Professional Documents
Culture Documents
Abstract: Soil particle roundness (R) and sphericity (S) are two important intrinsic properties that govern a soil’s mechanical behavior.
Although R and S have well-established mathematical definitions dating back to the 1930s, the values are much more typically estimated
Downloaded from ascelibrary.org by Hacettepe Universitesi on 06/28/16. Copyright ASCE. For personal use only; all rights reserved.
using charts developed in the 1940s and 1950s. The charts, are based on the earlier mathematical definitions. Using these charts, a class of
undergraduate civil engineering students at the University of Michigan were asked to estimate (Rc and Sc ) from images of twenty geologically
and geographically diverse sands spanning a range of actual R and S values. The images were of three-dimensional (3D) assemblies of the
sands as they would be found in images taken remotely or in situ. The students’ estimates were statistically analyzed and compared with
rigorously determined R and S using a recently developed computational geometry algorithm. Overall, the students’ estimates were scattered,
particularly for natural sands exhibiting intermediate values of R and low values of S. On average, the students underestimated Rc and Sc .
Reasons for the diverse responses and underestimates of the actual R and S are proposed. DOI: 10.1061/(ASCE)GT.1943-5606.0001485.
© 2016 American Society of Civil Engineers.
Introduction This paper compares the results of particle sphericity and roundness
estimation from such captured images as performed by individuals
It is well-established that particle roundness (R) and sphericity (S) using standard charts to results obtained by recently developed
affect most macroscale mechanical properties of soil, including computer methods that precisely determine R and S by their rig-
strength, compressibility, small-strain modulus, shear wave veloc- orous mathematical definitions.
ity, packing density, and the critical state parameters. Such relation- It is emphasized at the outset that estimating R and S from
ships have been investigated by Morris (1960), Kolbuszewski images of particle assemblies is not the conventional method
and Fredrick (1963), Eisma (1965), Koerner (1970), Oda (1972), that would examine the physical, single-detached particles.
Holubec and d’Appolonia (1973), Youd (1973), Edil et al. (1975), However, when a physical soil specimen is unavailable, computer
Oda et al. (1978), Sladan et al. (1985), Vepraskas and Cassel image analysis and visual comparison to charts are the only
(1987), Moroto and Ishii (1990), Sukumaran and Ashmawy options.
(2001, 2003), Yasin and Safiullah (2003), Santamarina and Cho
(2004), Cho et al. (2006), Rouse et al. (2008), Bareither et al.
Definitions of Roundness and Sphericity
(2008), Azami et al. (2010), Maeda et al. (2010), Cavarretta et al.
(2010), Chapuis (2012), Cabalar et al. (2013), Shin and Santamar- Wadell (1932) defined roundness as the ratio of the average radius
ina (2013), Kandasami and Murthy (2014), Oh et al. (2014), and of curvature of the corners of a particle (ri where i = corner num-
many others. Another body of literature related to aggregate in ber) to the radius of the maximum inscribed circle (rins ), as shown
pavements is summarized by Masad et al. (2007). It reveals the in Fig. 1. It is the most common definition of R and has been used
critical effects that aggregate shape (sphericity) and roundness have widely in a variety of disciplines (Eisma 1965; Edil et al. 1975;
on the performance of concrete and hot-mix asphalt mixtures. Frossard 1979; Vepraskas and Cassel 1987; Sagga 1993; Santamar-
With rapid advances in the ability to obtain high resolution im- ina and Cho 2004; Chapuis 2012; Shin and Santamarina 2013;
ages of geomaterials remotely, for example, by airborne drones or Cabalar et al. 2013; Kandasami and Murthy 2014).
in situ using subsurface probes such as the Vision Cone Penetrom- Krumbein and Sloss (1951) defined sphericity (S) as the ratio of
eter or even in the laboratory following sedimentation (Ohm and particle width (d2 ) to particle length (d1 ) as shown in Fig. 1. Other
Hryciw 2014), a need has developed to detect and analyze grain definitions of S have been proposed by Cox (1927), Tickell (1931),
shapes from images of three-dimensional (3D) particle assemblies. Wadell (1933), Riley (1941), Kuo and Freeman (2000), Santamar-
ina and Cho (2004), Cho et al. (2006), and Altuhafi et al. (2013).
