Database and spatial distribution of landslides triggered by the Lushan, China Mw 6.6 earthquake of 20 April 2013

Abstract On 20 April 2013, an earthquake with Mw 6.6 (Ms 7.0) hit Lushan County, Sichuan Province of China (30.3°N, 103°E), only 85 km southwest of the 2008 Wenchuan Mw 7.9 event. The Lushan shock triggered a large number of landslides with various types, including highly disrupted shallow slides and rock falls, deep-seated landslides and large-scale rock avalanches. Post-earthquake high resolution aerial photographs and satellite images, as well as a series of pre-earthquake high resolution satellite images were collected to construct a detailed, accurate, objective, and complete coseismic landslide database/inventory, which was validated by field investigation in selected areas. Based on the aerial photographs (0.2 or 0.6 m resolution) that partly cover the affected area and pre-earthquake satellite images, 14,580 coseismic landslides were recognized and mapped. In those areas without post-earthquake aerial photographs coverage, we mapped 7,948 more coseismic landslides based on post-earthquake satellite images (Rapideye images with 5 m resolution and ZY-3 images including panchromatic images with 2.1 m resolution and multi-spectral images with 5.8 m resolution) and pre-earthquake satellite images. The 22,528 coseismic landslides, with a horizontal projection area of 18.88 km 2 and an estimated total volume of 41.56 × 10 6  m 3 , were distributed in a nearly elliptic area of about 5,400 km 2 . Correlations between coseismic landslide abundances and topographic, geologic, and seismic factors were analyzed on 15,546 landslides of area ≥ 100 m 2 . Scatter plots between regional slope angle and landslide abundances in generalized homogeneous lithology combinations show no obvious relationship is present between them, which challenges the conclusions of previous studies. Combining the Lushan earthquake-triggered landslides with other events on reverse and/or thrust faults such as the 1999 Chi-Chi and 2008 Wenchuan shocks, we found that coseismic landslides tend to occur in areas enveloped by two or more imbricated reverse and/or thrust faults, rather than only along the seismogenic fault. The Lushan event shows a higher ability of triggering landslides compared with other earthquakes of similar magnitudes, which is likely due to the steep and rugged topography and fractured and densely jointed lithology resulted from long-term tectonic activity. The blind reverse seismogenic fault is likely also another significant factor. Based on detailed coseismic landslide inventory maps and databases of several recent major earthquakes such as the 2008 Wenchuan, 2010 Haiti, 2010 Yushu, and 2013 Lushan temblors, we suggest that the empirical correlations between earthquake magnitude and coseismic landslides need to be updated.

Lushan earthquake to trigger landslides. The collision of the Indian and Eurasian plates resulted in the uplift of the Tibetan Plateau and eastward motion of a series of blocks in this highland (Tapponnier et al., 2001;Chen et al., 2000;Wang et al., 2001;Gan et al., 2007). Obstructed by the rigid Sichuan Basin, the Longmenshan thrust zone formed along the boundary between the Tibetan Plateau and Sichuan Basin , at which the accumulated strain is released during frequent earthquakes (Xu et al., 2009c;Zhang et al., 2010). Both the 2008 Wenchuan and the 2013 Lushan temblors occurred on this northeast-trending fault zone. The epicenter of the Lushan event is located only 85 km southwest of the 2008 Wenchuan shock (Fig. 1).
The Longmenshan thrust zone is composed of three main thrust fault systems, which are the Maoxian-Wenchuan fault (also called back-range fault), the Yingxiu-Beichuan fault (also called central fault), and the Guanxian-Jiangyou fault (also called front-range fault) from northwest to southeast. In addition, a fault buried by the piedmont, southeast of the three faults, showing features of forward-propagating deformation, is also considered as a secondary fault of the Longmenshan thrust zone (Chen et al., 1994;Deng et al., 1994;Royden et al., 1997;Clark and Royden, 2000). Furthermore, the three fault systems aforementioned contain sub-faults ( 10 about 85 km apart, the two events are not considered to be generated by the same fault. The Wenchuan earthquake was spawned by the Yingxiu-Beichuan fault, with three surface ruptures on three faults, i.e. the Yingxiu-Beichuan, Guanxian-Jiangyou, and Xiaoyudong faults (Xu et al., 2009c;Tan et al., 2012). The epicenter of the Lushan event is, on the other hand, located near the Shuangshi-Dachuan fault. Since there is a nearly 40-km-long rupture gap between the Lushan and the Wenchuan earthquakes (Chen et al., 2013), most researchers do not consider the Lushan event to be an aftershock of the Wenchuan shock, but an independent one (Xu et al., 2013g;Jia et al., 2014;Chen et al., 2014c). On the other hand, the epicenter of the Lushan quake is located at the Coulomb stress increasing zones after the 2008 Wenchuan event (Parsons et al., 2008;Toda et al., 2008;Shan et al., 2013). Therefore, the latter might be an important trigger for the former. In other words, the 2008 Wenchuan great shock may have speeded up the occurrence of the 2013 Lushan earthquake.
In the Lushan temblor affected area, there are three major northeast-striking faults: the Yanjing-Wulong, the Shuanshi-Dachuan, and the Western Shangli faults from northwest to southeast (Fig. 2). It was originally proposed that the Shuangshi-Dachuan fault was the seismogenic fault of the Lushan event based on preferred nodal planes of focal mechanism solutions and regional tectonic setting. However, after subsequent field investigations, more detailed analyses of seismic waves, focal mechanism solutions and spatial distribution of aftershocks, it is suggested that the seismogenic fault of the Lushan quake is a blind reverse fault southeast of A C C E P T E D M A N U S C R I P T 3 Data and methods

