SlideShare a Scribd company logo
1 of 24
Download to read offline
Hyperspectral Imaging Features for Mortar Classification and Compressive Strength
1
Assessment
2
Liang Fan 1, Ming Fan 2, Abdullah Alhaj 1, Genda Chen 1,* and Hongyan Ma 1
3
1
Department of Civil, Architectural and Environmental Engineering, Missouri University
4
of Science and Technology, Rolla, MO 65401, United States; lf7h2@mst.edu;
5
ahanbc@mst.edu; mahon@mst.edu
6
2
Department of Mining and Minerals Engineering, Virginia Polytechnic Institute and
7
State University, Blacksburg, VA 24060; mingfan@vt.edu
8
* Correspondence: gchen@mst.edu; Tel.: +01-573-341-4462
9
Abstract: In this study, hyperspectral imagery with two computational algorithms are
10
proposed to classify the type of mortar and assess the in-situ strength of fresh mortar in near
11
real time. Each scanning on a mortar surface includes 30 spatial pixels selected for analysis,
12
each assigned with a light reflectance spectrum over 400 - 2500 nm. Three groups of mortar
13
samples with a water-to-cement (W/C) ratio of 0.6, 0.5 and 0.4, respectively, were cast and
14
scanned from Day 1 to 14 of curing. Reflectance data at a wavelength range of 1920 nm to
15
1980 nm, associated with the O-H chemical bond, were extracted and averaged to classify
16
the different mortar types with K-Nearest Neighbors (KNN) and Support Vector Machine
17
(SVM) algorithms and to predict their compressive strength from a regression equation. The
18
results showed that the average reflectance increased with time due to water molecules
19
reaction during curing process. The KNN classification model with K=5 had a prediction
20
accuracy of 70% to 75%, and the SVM classification model with C=1000 and ฯƒ=10 showed
21
a prediction accuracy of approximately 90%. Therefore, the SVM classification algorithm
22
is recommended for use in mortar classification. The compressive strength is well correlated
23
with the average reflectance with a coefficient of over 0.98.
24
Keywords: Hyperspectral imaging; W/C ratio; reflectance; KNN; SVM; compressive
25
strength
26
27
1. Introduction
28
Concrete is a mixture of aggregate, sand, cement and water in a certain proportion. The
29
cement and water together, referred to as cement paste, hardens through hydration reactions
30
and binds the aggregate and sand to achieve the strength of concrete in a curing process over
31
time [1]. Concrete hydration process starts immediately after concrete casting. Cement reacts
32
with water to generate hydration products like C-S-H gel and calcium hydroxide. The
33
hydration products grow, interconnect, and bond aggregate and sand. Concrete is formless
34
and shaped to various forms of interest, when newly mixed, and durable, when cured and
35
hardened. In the construction of highway pavements, ACI Code 301-72 requires a minimum
36
of curing period for concrete pavement to ensure that the constructed roadway is safe to
37
traffic without damage [2]. In the repair and resurface of existing roadways and their
38
transportation network in an urban environment, it is imperative to determine the early-age
39
strength of concrete pavements so that the impact of roadway construction on traffic is
40
minimized.
41
Coring and pullout test are two of the conventional approaches that have been used for
42
on-site evaluation of the compressive strength of concrete. With the coring method, concrete
43
cores are acquired by drilling a concrete structure at selected locations, and tested for their
44
compressive strength [3]. During the pullout test, a metal disk is attached to the concrete
45
surface with super glue. After a short curing period, the metal disk is pulled perpendicularly
46
off the surface and the pullout force can be used to calculate the compressive strength of the
47
concrete structure [4]. The pullout force can be related to the compressive strength of
48
concrete based on a pre-determined calibration curve. Both the coring and pullout test are
49
destructive, potentially compromising the integrity of concrete structures.
50
Nondestructive approaches such as the maturity method and the ultrasonic pulse velocity
51
(UPV) have also been used to determine the compressive strength of concrete. The maturity
52
method allows the estimate of early-age compressive strength of in-place concrete in real
53
time. A maturity index as a function of curing time and temperature is determined according
54
to the ASTM C1074 Standards [5]. In applications, a reference strength-maturity curve must
55
be developed for each project-specific material in advance. With the UPV method, the
56
velocity of an ultrasonic pulse that travels through concrete is measured and converted to the
57
strength of concrete based on their pre-determined calibration curve [6]. The field application
58
of this method is limited due to the effects of voids, cracks and steel bars.
59
Hyperspectral imagery has been used to assess various conditions of concrete by imaging
60
a concrete surface and analyzing the light reflectance as a function of wavelength for each
61
pixel in an image. Such a reflectance-wavelength spectrum can be divided into many narrow
62
and continuous wavelength bands for their correlation to specific materials on the concrete
63
surface [7]. By analyzing the change of reflectance values at these prominent bands, different
64
materials can be discriminated and classified. For instance, dark gray, light gray and
65
dolomitic limestone were distinguished in the selection of Portland cement clinkers based on
66
the reflectance variations of carbonate (CO3) and Al-OH in wavelength ranges of 2125โ€“2400
67
nm and 2170โ€“2250 nm, respectively [8]. The carbonation degradation depth of concrete was
68
estimated from reflectance values at a wavelength of 440 nm, 1500 nm, and 2340 nm [9].