1
Professor, Dept. of Civil and Environmental Engineering, Univ. of However, Zheng and Hryciw (2015) found the Krumbein-Sloss
Michigan, 2366 GG Brown, Ann Arbor, MI 48109-2125 (corresponding definition to be the most practical and easy to determine. For these
author). E-mail: romanh@umich.edu reasons, this paper uses the Wadell (1932) definition of R and the
2
Graduate Student Research Assistant, Dept. of Civil and Environmen- Krumbein-Sloss (1951) definition of S.
tal Engineering, Univ. of Michigan, 2340 GG Brown, Ann Arbor, MI
48109-2125. E-mail: junxing@umich.edu
3
Asssitant in Research, Dept. of Civil and Environmental Engineering, Chart Methods for Estimating Roundness
Univ. of Michigan, 2340 GG Brown, Ann Arbor, MI 48109-2125. E-mail: and Sphericity
kshetler@umich.edu
Note. This manuscript was submitted on April 15, 2015; approved on Determinations of R and S require considerable manual effort. In
December 16, 2015; published online on April 26, 2016. Discussion period Wadell’s (1935) original procedure for determination of R, the par-
open until September 26, 2016; separate discussions must be submitted for ticle outline was traced on paper and compared with transparent
individual papers. This paper is part of the Journal of Geotechnical and templates to find the maximum-sized circle that would fit inside
Geoenvironmental Engineering, © ASCE, ISSN 1090-0241. each corner. The maximum inscribed circle was found similarly.
(a) (b)
Fig. 1. Definitions of (a) roundness (R); (b) sphericity (S)
Fig. 2. Chart for estimating particle roundness (Rc ) [reprinted from Krumbein 1941, with permission from SEPM (Society for Sedimentary Geology)]
To accelerate this process and reduce the required effort, standard provides 20 reference particles having S from 0.3 to 0.9 and R from
charts were developed by Krumbein (1941), Krumbein and Sloss 0.1 to 0.9, both in increments of 0.2. The Krumbein–Sloss chart
(1951), and Powers (1953) to estimate R and S. These charts are pointed out that R and S are independent parameters because
shown in Figs. 2–4. one could vary without change in the other. The Krumbein–Sloss
The Krumbein (1941) chart, shown in Fig. 2, contains 81 chart may be the most widely used chart because it provides an
particles redrawn from pebbles for which R had been determined opportunity to simultaneously estimate both R and S. The Powers
by Wadell’s method. Krumbein placed the particles in 9 bins with R (1953) chart separates soil particles having Wadell roundness val-
ranging from 0.1 to 0.9 in increments of 0.1. Several broken par- ues from 0.12 to 1.00 into six classes, as shown in Fig. 4. The ratio
ticles are also shown in the figure to indicate the effects of breakage of the upper limit to the lower limit of R in each class is approx-
on R. Compared with the other two charts, Krumbein’s provides imately 1.4. Each R range is exemplified by two particles: one with
more reference particles and thus more opportunities for visual high S and one with low S. The Powers chart may have inspired the
comparisons. The Krumbein–Sloss (1951) chart, shown in Fig. 3, word descriptions of R in the current ASTM D2488 (ASTM 2009)
Fig. 4. Chart for qualitatively describing roundness (Rc ) and sphericity (Sc ) [reprinted from Powers 1953, with permission from SEPM (Society for
Sedimentary Geology)]
Fig. 5. Comparison of Wadell’s manual method and computational geometry results for roundness (reprinted from Zheng and Hryciw 2016, with
permission)
maximum inscribed circles are also displayed in Fig. 5, in which the sands were sieved to a narrow size range to create relatively
Wadell’s results are in millimeters whereas Zheng and Hryciw’s are uniform-sized specimens. Each was prepared on a flat surface.