Remote sensing images
In this study, we utilize aerial photographs and satellite images to construct coseismic landslide databases.
The pre-earthquake images used include SPOT-5 images (panchromatic images of 2.5 m and multi-spectral of  (Fig. 3). The aerial photographs were ortho-rectified and geo-referenced, and can be directly used to delineate coseismic landslides. Before landslide interpretation, the Rapideye and ZY-3 images were processed with some methods, including system calibration, ortho-rectification, geometric correction, data fusion, and true-color composite. All these procedures were performed on the ENVI platform.

Topographic, geologic, and seismic data
Available digital elevation models (DEMs) of the study area includes the SRTM DEM of 90 m and the ASTER GDEM of 30 m resolutions. Although the ASTER GDEM has higher resolution, it contains more noise and distortion. Therefore, we selected the SRTM DEM for this study. We resampled this DEM to a DEM of 10 m resolution using the bilinear method to improve statistical precision on areas of small-scale landslides. The resampling process would not add local topographic information, neither would it change the general topographic information of the original SRTM DEM data. The common topographic thematic maps of slope gradients and aspects were derived from the processed DEM of 10 m resolution. Slope position is a newly proposed factor that may affect coseismic landslide occurrence (Weiss, 2001;Jenness et al., 2013). The slope position thematic layer of the study area was extracted from a database of slope localities of China in Geospatial Data Cloud (www.gscloud.cn). Drainages of the affected area are extracted from the 10 m-resolution DEM and manually corrected based on high-resolution remote sensing images.

A C C E P T E D M A N U S C R I P T
ACCEPTED MANUSCRIPT 14 C 1 ): quartz sandstone, siltstone, limestone, and shale; (7) Devonian (D), (D 3 , D 3 s, D 2-3 , D 2 , D 2 y, D 2 g, D 1 p, D 1 g, D 1 , D 1 2 , D 1 1 ): limestone, dolostone, quartz and sandstone; (8) Silurian (S), (S, S 3 , S 2-3 , S 2 , S 2 l, S 1 , S 1 l): sandy shale, dolomitic limestone, and mudstone; (9)   Seismic data used in this study include the location of the epicenter, regional peak ground acceleration (PGA), seismic intensity, and the parameters of the seismogenic fault of the Lushan earthquake. The epicenter of the Lushan earthquake is located at 30.3°N, 103°E according to the China Earthquake Network Center (CENC, www.cenc.ac.cn). Although there were a series of real-time vibration data from seismic stations recorded during the main shock (CSMNC, www.csmnc.net), only a few of which are near the epicenter (Fig. 2).
Therefore, it is difficult to get PGA values based only on data interpolation from seismic stations in operation.
The US Geological Survey (USGS, www.usgs.gov) released a series (4 versions) of PGA maps related to the Lushan earthquake (US Geological Survey, 2013). The Version 1 was only based on a simulation-based method without considering the geometry of the seismogenic fault, whereas Version 4 with a moderate objectivity was obtained from a combination of regional tectonic settings, records from a few seismic stations, and numerical simulations. The PGA map of version 4 with a moderate uncertainty is more objective and accurate than the previous versions. Therefore, we used the Version 4 PGA map to correlate coseismic landslides with the ground shaking. The -Seismic Intensity Map of the 2013 Lushan Earthquake‖ was released by the China Earthquake Administration several days after the earthquake, mainly based on field investigations. Since the Lushan event was presumably caused by breaking of a blind reverse fault without surface ruptures, we used the preferred nodal plane from several released focal mechanism solutions of the main shock to represent the seismogenic fault of the earthquake.