69
The total chloride content in mortar specimens was linearly related to the reflectance at a
70
wavelength of approximately 2260 nm [10]. The status of concrete (hydration, curing and
71
hardening) was determined by constructing a logistic regression model with reflectance
72
spectra [11].
73
In the past decade, hyperspectral imaging has also been used to estimate the compressive
74
strength of concrete. For example, a partial least square regression model was developed to
75
establish the relation between concrete strength (7, 14 and 28 days) and its corresponding
76
reflectance over the entire wavelength range [12]. The reflectance spectrum of eight concrete
77
samples with various W/C ratios and curing ages moved upward with an increase of
78
compression strength [13]. In both studies, the mix designs of concrete were not introduced
79
and the relation between the compression strength and the reflectance at a characteristic
80
wavelength range was not clearly interpreted. Three groups of 28-day cured concrete
81
specimens with a W/C ratio of 0.5, 0.65 and 0.8 were differentiable by comparing absorbance
82
values (complimentary to reflectance) in a wavelength range of 1940-1970 nm [14]. In that
83
study, the relation between absorbance and compression strength was not discussed.
84
The ultimate goal of this study is to rapidly classify the type of concrete with various
85
W/C ratios in pavement construction of highways through hyperspectral imaging from an
86
unmanned aerial vehicle, and determine the early-age compressive strength of concrete
87
pavements from light reflectance spectra. The focus of this paper is to develop a dataset with
88
light reflectance and its corresponding compressive strength of mortar of various types, a
89
classification model for mortar type, and a regression curve of reflectance versus compressive
90
strength corresponding to a specific mortar type. Specifically, three groups of mortar cuboid
91
samples with a W/C ratio of 0.4, 0.5 and 0.6 were cast. For each group, five mortar samples
92
were tested for compressive strength after 1, 3, 5 7, 9, 11, 13, or 14 days of curing. Another
93
nine samples were scanned using a hyperspectral camera from Day 1 to 14. A large set of
94
reflectance data were extracted from the scanned images and used to train Nearest Neighbors
95
(KNN) and Support Vector Machine (SVM) classifiers for discrimination of three mortar
96
types. The compressive strength of each type of mortar samples was measured corresponding
97
to the hyperspectral imaging schedule and related to the light reflectance by an exponential
98
regression model developed.
99
100
2. Experiment Setup
101
2.1. Sample preparation
102
Three types of mortar samples were prepared and designated as C1, C2, and C3 in Table
103
1. They are a mixture of water, ordinary Portland cement and Missouri river sand with a
104
W/C/Sand weight ratio of 0.194/0.324/1.0, 0.182/0.364/1.0, and 0.165/0.415/1.0, or a W/C
105
ratio of 0.6, 0.5, and 0.4, respectively. Type I Portland cement was used as detailed in Table
106
2 for its chemical composition. The Missouri river sand used had the maximum particle size
107
of 4.75 mm, a specific gravity of 2.64, and a fineness modulus of 2.71. Freshly mixed mortar
108
was poured into standard cubic steel molds that are 50 mm ร— 50 mm ร— 50 mm in size. After
109
casting, the specimens were covered with wet burlaps and plastic sheets to prevent surface
110
cracking due to shrinkage. After 24 hours of curing, they were demolded for compressive
111
tests and hyperspectral scanning. For each type of mortar mixture, compressive tests of 40
112
samples were conducted according to the ASTM Standard C39 [15], 5 samples tested after
113
1, 3, 5, 7, 9, 11, 13, and 14 days of curing and hardening. All the samples were cured in air
114
with a temperature of 23โ€ฏยฑโ€ฏ1.7โ€ฏยฐC and a relative humidity (RH) of 50โ€ฏยฑโ€ฏ5%. Hyperspectral
115
scanning on 9 samples with each mortar mixture was conducted continuously for 13 days
116
from the end of 1st day to 14th day of curing and hardening. For each cuboid sample, only
117
the four vertical side surfaces were scanned since the horizontal top surface was relatively
118
uneven. In addition, the top surface had a thin layer of cement paste due to water bleeding
119
during mortar settlement, which made its composition different from the side surfaces [16].