in pixel units. Comparisons of results show that the positions and A 36.3-megapixel Nikon D800 SLR camera with different magni-
sizes of the circles were identical; the computer-generated circles fying lenses, which could provide 2, 4, and 8× magnifications, was
were perfectly fitted to the corners. As a result, the values of R by used to photograph the soils. Each image was 7,360 by 4,912 pix-
the two methods were also in excellent agreement. els. For this paper, small square patches were cropped from each
Zheng and Hryciw (2016) also computed Wadell’s R value image and compiled in Fig. 6. Each of these patches represents be-
for the particles shown in the three standard charts in Figs. 2–4. Their tween 20 and 25% of a full image. Table 1 provides the origin,
results were again remarkably similar to the Rc and Sc values in the specific gravity, D50 , Cu , color, and translucency of the original
original charts. Readers can refer to Zheng and Hryciw (2015, 2016) collected soils. It also lists the number of fully visible particles
for details of the computational algorithms and comparisons of the in each image, the corresponding D50 and Cu values of this sub-
computationally determined R and S values with the chart values, group, and their average computed R and S values.
Rc and Sc .
13. Muskegon, MI 14. Nevada Sand 15. New Madrid, MO 16. Oakland Co., MI
course were asked to estimate the Rc and Sc values of the 20 Results of Chart-Based Estimates of R c and S c
soils using the Krumbein (1941) and Krumbein-Sloss (1951)
Histograms of the chart-based estimates of Rc and Sc for each soil
charts. Each student was provided with a booklet containing 20 are shown in Figs. 7 and 8, respectively. The estimates follow
(8.0 × 10.5-in.) color images of the specimens and the two normal distributions as shown by the solid curves superimposed
charts. The Sc values were determined using the Krumbein–Sloss over the histograms. The mean values of the estimates are denoted
chart. Although Rc could be determined from either chart, the students by Rc and Sc and their standard deviations (σR and σS ) were com-
were encouraged to use Krumbein’s to obtain more precise Rc esti- puted from the normal distributions.
mates. Only one significant digit was to be reported for both Rc The standard deviations quantify the spreads in the chart-based
and Sc . estimates. As such, they reflect the uncertainties in the students’
Because the particles were in 3D assemblies, some showed a full estimates of Rc and Sc . Fig. 9(a) plots Rc σR with increasing
outline of their projected perimeter although others were partially Rc , and Fig. 9(b) shows Sc σS with increasing Sc for each soil.
occluded by foreground particles. The students were told to use The figures also show the full range of estimates from Rcmax to Rcmin
only the particles whose outlines were fully visible to make their and from Scmax to Scmin . Except for the glass beads, the estimated Rc
estimates. The number of particles per image to be considered by and Sc values are quite scattered.
the students for evaluation was not specified. However, the students
were encouraged to consider as many particles as they felt were
necessary to obtain good estimates of Rc and Sc . The students were Discussion of Chart-Based Estimates of R c and S c
given one hour of a two-hour lab period to complete the assign- Before presenting the computer-based results, this paper pauses to
ment. All of the students worked independently. analyze the students’ estimates alone. Having observed the spreads
Fused aluminum oxide Coarse commercial abrasive Brown mottled, opaque 3.86 1.80 1.6 125 2.20 1.2 0.30 0.69
Michigan crushed gabbro Crushed igneous rock Gray to brown, opaque 3.30 0.80 5.5 134 1.50 2.3 0.23 0.62
Glass beads Synthetic silica spheres Translucent and reflective 2.53 0.97 1.1 125 0.97 1.1 0.96 0.93
Griffin, Indiana Quarried glacio-fluvial Multicolored, opaque 2.66 0.74 4.3 198 1.02 1.5 0.54 0.74
deposit
Chesterton, Indiana beach Recent beach sand Multicolored, opaque 2.65 0.64 2.9 120 0.86 1.7 0.56 0.72
Chesterton, Indiana dune Recent eolian deposit brown to white, 2.68 0.26 1.3 125 0.28 1.1 0.62 0.76