Landslides detection
The methods for landslide detection include field investigations, visual interpretation of paper-based stereo-pair aerial photographs with the aid of stereo microscopes, identification of landslides from a high resolution DEM (e.g., LiDAR), visual interpretation of high resolution satellite ortho-images, deformation analysis from multi-temporal radar images based on InSAR technology, and automatic identification of landslides from optical remote sensing images (e.g. Keefer, 2002;Harp et al., 2011;Guzzetti et al., 2012;Xu, 2014a). For earthquake-triggered landslides, especially shortly after the earthquake, areas affected by coseismic landslides would always show an obviously different appearance from the surrounding areas on optical remote sensing images. Coseismic landslides can be easily recognized and detected on images with high resolutions (~1-10 m or higher), which are almost comparable with results from field investigations. In addition, landslides triggered by large earthquakes are usually large in number with high density and broad distribution, thus it is not realistic to prepare a detailed coseismic landslide database based only on field investigations. Therefore, the most suitable method of this study is visual interpretation of high-resolution remote sensing images, which is then verified by field inspection in selected areas. With the aid of computer, GIS and remote sensing technologies, we used a method of syntheses of computer screen-based visual interpretation of high-resolution aerial photographs and satellite images on the GIS platform, which has been shown to be the most objective and best method for this issue (Liao and Lee, 2000;Harp et al., 2011;Dai et al., 2011a;Gorum et al., 2013;Xu, 2014a;Xu et al., 2014b, c, d).
The quality of a coseismic landslide inventory map or database depends upon three aspects, which are the accuracy of geographic location of landslides, the amount of omission (false negative) errors and quantity of commission (false positive) errors. In order to ensure the geographic location accuracy of the landslides, we collected a series of actual GPS routes and points from field investigations to correct the remote sensing images.
The deviations between the GPS data and remote sensing images were lower than one grid of the Rapideye and ZY-3 images (~5 m). In order to limit omission errors, we collected aerial photographs and satellite images of high resolutions that cover the entire earthquake affected area. Although the aerial photographs of 0.2 and 0.6 m resolutions do not cover the entire earthquake affected area, we obtained supplementary Rapideye and ZY-3 satellite images with resolutions of 5 and 2.1 m. This ensured the landslides larger than 100 m 2 can clearly be shown on the images. In addition, we used a computer with screen height of about 300 mm, and zoomed in the image so that the actual length on the image in a computer screen was about 300 m. This is equivalent to working on images of a scale 1:1,000, and ensures that almost all landslides clearly displayed on the images can be recognized, and delineated. To limit commission errors, we checked every landslide on pre-earthquake images to distinguish pre-earthquake landslides from coseismic landslides by the methods mentioned by Xu (2014a). Since the post-earthquake aerial photographs were taken within three days after the main shock, and the satellite images were taken within about one month after the main shock, very few non-seismic landslides occurred after the main shock and appeared on the images. Although rainfalls did occur during those days, the number of the landslides triggered by these rainfalls or other factors after the main shock would be quite limited.
Furthermore, since the post-earthquake images we used were taken shortly after the main shock, most of the coseismic landslides would preserve their original shapes, with very few new vegetation covers on landslide bodies that could affect the visual interpretation.

Field inspection
Although a large number of high-resolution remote sensing images were available, and the accuracy of A C C E P T E D M A N U S C R I P T

ACCEPTED MANUSCRIPT
17 inventory maps and databases based only on visual interpretation of pre-and post-earthquake images has already become satisfactory (Xu et al., 2014c), we still carried out filed investigations to validate the coseismic landslide inventory map. Two stages of filed investigations were performed, which are the emergent field reconnaissance from the day of the main shock to 1 May 2013 (Xu et al., 2013f) and the detailed filed investigation lasting 27 days from 25 May 2013 (Xu et al., 2015). In total, one-way accumulated expedition routes excluding repeated survey routes during the two stages of field investigations exceed 1,000 km in length ( Fig. 3; Xu et al., 2015). The major access of the field area was by car, next by walking. We were able to observe all coseismic landslides mapped on remote sensing images that we selected in the field inspection.
However, some very small-scale landslides (< 50 m 2 in area), mostly as blocks rolling downhill, cannot be observed on the 5 m resolution Rapideye satellite images, but can usually be seen on the 0.2 and 0.6 m resolution aerial photographs. Therefore, in order to eliminate the uneven omission errors with those very small-scaled landslides in the subsequent spatial distribution analyses, all coseismic landslides of an area less than 100 m 2 were excluded in the subsequent analysis.

