and Allen (2007) is thirty timescoarser and is less suitable for microzonation pur-poses (Tinsley et al 2004). For regions with verysteep topographical slopes that are facing awayfrom the radar during the SRTM mission, and forregions with smooth areas, such as water or sand,the SAR technique cannot accurately measure theterrain properly (Farr et al 2007). Also, SRTMDEMs do not have global coverage, so the datacontain no information for latitudes north of 60◦N,which coincide with the seismically active regionsof Alaska, and south of 59◦S.

A disadvantage of using our approach is thatit is more computationally intensive than a sim-ple slope analysis. Although coarse and impre-cise, the Wald and Allen (2007) approach is fasterand more appropriate for input into an earth-quake response product, such as ShakeMap (Waldet al 2003). Despite differences in spatial resolu-tion, both Wald and Allen (2007) and this studyshare a common disadvantage: variations in topo-graphic slope do not always directly relate to thecompositional rigidity of the terrain. For example,certain exceptional terrains, such as the flat ModocPlateau and steep Mendocino Hills in California,have compositional rigidity that are not expectedas a result of their topographic expressions.

Finally, while our maps (figures 13, 14, and15) provide a basis for estimating site responsein the Mozambique, Pakistan and Turkey regions,any site characterization map needs to be com-pared and validated with ground-motion record-ings. Recorded ground motions from the 1999M 7.4 Izmit, 2005 M 7.6 Muzaffarabad, and 2006M 7.0 Beira earthquakes would provide indepen-dent estimation of amplification factors. Thesedata have not yet been made available to the inter-national community. In the absence of instrumen-tally recorded ground motions, a detailed intensitysurvey of shaking in the regions during theseearthquakes would also provide a useful basis forcomparison.


Acknowledgements

We greatly appreciate the helpful comments byRobert S Dollar and Karen R Felzer, in additionto earlier discussions with Edward E Field andKen W Hudnut. We also greatly appreciate techni-cal guidance from Matthias Stolz and John Parkerof Definiens AG. Work done by Michael J Abramswas performed at the Jet Propulsion Laboratory/California Institute of Technology, under contractto the National Aeronautics and Space Adminis-tration. We thank Linda Gundersen for the partialsupport through U.S. Geological Survey WorkingCapital Fund.

References

Abrams M 2000 The advanced spaceborne thermal emissionand reflectance radiometer (ASTER): Data products forthe high spatial resolution imager on NASA’s Terra plat-form; Int. J. Remote Sens. 21 847–859.

Abrams M and Hook S J 1995 Simulated Aster data forgeologic studies; IEEE Trans. Geosci. Remote Sens. 33692–699.

Abrams M, Hook S and Ramachandran B 2002 ASTER UserHandbook, Jet Propulsion Laboratory/California Insti-tute of Technology, Pasadena, 135 pp.

Anderson J G, Lee Y, Zeng Y and Day S 1996 Control ofstrong motion by the upper 30 meters; Bull. Seismol. Soc.Am. 86 1749–1759.

Bard P Y and Bouchon M 1980 The seismic response ofsediment-filled valleys. Part 1. The case of incident SHwaves; Bull. Seismol. Soc. Am. 70 1263–1286.

Bard P Y and Gabriel J C 1986 The seismic responseof two-dimensional sedimentary deposits with large ver-tical velocity gradients; Bull. Seismol. Soc. Am. 76343–346.

Bard P Y and Tucker B E 1985 Underground and ridge siteeffects: A comparison of observation and theory; Bull.Seismol. Soc. Am. 75 905–922.

Bolongaro-Crevenna A, Torres-Rodrıguez V, Sorani V,Frame D and Ortiz M A 2005 Geomorphometric analy-sis for characterizing landforms in Morelos State, Mexico;Geomorphology 67 407–422.

Boore D M and Joyner W B 1997 Site amplificationfor generic rock sites; Bull. Seismol. Soc. Am. 87327–341.

Borcherdt R D 1970 Effects of local geology on groundmotion near San Francisco Bay; Bull. Seismol. Soc. Am.60 29–61.

Borcherdt R D 1994 Estimates of site-dependent responsespectra for design (methodology and justification); Earth-quake Spectra 10 617–653.

Clark P J and Evans F C 1954 Distance to nearest neighboras a measure of spatial relations in populations; Ecol. 35445–453.

Definiens 2006 Definiens Professional 5 Basic TrainingMaterial; Definiens AG, Munchen, 69 pp.

Dehn M, Gartner H and Dikau R 1999 Principles of seman-tic modeling of landform structures, 4th InternationalConference on GeoComputation 1999, July 25–28:Fredericksburg, Virginia.

Dikau R, Brabb E E and Mark R M 1991 Landform clas-sification of New Mexico by computer; U.S. Geol. Surv.Open-File Rept. 634 15 pp.

Farr T G, Rosen P A, Caro E, Crippen R, Duren R,Hensley S, Kobrick M, Paller M, Rodriquez E, Roth L,Seal D, Shaffer S, Shimada J, Werner M, Oskin M,Burbank D and Alsdorf D 2007 The shuttle radar topog-raphy mission; Jet Propulsion Laboratory/CaliforniaInstitute of Technology, Pasadena, 43 pp.

Hammond E H 1964 Classes of land surface form in theforty-eight states, U.S.A.; scale 1:5,000,000.