semi-opaque
Michigan dunes Recent eolian deposit Brown to white, 2.68 0.30 1.5 119 0.40 1.1 0.52 0.78
semi-opaque
Muskegon, Michigan Lake Michigan beach sand Multicolored, semi-opaque 2.66 0.40 1.6 159 0.52 1.2 0.56 0.77
Nevada sand Fine eolian silica dune sand White, semi-translucent 2.65 0.16 1.3 169 0.20 1.1 0.62 0.76
New Madrid, Missouri Old Mississippi river Multicolored, semi opaque 2.59 0.32 2.2 102 0.42 1.4 0.57 0.78
alluvium
Oakland County, Michigan Glacial Tan to multicolored, 2.63 0.31 1.6 101 0.39 1.5 0.56 0.77
semi-opaque
Ottawa 20–30 Silica mined from St. Peter’s White, semi-translucent 2.65 0.71 1.4 139 0.75 1.1 0.75 0.85
Sandstone
Rincon, New Mexico Recent roadside eolian White to tan, 2.61 0.36 3.0 117 0.62 1.5 0.49 0.73
deposit semi-translucent
Scotts Valley, California Weathered and eroded White, opaque 2.85 0.33 1.5 108 0.43 1.1 0.45 0.74
seabed sandstone
Upper Peninsula, Michigan Glacial Brown, opaque 2.72 0.60 2.8 112 0.78 1.6 0.47 0.73
and standard deviations in the students’ chart-based estimates of studied only a few particles, whereas more diligent students inspected
Rc and Sc , it is useful to consider why they occurred. There are more.
several factors that may have contributed including time constraints
and task fatigue, diligence, analyzing images of 3D particle assem-
Analyzing Images of 3D Particle Assemblies
blies, and the ability to resolve chart differences by eye. We con-
sider each of these potential factors in turn. The students were provided with the same images as the computer
analyzed. Furthermore, they were instructed to focus their attention
on the particles whose full perimeters were in view, as these are the
Time Constraints and Task Fatigue
particles that the computer would be analyzing. As such, despite the
All students were given an hour to complete the assignment fact that the particles were in a 3D assembly, the students were
and were not permitted to leave early. Therefore, there was neither observing exactly the same particles in precisely the same orienta-
incentive nor pressure to finish in a shorter period of time. tion as the computer did. Theoretically, the other particles in the
They were permitted to take breaks as needed and none reported background should have no effect on the student’s estimates of
having inadequate time to complete the task. Virtually all of the Rc and Sc , however, it is very difficult to avoid being distracted
students finished the assignment in under an hour then waited by the surroundings of the particle being analyzed. It is also pos-
for the lab section to be dismissed. This suggests that time was sible that students’ attention was drawn to the more unusual par-
not a factor. ticles and, therefore, their estimates could have been more heavily
influenced by such odd-shaped particles.
Diligence
Ability to Resolve Chart Differences by Eye
The students were instructed to look at as many particles as necessary
to obtain good estimates. They were told that the assignment would be The standard deviations in the students’ Rc estimates were smallest
graded and that the students whose estimates of Rc and Sc were, on for soils having either very high or very low particle roundness.
average, closest to the computer-determined R and S would obtain This is most easily observed in Fig. 9(a). By contrast, σR and
small prizes (approximately 25% of the students later received gift the overall ranges of estimates were largest for soils in the inter-
certificates to a local coffee shop). However, these incentives may mediate (Rc ¼ 0.4 to 0.6) range. This may be explained by refer-
not have been sufficient. Disinterested students are likely to have ence to Figs. 2 and 3, which show that it is easier to visually
8 8 8 8
4 4 4 4
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
20 20 20 50
5. Fort Davis, TX 6. Fused Aluminum Oxide 7. Michigan Crushed
16 R c = 0.20 16 R c = 0.25 16 Gabbro 40 8. Glass Beads
σR = 0.09 σR = 0.11 R c = 0.22 R c = 0.91
Downloaded from ascelibrary.org by Hacettepe Universitesi on 06/28/16. Copyright ASCE. For personal use only; all rights reserved.