GIS spatial analysis of coseismic landslides
Landslide abundance (or landslide density) is a common index to measure patterns and characteristics of coseismic landslides. It generally includes three different indexes, which are the landslide number abundance (LNA), area abundance (LAA), and volume abundance (LVA). LNA is defined as the number of landslides per square kilometer, LAA is the percentage of the area affected by the landslides, and LVA is the average thickness of landslide materials in a given area, which is calculated by dividing the volume of coseismic landslides with the given area. LNA and LAA are commonly used in previous studies. In earlier studies, due to the lack of advanced remote sensing and GIS technologies, it was difficult to determine the boundary and area of coseismic landslides, thus usually only LNA was calculated (Jibson and Keefer, 1989;Keefer, 2000), with occasionally LAA reported (Tibaldi et al., 1995). Although LNA is easy to calculate, the scales of coseismic landslides cannot

ACCEPTED MANUSCRIPT
18 be reflected in this parameter. With the maturation of GIS and remote sensing technologies, more and more coseismic landslide inventory maps are polygon-based. Therefore, with a few exceptions (Sato et al., 2007;Qi et al., 2010;Gorum et al., 2011), most recent studies report LAA as the statistical index (Liao and Lee, 2000; Sepúlveda et al., 2010;Dai et al., 2011a;Gorum et al., 2013Gorum et al., , 2014Xu et al., 2013dBasharat et al., 2014;. Although making a polygon-based landslide inventory map is time-consuming, it is warranted because it is more accurate than point-based inventory maps for landslides susceptibility and hazard assessments (Xu et al., 2012c). Moreover, landslide -area to volume‖ power-law relations have been recently employed in many regions of the world (ten Brink et al., 2006;Guzzetti et al., 2009;Larsen et al., 2010;Parker et al., 2011), which make LVA in coseismic landslides distribution analyses available Xu et al., 2014c). Therefore, we selected the three indexes to perform spatial distribution analyses of the landslides. It should be noted that landslide volume was calculated based on a simple scaling relationship to convert an individual landslide area into individual landslide volume: where V i represents the volume of a landslide (i-th landslide) and A i is the area occupied by the landslide. The two scaling parameters α and γ vary with different landslide types and cases. We assigned α = 0.146 and γ = 1.332, which are derived from various types of landslides (Larsen et al., 2010) and have been used in previous studies (Parker et al., 2011;Xu et al., 2014c,e).

Database of landslides triggered by the Lushan earthquake
We delineated in total 27,259 landslides based on post-earthquake aerial photographs and satellite images (Rapideye and ZY-3) in the earthquake affected area (Fig. 3). Of these, 4,731 landslides were confirmed to be present before the earthquake based on pre-earthquake images and do not show any signs of enlargement or

A C C E P T E D M A N U S C R I P T
remobilization during the event. The remaining 22,528 landslides, either newly formed landslides or remobilized or enlarged of old landslides, are considered as coseismic landslides (Fig. 4). The 22,528 landslides, with 18.88 km 2 in total area and 41.56 × 10 6 m 3 in total volume, are distributed in an area about 5,400 km 2 (Fig.   4). Compared with the landslides triggered by the 2008 Wenchuan earthquake (Xu, 2013b; which are 197,481 in number, about 1,160 km 2 in occupying area, about 6 km 3 in landslide volume, and about 110,000 km 2 in landslide distribution area, the landslides triggered by the Lushan earthquake equal to about 1/9 in number, 1/60 in occupying area, 1/150 in volume, and 1/20 in distribution area of those of the Wenchuan event, respectively. Among the 22,528 coseismic landslides induced by the Lushan earthquake, 14,580 landslides were detected from the high resolution aerial photographs, while 7,948 landslides were recognized on Rapideye or ZY-3 satellite images. Of these, 162 landslides are of area ≥ 10,000 m 2 , 4,378 landslides are between 1,000 and 10,000 m 2 , 11,006 landslides are between 100 and 1,000 m 2 , and the remaining 6,982 are <100 m 2 . Most of the landslides are relatively shallow slides, typically 1-5 m in depth. Although large-scale deep-seated landslides and rock avalanches are less than small shallow disrupted landslides, the former contribute to the total volume of landslide material due to their large sizes. The largest landslide triggered by the Lushan earthquake is the Pingtou rock avalanche (or double rock avalanches) that occurred in the Tangjiagou gully, Damiao village, Baoxing county, lying on a ridge (30°10′43″N, 102°45′40″E). The rock avalanche, with a total volume about 1.5×10 6 m 3 , originated from two connected parts, formed two movement paths and finally gathered together in the accumulated area (Xu et al., 2015).
Since almost all the 6,982 small landslides of area < 100 m 2 were detected on the aerial photographs with high resolution of 0.2 or 0.6 m, we exclude those small ones in the subsequent analysis. The 6,982 landslides only cover 0.33 km 2 and moved 0.185×10 6 m 3 materials. Although the remaining 15,546 large landslides (area ≥100 m 2 , see the Supplementary material) only accounted for 69% of the total coseismic landslide number, they dominated the landslide area and volume, which are 18.55 km 2 and 41.375×10 6 m 3 , or 98.25% and 99.55% of the total, respectively. Therefore, the subsequent spatial analyses with only the large landslides can reflect the predominant situation of the Lushan earthquake-triggered landslides. The curves depicting the correlations between the cumulative landslide number and the landslide area related to three landslide inventories were prepared (Fig. 5). The first is the landslide database obtained from the aerial photographs (contains 14,580 landslides), the second is from Rapideye or ZY-3 images (contains 7,948 landslides), and the third is the final merged inventory of the first two inventories (contains 22,528 landslides).
Although all three curves bend towards horizontal in small landslides areas, the curve related to the aerial photographs bends towards horizontal when the area is smaller than about 10-20 m 2 , whereas the curve related to the satellite images bends so when the area is smaller than about 100-200 m 2 . This difference may be attributed to the omission of small-scale landslides in the inventory based on satellite images. Such omission of landslides with area < 200 m 2 is perhaps due to coalescing small landslides mapped as one large landslide, or the difficulty to observe smaller landslides on the images with a resolution of about 5 m. Therefore, we selected 100 m 2 as the landslide area threshold, which means only landslides of area ≥ 100 m 2 are used in the following analyses.