Hirano A, Welch R and Lang H 2003 Mapping fromASTER stereo image data: DEM validation andaccuracy assessment; Photogramm. Remote Sens. 57356–370.

I.C.B.O. (International Conference of Building Officials)1997 Uniform Building Code, International Conference ofBuilding Officials, Whittier, 1411 pp.

Kaab A, Huggel C, Paul F, Wessels R, Raup B, KiefferH and Kargel J 2002 Glacier monitoring from ASTERimagery: accuracy and applications; In: Proceedings ofEARSeL-LISSIG-Workshop Observing our Cryospherefrom Space, March 11–13, Bern, Switzerland.

Kurtz R, Kline K and Lacasse J 2006 Preliminary review ofthe new Landsat 7 segment-based gap-fill; In: Proc. Thir-teenth Australasian Remote Sensing and PhotogrammetryConference, November 20–24, Canberra, Australia.

Munir M and Butt A A 2007 The paleogene of AzadKashmir, Hazara Kashmir syntaxis, Pakistan; In: Proc.Geophysical Research Abstracts of 2007 European Geo-sciences Union, Volume 9-11052, April 15–20, Vienna,Austria.

Navulur K 2006 Multispectral image analysis using theobject-oriented paradigm, CRC Press, Boca Raton,165 pp.

Nawaz F, Hamidullah F and Fayaz A 2004 The effectof mining on geomorphology; detection of changes byusing remote sensing techniques; In: Proceedings of theISRPS 2004 Annual Conference, July 19-23, Istanbul,Turkey.

Park S and Elrick S 1998 Predictions of shear-wave veloc-ities in southern California using surface geology; Bull.Seismol. Soc. Am. 88 677–685.

Pike R J 2002 A bibliography of terrain modeling (geomor-phometry), the quantitative representation of topogra-phy; Supplement 4.0; U.S. Geol. Surv. Open-File Rept.465 158 pp.

Romero S M and Rix G J 2001 Ground motion amplificationof soils in the upper Mississippi embayment, Report No.GIT-CEE/GEO-01-1, 461 pp.

Sanchez-Sesma F J and Campillo M 1993 Topographiceffects for incident P, SV and Rayleigh waves; Tectono-phys. 218 113–125.

Tinsley J, Hough S E, Yong A, Kanamori H, Yu E, Appel Vand Wills C J 2004 Geotechnical Characterization ofTriNet Sites: A Status Report; Seismol. Res. Lett. 75505.

Wald D J and Allen T I 2007 Topographic slope as a proxyfor seismic site conditions and amplification; Bull. Seis-mol. Soc. Am. 97 1379–1395.

Wald D J, Wald L, Worden B and Goltz J 2003 ShakeMap –a tool for earthquake response; U.S. Geol. Surv. Fact Sht.87 4 pp.

Williams V S, Pasha M K and Sheikh I M 1999 Geologicmap of the Islamabad–Rawalpindi area, Punjab, northernPakistan (map supplement); U.S. Geol. Surv. Open-FileRept. 47 16 pp.

Wills C J and Clahan K B 2006 Developing a map of geolog-ical defined site-condition categories for California; Bull.Seismol. Soc. Am. 96 1483–1501.

Wills C J, Petersen M, Bryant W A, Reichle M,Saucedo G J, Tan S, Taylor G and Treiman J 2000

Alan Yong et al

A site-conditions map for the California based on geol-ogy and shear-wave velocity; Bull. Seismol. Soc. Am. 90S187–S208.

Wills C J, Silva M and Yong A 2007 Progress toward moredetailed site-conditions maps in southern California; In:Proc. SCEC 2006 Annual Meeting, September 8–12, PalmSprings.

Wills C J and Silva W 1998 Shear-wave velocity character-istics of geologic units in California; Earthquake Spectra14 533–556.

Yamaguchi Y, Kahle A B, Tsu H, Kawakami T and Pniel M1998 Overview of advanced spaceborne thermal emissionand reflection radiometer (ASTER); IEEE Trans. Geosci.Remote Sens. 36 1062–1071.

Yong A, Hough S E, Hudnut K, Healy S, Giberson J andSimila G 2005 Geotechnical site characterization usingremote sensing data: A feasibility study; In: Proc. SSA2005 Annual Meeting, Lake Tahoe, NV.

Yong A, Hough S E, Wills C J and Abrams M 2006aA report on the geotechnical site characterization effortsin southern California; In: Proc. SCEC 2006 AnnualMeeting, September 9–14, Palm Springs.

Yong A, Hough S E, Cox H, Tiampo K, Braverman A,Harvey J, Hook S, Hudnut K and Simila G 2006b Bal-loons to satellites: a century of progress in geotechnicalsite characterization; In: Proc. SSA 2006 Annual Meet-ing, San Francisco, CA.

Yong A, Hough S E, Wills C J and Abrams M J2006c Site Characterization Using Satellite Imagery;In: Proc. AGU 2006 Fall Meeting, December 11–15San Francisco, CA.

Yong A, Hough S E, Abrams M J, Cox H M, Wills C J andSimila G W 2008 Quantification of site characterizationusing integrated imaging analysis methods on satellitedata of the Islamabad, Pakistan region Bull. Seismol.Soc. Am. 98(6) (in press)

MS received 28 September 2007; revised 25 March 2008; accepted 1 April 2008

Documents

Preliminary results for a semi-automated quantiﬁcation of