12 12 12 σR = 0.09 30 σR = 0.02
8 8 8 20
4 4 4 10
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
20 30 20 20
Number of Responses
9. Griffin, IN 10. Chesterton, IN Beach 11. Chesterton, IN Dunes 12. Michigan Dunes
16 25 16 R c = 0.58 16 R c = 0.50
R c = 0.55 R c = 0.50
σR = 0.10 20 σR = 0.09 σ = 0.14 σR = 0.11
12 12 R 12
15
8 8 8
10
4 5 4 4
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
20 20 20 20
13. Muskegon, MI 14. Nevada Sand 15. New Madrid, MO 16. Oakland Co., MI
16 R c = 0.54 16 R c = 0.56 16 R c = 0.45 16 R c = 0.50
σ = 0.12 σR = 0.14 σR = 0.11 σR = 0.10
12 R 12 12 12
8 8 8 8
4 4 4 4
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
20 20 20 20
17. Ottawa 20-30 18. Rincon, NM 19. Scotts Valley, CA 20. Upper Peninsula, MI
16 R c = 0.80 16 R c = 0.34 16 R c = 0.37 16 R c = 0.45
σR = 0.07 σR = 0.09 σR = 0.11 σR = 0.18
12 12 12 12
8 8 8 8
4 4 4 4
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
c
Chart-based Roundness, R
distinguish very angular from angular particles and rounded from shape. As a particle’s elongation increases, estimating the ratio
well-rounded than it is to distinguish subangular from subrounded of its dimensions becomes more difficult. Furthermore, as the
or subrounded from rounded. average sphericity of a sand decreases, the range of actual sphe-
There is a general trend of increasing σS and range of Sc with ricities of individual particles within the specimen also surely
decreasing S. Sands with very spherical particles are relatively increases. Therefore, an average value becomes more difficult
easy to identify and they tend to also be more uniform in to estimate.
8 8 8 8
4 4 4 4
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
20 20 20 80
5. Fort Davis, TX 6. Fused Aluminum Oxide 7. Michigan Crush Gabbro
16 S c = 0.68 16 S c = 0.63 16 S c = 0.37 8. Glass Beads
60
σS = 0.15 σS = 0.15 σS = 0.15 S c = 0.91
Downloaded from ascelibrary.org by Hacettepe Universitesi on 06/28/16. Copyright ASCE. For personal use only; all rights reserved.
12 12 12 σS = 0.02
40
8 8 8
20
4 4 4
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
20 20 20 25
Number of Responses
9. Griffin, IN 10. Chesterton, IN Beach 11. Chesterton, IN Dunes 12. Michigan Dunes
16 S c = 0.62 16 S c = 0.58 16 S c = 0.69 20 S c = 0.69
σS = 0.13 σS = 0.14 σ = 0.14 σS = 0.08
12 12 12 S 15
8 8 8 10
4 4 4 5
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
20 20 20 20
13. Muskegon, MI 14. Nevada Sand 15. New Madrid, MO 16. Oakland Co., MI
16 S c = 0.74 16 S c = 0.77 16 S c = 0.59 16 S c = 0.61
σS = 0.12 σS = 0.12 σS = 0.14 σS = 0.10
12 12 12 12
8 8 8 8
4 4 4 4
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
20 20 20 20
17. Ottawa 20-30 18. Rincon, NM 19. Scotts Valley, CA 20. Upper Peninsula, MI
15 S c = 0.83 16 S c = 0.62 16 S c = 0.64 16 S c = 0.67
σS = 0.07 σS = 0.12 σS = 0.13 σS = 0.13
12 12 12
10
8 8 8
5
4 4 4
0 0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
c
Chart-based Sphericity, S
Computational Geometry Results for R and S are shown in white in Fig. 10(a). The computer program automati-
cally identifies the labeled particles (white areas) and computes the
The same 20 images that were provided to the students were also R and S for each. The fitted maximum inscribed circles (shown in
evaluated for R and S by computer. In each image, all particles with dashed lines) and corner circles (shown in solid lines) are super-
full perimeters in view were manually selected and labeled. For imposed over the original image in Fig. 10(b). Two select areas
example, the 124 particles in the image of the Brady, TX sand are enlarged in Figs. 10(c and d).