Correlation of landslide density with controlling factors
To draw the curves of statistical correlations between controlling factors and the three indexes of landslide abundance on each panel, the units of LNA, LAA, and LVA are set to be km 2 or km −2 , ‰, and mm, respectively.
The topographic controlling factors include elevation, slope angle, slope aspect, slope position, and distance from drainages. The geologic factors refer to lithology types and distance from faults. The seismic factors are PGA, seismic intensity, distance from the epicenter, distance from the probable seismogenic fault, and distance along the probable seismogenic fault.

Topographic factors
The correlations between the five topographic factors and three landslide abundance indexes are displayed in Fig. 6. The elevations of the study area range from 543 to 4,852 m, while almost of the whole area (4,977 km 2 , 92.2% of the total) are lower than 3,000 m. All three curves in each graph show similar tendencies, which indicates that most landslides of different scales are roughly uniformly distributed in areas with different elevation intervals and all the three landslide abundance indexes can be used to carry out spatial distribution analysis. The elevation interval between 1,000 and 1,500 m has the largest landslide abundances, which are 4.59 km −2 in LNA, 5.61‰ in LAA, and 13.21 mm in LVA (Fig. 6a). The landslide abundance values gradually decrease at the elevations higher than 1,500 m.
The slope angle is an essential impact parameter to coseismic landslides occurrence. In general, the higher the slope angle, the higher the probability of earthquake-triggered landslides. The slope angle of the study area ranges from 0° to 82° which is divided into 11 classes with an interval of 5°. Slope angles of most of the study area (4,182 km 2 , 77.5%) are between 10° and 40° (Fig. 6b). It seems that the three landslide abundance values increase with the slope angle, particularly for slope angles exceeding 30°. The maximum LNA, LAA, and LVA occurred at slope angles higher than 50°, and their values are 5.57 km −2 , 9.03‰, and 20.94 mm, respectively. Different slope aspects may have different effects on earthquake-triggered landslides, because they have varied responses to the slip direction of the seismogenic fault or the propagating direction of seismic waves (Xu et al., 2014c). Probably, there are differences in vegetation covers, sunlight and evaporations, precipitations, and soil conditions on different slope aspects (Kamp et al., 2008;Yalcin, 2008). The slope aspects in the study are divided into nine kinds, i.e. flat, north (N), northeast (NE), east (E), southeast (SE), south (S), southwest (SW), west (W), and northwest (NW). The slope aspect of SE has the largest area: 923 km 2 , accounting for 17.1% of the entire study area (Fig. 6c). The slope aspect of SW has the largest LAA and LVA (4.08‰ and 9.52 mm), whereas the maximum LNA occurred at the slope aspect of NE (3.27 km −2 ). Different from several other earthquakes such as the 2008 Wenchuan (Huang and Li, 2009;, 2010 Haiti (Xu et al., 2014c), and 2010 Yushu  earthquakes, no single slope aspect had an absolutely high landslide abundance for the Lushan earthquake. A possible reason is that the Lushan earthquake was generated by a blind reverse fault (Xu et al., 2013e). A comparison of coseismic landslides triggered by blind fault earthquakes and surface rupture events shows that the former may trigger abundant landslides in a larger area (Xu, 2014b), which may make the slope aspect effect less obvious.
Landslide abundances generally decrease with increasing distance from the drainage, except for the distance greater than 2,000 m. The slope position and distance from drainages have somewhat similar meanings (Xu et al., 2014c). Ridges generally have long distance from drainages, thus the area of over 2,000 m distant from drainages has higher landslide abundances. Fig. 6 here.

Geologic factors
Lithology is generally considered as an important factor of landslide occurrence, for both earthquake-triggered and non-seismic ones. In the study area, the lithology classes of Cretaceous (K), Devonian (D), and intrusive rocks cover the largest areas, which are 906, 760, and 917 km 2 , or 16.8%, 14.1%, and 17%, respectively. The class of intrusive rocks has the highest LNA, LAA, and LVA values, which are 5.38 km −2 , 7.78‰, and 17.25 mm, respectively (Fig. 7a). Such a phenomenon is also observed in the Wenchuan earthquake-triggered landslides , probably because many intrusive rocks crops out along the Longmenshan fault, and the movement of the fault would have weakened the strength of these rocks. Similar tectonic weakening effects have also been noticed in previous studies (Kellogg, 2001;Fisher, 2003;Korup, 2004;Osmundsen et al., 2009;Shroder and Weihs, 2010).
Distance from faults or lineaments is often used in non-seismic landslide susceptibility assessments (Yilmaz, 2009;Demir et al., 2013;Pourghasemi et al., 2013), and it is considered that landslide abundances A C C E P T E D M A N U S C R I P T

ACCEPTED MANUSCRIPT
24 generally decrease with increasing distance from faults. We extracted faults in the study area from the three sheets of 1:200,000 geologic maps. We divided six classes of distance from faults in 100 m belts. The result shows no clear correlation between them and even an opposing tendency is observed (Fig. 7b). All the maximum LNA, LAA, and LVA occur at distances over 500m from faults, which are 3.06 km −2 , 3.7‰, and 8.41 mm, respectively. The possible reason is that coseismic landslides are controlled by the seismogenic faults rather than other faults in the area. We also noticed that the class of > 500 m distance from faults covers the largest area, which is 4,024 km 2 or 74.6% of the whole study area. The large differences among the areas of these classes of distances from faults probably resulted in a low statistical significance.

Seismic factors
We selected PGA, seismic intensity, distance from the epicenter, distance from the probable seismogenic of PGA ≥ 0.2 g are much higher than those in lower PGA classes (Fig. 8a). The total area of the PGA classes of ≥ 0.2 g covers 2,591 km 2 , and only accounts for about 48% of the total. However, the total number, area, and The seismic intensity information of the study area was extracted from the -Seismic Intensity Map of the 2013 Lushan Earthquake‖ produced by the China Earthquake Administration (CEA, 2013; Fig. 2). The areas of the seismic intensity zones increase with the decreasing seismic intensity (Fig. 8b). In general, LNA, LAA, and LVA increase as the seismic intensity increases except for the VIII zone, probably due to the area near the Baoxing County, located in the VII zone but with a steep topography and thus higher landslide susceptibility.

A C C E P T E D M A N U S C R I P T
The maximum landslide abundances occur in the IX zone, which are 10.72 km −2 , 10.77‰, and 25.02 mm, respectively.
We also constructed 68 buffers with 1 km distance intervals from the epicenter (30.3°N, 103°E). The outer belts were truncated where they intersect with the boundary of the study area. Generally, the landslide abundance values of the areas 40 km or less from the epicenter are much higher than other areas (Fig. 8c).
Although Therefore, an approximate average value of 214° was selected as the strike direction of the probable seismogenic fault of the Lushan earthquake. The epicenter (30.3°N, 103°E) was selected as the original point to construct belts that are parallel or perpendicular to the probable seismogenic fault. In other words, two groups of belts that strike in 214° and 124° were constructed. Considering the shape of the study area, the width of the 214° belts is set to be 1 km, whereas the width of the 124° belts was set to be 2 km. The 214°-trending 1-km-width belts were numbered from northwest to southeast as 1 to 68 (Fig. 8d) and the epicenter is located between the belts 47 and 48. The landslide abundances in the belts 24-54 are much higher than those in other areas (Fig. 8d). This 31 km wide, northeast-southwest trending area registers most of the coseismic landslides: 13,104 in number, 15.88 km 2 in area, and 35.93 × 10 6 m 3 in volume, or 84.3%, 85.7%, and 87.1% of the total, respectively. The 124° 2-km-width belts from southwest to northeast were numbered 1 to 51 (Fig. 8e) and the epicenter is located between the belts 32 and 33. The landslide abundances in areas of the belts 17-45 (29 belts)

ACCEPTED MANUSCRIPT
27 are much higher than those in other areas (Fig. 8e). This 58 km wide, northwest-southeast trending area registers most of the coseismic landslides: 14,461 in number, 17.02 km 2 in area, and 38.01×10 6 m 3 in volume, or 93%, 91.9%, and 92.1% of the total, respectively.