Fig. 9. Summary of students’ chart-based estimates for (a) roundness (Rc ); (b) sphericity (Sc )
Figs. 11 and 12 show the cumulative R and S distributions for and Glass Beads) have roundness values 0.67 ≤ R̄ ≤ 0.96 and
each soil. The number of particles that were used from each soil are of high sphericity (0.78 ≤ S̄ ≤ 0.93). Finally, the natural sands
image is provided in Table 1. The mean sphericity values (S̄) span a wide range of R (from 0.30 to 0.70) but a very narrow range
are plotted versus the mean roundness (R̄) for the 20 sands in of S (from 0.70 to 0.80).
Fig. 13. The R̄ values fall in a range approximately between
0.20 and 0.95; the S̄ values fall between approximately 0.60 and
0.95. The crushed sands (Michigan 30A, Fused Aluminum Oxide, Evaluation and Comparison of Chart-Based
and Michigan Crushed Gabbro) are typically angular (0.20 ≤ Estimates and Computed Values of Roundness
R̄ ≤ 0.30) and of relatively low sphericity (0.60 ≤ S̄ ≤ 0.70). This
is attributed to the brittle fracturing in formation of these materials. To quantify the spread in the computed values of R for each
The rounded and well-rounded sands (Brady TX, Ottawa 20-30, soil individually, this paper defines roundness uniformity as
Fig. 10. Identification of 124 Brady, TX sand particles with full perimeter projections and fitting of corner circles and maximum inscribed circles
20 20
0 0
1.0 0.8 0.6 0.4 0.2 0.0 1.0 0.8 0.6 0.4 0.2 0.0
(a) R (b) R
100 100
Chesterton, Oakland CO., MI
Particles with smaller R (%)
IN Dunes (125)
(101)
80 Michigan Dunes 80 Ottawa 20-30
(119)
(139)
Muskegon, MI Rincon, NM
60 (159) 60 (117)
Nevada Sand Scotts Valley, CA
(169)
(108)
40 New Madrid, MO 40 Upper Peninsula, MI
(102) (112)
20 20
0 0
1.0 0.8 0.6 0.4 0.2 0.0 1.0 0.8 0.6 0.4 0.2 0.0
(c) R (d) R
U R ¼ R90 –R10 , where R90 ¼ R value from a distribution curves at 10% (90 and 10%, respectively) of particles are excluded from de-
the 90% level; and R10 = value from the distribution curves at the fining the spread because these particles often have exceptional
10% levels. The 10 and 90% thresholds coincide with the levels at shapes compared with the majority of the particles, and therefore,
which the R distribution curves, shown in Fig. 12, often transition the lengths of the tails at both ends of the curve could be controlled
from being relatively flat to more steep. The highest and lowest by just a few unusual but inconsequential particles.
20 20
0 0
1.0 0.8 0.6 0.4 0.2 0.0 1.0 0.8 0.6 0.4 0.2 0.0
(a) S (b) S
Downloaded from ascelibrary.org by Hacettepe Universitesi on 06/28/16. Copyright ASCE. For personal use only; all rights reserved.
100 100
Chesterton, Oakland CO., MI
Particles with smaller S (%)
20 20
0 0
1.0 0.8 0.6 0.4 0.2 0.0 1.0 0.8 0.6 0.4 0.2 0.0
(c) S (d) S
The relationship between U R and R̄ is shown in Fig. 14(a). The (R̄>0.65) soils. On average, σR was 0.09 for very angular soils,
crushed and well-rounded sands generally possessed uniform R 0.13 for subangular to rounded soils (natural sands), and 0.08 for
distributions (U R is approximately 0.2), whereas the natural sands well-rounded soils (excluding the glass spheres). Figs. 14(a and b)
were more variable in roundness and thus had U R ranging from show that U R and σR were similarly related to R̄. As such, a strong
approximately 0.25–0.45. correlation should also be expected between σR and U R , as con-
Various correlations between the students’ chart-based estimates firmed by Fig. 14(c). This logically suggests that chart users have
and the values determined by computational geometry for round- a more difficult time estimating Rc when particles in a specimen
ness are shown in Figs. 14(b–d). The standard deviations (σR ) in the exhibit a larger range of R values. Soil No. 6 Fused Aluminum
students’ chart-based estimates of Rc were plotted against the com- Oxide (FAO) is a unique material because it is manufactured to
puter R̄ in Fig. 14(b). As previously discussed, relatively small σR be very angular and similar in its angularity, as evidenced by its
values were observed for angular (R̄<0.25) and for well-rounded steep cumulative R distribution curve in Fig. 11(b). Because of this
manufacturing process, the FAO does not fit the trend of the other
soils and was left out of the trend bounds in Fig. 14(c).