Lithology, topography, and coseismic landslides
Lithology and slope angle are considered to play important roles in occurrence of landslides, including both earthquake-triggered and non-seismic ones. It is generally accepted that the steeper the topography, or the lower the rock strength, the higher the probability of earthquake-triggered landslides. However, the areas where hard rocks crop out generally have steeper topography than trhe areas with soft rocks. Thus, on a regional scale, the rock strength and slope angle may be a couple of contradictory factors affecting landslide occurrence.
However, several previous studies (Korup, 2008;Korup and Schlunegger, 2009;Clarke and Burbank, 2010) pointed out that information of regional slope angle in selected homogeneous rock types can be used as a proxy for landslide susceptibility. Landslides tend to occur on slopes inclined slightly steeper (Gorum et al., 2013).
The peaks of distributions of slope angle are used as the information proxy of slope angle on regional scales (Korup, 2008;Korup and Schlunegger, 2009;Gorum et al., 2013). In order to comprehensively explore the correlations between the two contradictory factors for coseismic landslide occurrence, both peak and mean values of slope angle, LAA in generalized homogeneous lithology combinations were calculated (Table 1) and the values were plotted (Fig. 9a). Unlike previous studies, these points are unexpectedly very scattered, and there is no clear relationship between the mean or peak values of slope angle and landslide area abundance for generalized homogeneous lithology combinations (Fig. 9a). It should be noted that most of the previous studies

ACCEPTED MANUSCRIPT
28 on this issue (Korup, 2008;Korup and Schlunegger, 2009;Clarke and Burbank, 2010) did not focus on landslides triggered by an individual earthquake. Compared with historical or pre-historical landslides on regional scales, landslides triggered by an individual earthquake were not only affected by slope and lithology type, but also strongly affected by earthquake sources. Therefore, we normalized the landslide area abundance values by PGA values and distance from the epicenter. The method for normalizing is to take the mean and peak values divided by the PGA values (the real values from 0.08 g to ≥ 0.52 g are substituted by the class number from 1 to 12) or the distance from the epicenter in homogeneous lithology types. In this way, we can eliminate the effects of the earthquake sources on LAA in homogeneous lithology types. However, the resulting points are still very scattered, without any obvious correlations (Fig. 9b,c). Therefore, we suggest that although the information of slope angle of homogeneous lithology types can represent the rock strength, it may not represent the landslide susceptibility of the lithology type.

Imbricated reverse and/or thrust seismogenic faults and coseismic landslides
There Similar to the Wenchuan event, the spatial distribution of the Lushan earthquake-triggered landslides can be explained using the spatial distribution of the structures in the affected area. The energy may have been released not only by the blind reverse seismogenic fault (Xu et al., 2013e), but also by the neighboring Shuangshi-Daichuan fault , as well as other sub-faults parallel to the seismogenic fault in the vicinity.
Besides the 2008 Wenchuan earthquake and the 2013 Lushan earthquake, the 1999 Chi-Chi earthquake

ACCEPTED MANUSCRIPT
30 also shows similarity of the fault structure, topography, and coseismic landslides. The Chelungpu fault, the seismogenic fault of the Chi-Chi event, ruptured the surface for over 90 km (Chen et al., 2001;Lee et al., 2002).
Most of the coseismic landslides, however, occurred in the area between the Shuangtung, Shuilikeng, and Lishan faults (Liao and Lee, 2000;Wang et al., 2002), which form a typical imbricated fold-and-thrust belt together with the Chelungpu fault. Although these faults did not rupture during the earthquake, they may also contribute to the coseismic landslide occurrence by absorbing seismic energy (Wang et al., 2000) or releasing stress.
Based on the correlations between the regional tectonics and the coseismic landslides related to the three earthquakes (Wenchuan, Lushan, and Chi-chi), we suggest that earthquakes that occurred on an imbricated reverse fault system may be affected by multiple sub-faults, and therefore the coseismic landslides would occur along the imbricated fault belt, rather than only one sub-fault (Fig. 10). The spatial distribution patterns of coseismic landslides would be affected by the differences of earthquake energy release on different sub-faults of the imbricated reverse fault system.

Landslide triggering ability of the blind reverse fault during the Lushan earthquake
The coseismic landslide distribution area is often used to measure the landslide triggering ability of an earthquake. Here the correlation between earthquake magnitudes and distribution areas of coseismic landslides of earthquakes worldwide is presented (Fig. 11), where the solid and dashed lines are the two upper margins determined by Keefer (1984) and Rodriguez et al. (1999). Almost all earthquakes are located under the dashed line from Rodriguez et al. (1999) except for the 1988 Saguenay, Canada, Mw 5.8 earthquake and the 2008 Wenchuan, China, Mw 7.9 earthquake, with earthquake-triggered landslides distribution areas of 45,000 km 2 (Rodriguez et al., 1999) and 110,000 km 2 , respectively. The coseismic landslide distribution

ACCEPTED MANUSCRIPT
31 area related to the Lushan earthquake is about 5,400 km 2 . This event is located beneath, but very close to the upper margin drawn by Keefer (1984). Although the point of the Lushan event is beneath the two proposed upper margins, the area is larger than the most of the earthquakes with similar magnitudes. However, the absolute value of coseismic landslide distribution area related to an earthquake is difficult to obtain, since sometimes there will be a few coseismic landslides on high landslide susceptibility slopes very far away from the earthquake source (Alfaro et al., 2012;Jibson and Harp, 2012;Xu et al., 2014b,d). Therefore, the landslide area and landslide number are also used to compare the landslides triggered by the Lushan earthquake with those by other earthquakes. The published information of landslide area related to individual earthquakes is much less than the landslide distribution area. We collected the data of 13 earthquakes from the literature (Harp et al., 1984;Pearce and O'Loughlin, 1985;Harp and Jibson, 1995;Murphy, 1995;Liao and Lee, 2000;Jibson et al., 2004b;Sato et al., 2005Sato et al., , 2007Wang et al., 2007;Kamp et al., 2008;Yagi et al., 2009;Sepúlveda et al., 2010;Gorum et al., 2014) and our previous studies (Xu, 2014a(Xu, , 2014bXu et al., 2014b, c, d) (Fig. 12). Then, we correlated the total landslide areas with earthquake magnitudes based on these 13 individual earthquake events, such as the 1929 Murchison, New Zealand, Mw 7.7 earthquake (Pearce and O'Loughlin, 1985) and the 2013 Minxian, China, Mw 5.9 earthquake (Xu et al., 2014d).
The points are drawn and the power-law regression curve is produced using the least square method (Fig. 12a).
The curve has an R 2 value of 0.6914. The point for the Lushan earthquake-triggered landslides is above the regression curve. Based on the line, the coseismic landslide area related to an Mw 6.6 earthquake should be about 10 km 2 , which is about half of the actual landslide area (18.88 km 2 ) related to the Lushan earthquake. The correlation between the landslide number and earthquake magnitudes is also shown (Fig. 12b). The number of the landslides (area ≥ 100 m 2 , 15,546 pieces) triggered by the Lushan earthquake is also above the regression

ACCEPTED MANUSCRIPT
32 line from the 13 events. However, the regression is not very good, with R 2 only 0.1931. This is likely because the landslide number may differ significantly among landslide inventory maps prepared by different interpreters, since a coalescing landslide may be considered as either a single landslide or several landslides, whereas no such differences would occur to the landslide area (Xu, 2014b). In addition, Keefer (1994)  where M o is measured in dyne cm -1 and V is in m 3 . The correlation has been applied to several studies, such as Parise (2000), and also been used to compare results from the formula with actual landslide volumes from detailed landslide databases related to individual earthquakes, such as the 2010 Haiti event (Xu et al., 2014c).
The seismic moment of the Lushan earthquake is 1.02×10 26 dyne cm -1 (Global Centroid Moment Tensor Catalog, http://www.globalcmt.org/CMTsearch.html, last accessed in July 2014). Based on the regression formula, the total coseismic landslide volume of the Lushan earthquake can be calculated as 12.84×10 6 m 3 (9.52×10 6 -17.32×10 6 m 3 ). The volume calculated from our detailed coseismic landslide inventory and the -area to volume‖ scaling relationship (Larsen et al., 2010) suggest the volume of landslides triggered by the Lushan earthquake is 41.56×10 6 m 3 , which is about three times that from the regression formula (Keefer, 1994).
Therefore, the distribution area, areal area, number, and volume of the landslides triggered by the Lushan earthquake are higher than the values from most of earthquakes with similar magnitudes or calculated by the regression formula. This indicates the Lushan earthquake has higher landslide triggering ability. We infer that the blind reverse seismogenic fault (Xu, 2014b), the imbricated structure in the earthquake source area, the steep and rugged topography, and a fractured and densely jointed lithology resulted from prolonged tectonic A C C E P T E D M A N U S C R I P T

ACCEPTED MANUSCRIPT
57 A C C E P T E D M A N U S C R I P T

ACCEPTED MANUSCRIPT
58 Table 2 Generalized lithology combinations in the study area.