Finally, Fig. 14(d) compares the mean of the students’ chart-
based estimates of roundness ( Rc ) to the mean values by computa-
tional geometry (R̄). Despite the large σR values in estimates by the
students [as shown in Figs. 14(b and c)], 75% of the sands’ Rc val-
ues were within 0.10 of R̄, and 100% were within 0.15, as indicated
by the dashed lines in Fig. 14(d). The differences were smallest for
angular and very rounded soil particles. For the natural soils, the
differences were larger and the students almost always underesti-
mated roundness. As discussed previously, the authors suspect that
a student’s eye is naturally drawn to the more unusual angular par-
ticles and tends to compare these more angular particles to those in
the charts.
Fig. 14. Summary and comparison of chart-based estimates and computational geometry values for roundness
U S ¼ S90 –S10 , in which S90 and S10 were obtained from the S dis- Practical Significance of Findings
tribution curves. As with the R distribution curves, S10 appears to
typically coincide with the transition from a steep to a flatter portion The results of this study show that chart estimates of Rc using im-
of the curve. However, no such transition occurred at the high S ages of 3D particle assemblies were in error by as much as 0.3 to
end. This is logical as some particles with length to width ratios 0.5 for subangular to subrounded and even rounded natural sands,
approaching 1.0 will always exist in a natural sand. Nevertheless, as observed in Fig. 9. When estimates performed by 38 students in
the authors used S90 for consistency with the previous definition of a laboratory setting were averaged, the Rc values underpredicted R
U R . Fig. 15(a) confirms that less spherical sands display a much by up to 0.15, as shown in Fig. 14(d). Such misestimates of particle
wider range of S (larger U s ) than spherical particles do. As a prac- roundness can be significant, even costly. Santamarina and Cho
tical matter, this indicates that fewer spherical particles are needed (2004) showed that for every 0.10 change in R, the angle of repose,
to establish statistically valid values of S. which is approximately the residual or constant volume angle of
Correlations between the students’ chart-based estimates for sphe- internal friction, changes by approximately 1.7°. Thus, underesti-
ricity and the values determined by computational geometry are mates of R of even 0.15 translate to overestimates of friction angle
shown in Figs. 15(b–d). The σS ’s are plotted versus the mean com- by 2–3° degrees. The implications of the larger, individual mises-
puter value, S̄, in Fig. 15(b). The average standard deviation for the timates of 0.3 to 0.5 can lead to friction angle misestimates of up to
natural sands was 0.13. Just as with roundness, the very strong cor- 5–9° degrees.
relations between U S and S̄ and between σS and S̄ lead to the similarly This study also showed that S of natural sands was underesti-
strong correlation between σS and U S as shown in Fig. 15(c). mated by as much 0.15 or more, as shown in Fig. 15(d). Further-
As expected, these figures indicate that S is more difficult to more, the spreads in Sc estimates, as observed in Fig. 10, are even
estimate by charts when particles are elongated (i.e., low sphericity) larger than they were for Rc. Sands having low sphericity particles
than when the soils are spherical. For the majority of soils exhibit- can exhibit significant fabric and, therefore, strong anisotropic
ing S̄ in a range of 0.7 to 0.8, the σS ranged narrowly between behavior. Azami et al. (2010) showed that the friction angle of fine
approximately 0.12 and 0.15. to medium sand, consisting of elongated crushed limestone par-
Finally, Fig. 15(d) compares the mean chart-based estimates of ticles, varied by 12° depending on the orientation of the particles.
sphericity ( Sc ) with the mean values by computational geometry It would be logical to conclude that, with all other factors being
(S̄). The trends are as expected: Sc diverges very significantly from equal, the range of friction angles would depend on the sphericity
S̄ when soil becomes elongated (low sphericity). The students of the particles. Particles approaching perfect spheres (S ¼ 1.0)
clearly overestimated the elongation of particles. This supports should display virtually no range because such soils would not
the authors’ previous hypothesis that the students’ attentions were exhibit anisotropy. As sphericity decreases, three things actually hap-
drawn to the more unusual particles in the images, even when they pen: undeprediction of Sc becomes larger [observed in Fig. 15(d)];
were less abundant. the range of Sc estimates increases [observed in Fig. 15(b)]; and the
Fig. 15. Summary and comparison of chart-based estimates and computational geometry values for sphericity
range of potential friction angles, because it depends on fabric ori- 3D assemblies (possessing a range of particle shapes) were col-
entation, increases. In other words, the uncertainties are compounded lected, and a group of 38 undergraduate civil engineering students
as S decreases. were asked to estimate the values of Rc and Sc using the chart
method. The results were compared with results by the previously
proven computer methods. Significant differences between the
Discussion computer and student-estimated Rc values were observed for par-
ticles in the mid-range of Rc , between 0.30 and 0.65. The differ-
Attributable to the complex, subjective, and tedious nature of the ences were smaller for very round and angular particles. This is
task, the previously demonstrated accuracy of the computer meth- because very round and angular particles show greater visually per-
ods, and the results of this study, it should be evident that computers ceptible changes with incremental changes in R. Particles in the
will do a better job of determining R and S from 3D assemblies of 0.30 to 0.65 range appear to be more similar to the human eye.
particles than human evaluators could. In the end, computers are For S, the agreement between computer and student-estimated val-
actually computing R and S whereas humans, regardless of their ues was best at the highest values of S and logically decreased as S
experience and training, are estimating these parameters from decreased. For both R and S, the differences between computer and
charts. Nevertheless, the authors are not experts in cognitive sci- student-estimated values increased as the range of actual R and S
ence and do not profess to be conducting research in engineering values increased. Finally, students tended to underestimate both R
education. As such, the results of this study should not be quanti- and S of natural soils. It was hypothesized that when observing an
tatively extrapolated to other populations such as different groups image containing several hundred particles in a 3D assembly, at-
of students, laboratory technicians, and more experienced engi- tention may be drawn to the more unusual (more angular and less
neers or geologists. It would be interesting to repeat this exercise spherical) particles, even when such particles may be less abundant.
with such groups and the results may (or may not) be better The practical implication of misestimating Rc is that the mechanical
(i.e., closer to computer results) than obtained by undergraduate properties of sands could be significantly mispredicted. Computer-
geotechnical engineering students. based methods that analyze images of soils are much faster and can
analyze many more particles to obtain more accurate distributions
and average values of R and S.
Conclusions
Computer algorithms using computational geometry were devel- Acknowledgments
oped to automate and replace the laborious manual task of meas-
uring, or crudely estimating from charts, values of soil particle This material is based on work supported by the U.S. National
roundness, R, and sphericity, S. Images of 20 different soils in Science Foundation under Grant No. CMMI 1300010. Any
Highway Administration, Washington, DC. particles in assemblies by computational geometry.” J. Comput. Civ.
Vepraskas, M. J., and Cassel, D. K. (1987). “Sphericity and roundness of Eng., 10.1061/(ASCE)CP.1943-5487.0000578, 04016021.
sand in coastal plain soils and relationships with soil physical proper- Zheng, J., Hryciw, R. D., and Ohm, H.-S. (2014). “Three-dimensional
ties.” Soil Sci. Soc. Am. J., 51(5), 1108–1112. translucent segregation table (3D-TST) test for soil particle size
Wadell, H. (1932). “Volume, shape, and roundness of rock particles.” and shape distribution.” Proc., Geomechanics from Micro to Macro,
J. Geol., 40(5), 443–451. Taylor & Francis Group, Abingdon, U.K., 1037–1042.