120
Table 1. Mix proportions of three types of mortar samples by weight (kg/m3
)
121
Types of mortar samples C1 C2 C3
Water 288 270 245
Ordinary Portland cement 480 540 615
Missouri river sand 1482 1482 1482
Table 2. Mass percentage (%) of oxides in cement
122
SiO2 CaO Al2O3 Fe2O3 MgO SO3 Loss of ignition
19.8 64.2 4.5 3.2 2.7 3.4 2.6
123
2.2. Hyperspectral scanning
124
A wideband hyperspectral camera (Headwall Hyperspec VNIR-SWIR dual sensor) was
125
used to scan the mortar samples. The co-aligned VNIR-SWIR sensor has a broad wavelength
126
range of 400 - 2500 nm. The VNIR sensor has a spectral range of 400-1000 nm with 2.2 nm
127
in spectral resolution and the SWIR sensor has a spectral range of 900-2500 nm with 6 nm
128
in spectral resolution. Figure 1 shows the experimental setup of a cuboid mortar sample. A
129
light source (LED illumination) was set at 0.5 m away from the mortar sample and lit the
130
sample from one side (left in the photo). The hyperspectral camera was set right in front of
131
the mortar sample at 1.2 m standoff distance from the front vertical side of the mortar sample
132
for better resolution of near-distance imaging. The camera was installed on a tripod, both
133
connected to a laptop installed with Hyperspec III software to control the cameraโ€™s rotation
134
(ยฑ5ยฐ) in the horizontal plane and collect images continuously. A grey tarp was set right behind
135
the mortar sample as a reference.
136
137
138
Figure 1. Test setup of a cuboid mortar sample with illumination light, a hyperspectral
139
camera, a laptop computer, and a grey tarp.
140
Prior to each sample scanning, the hyperspectral camera was calibrated through the
141
collection and processing of dark and white reference data. Measuring electric current in the
142
camera system, the dark reference was collected with the camera lens covered, and deducted
143
from any scanned image to cleanse noise. The white reference was used to get a white balance
144
to enhance imaging quality. It was collected by aiming the camera lens at the grey tarp with
145
a reflectance of 32%. The grey tarp was chosen in this study since its color was close to that
146
of the mortar samples. Frame period and exposure time were adjusted to ensure that 60% of
147
the saturated light intensity was detected by using the grey tarp since a lack of light intensity
148
can generate too many bad pixels to correct mathematically. The rotation angle was adjusted
149
so that the camera can scan the mortar surface area of interest at a fixed standoff distance of
150
1.2 m. The rotation speed of the camera was also adjusted until no distorted shapes or forms
151
were seen in the captured image.
152
At the completion of each sample scanning, the scanned data files were transferred from
153
the camera (480 GB solid-state drive) to the laptop computer. SpectralView software was
154
then used to extract the reflectance spectrum for each pixel in the image by:
155
Calibrated Reflectance = ๐‘…๐‘Ž๐‘คโˆ’๐ท๐‘Ž๐‘Ÿ๐‘˜
๐‘Šโ„Ž๐‘–๐‘ก๐‘’โˆ’๐ท๐‘Ž๐‘Ÿ๐‘˜
ร— ๐‘Šโ„Ž๐‘–๐‘ก๐‘’ ๐‘…๐‘’๐‘“๐‘’๐‘Ÿ๐‘’๐‘›๐‘๐‘’ ๐‘…๐‘’๐‘“๐‘™๐‘’๐‘๐‘ก๐‘Ž๐‘›๐‘๐‘’ C๐‘Ž๐‘™๐‘–๐‘๐‘Ÿ๐‘Ž๐‘ก๐‘–๐‘œ๐‘› (1)
where Raw is the raw reflectance spectrum without processing, Dark means the dark
156
reference spectrum, White means the white reference spectrum, and White Reference
157
Reflectance Calibration denotes the maximum reflectance of white reference spectrum to
158
ensure no saturation in measurement. The software SpectralView automatically calculates
159
the normalized reflectance using the dark/white reference spectra.
160
2.3 Data classification techniques
161
Two classification models with KNN and SVM algorithms were established to
162
distinguish various types (W/C ratios) of mortar samples from the reflectance dataset as
163
shown in Figure 2. The reflectance dataset is a group of data with each datum showing a
164
reflectance value and its corresponding class label (W/C ratio). In this study, 80% of the
165
reflectance data were used for training and the remaining 20% were used for testing of the
166
classification models. Both KNN and SVM algorithms were trained in Python to construct
167
the classification models. The established classification models were then used to predict the
168
label (Class A, Class B or Class C) for a W/C ratio of 0.6, 0.5, or 0.4 given a test example of
169
known reflectance value.
170
171
Figure 2. Reflectance data classification using classification algorithms.
172
2.3.1 KNN
173
The KNN algorithm computes the proximity of a test example z to K data points in the
174
training set, which are closest to z. The test example is classified based on the majority class
175
label of its K nearest neighbors [17]. Weights are assigned to the contributions of the
176
neighbors so that the impact of data depends on their distances to the test example. Choosing
177
the right parameter K is important to ensure a better accuracy in classification. A small K can
178
result in overfitting due to noise in the training data, whereas a large K can lead to
179
misclassification because the nearest neighbors may include data that are located far away
180
from its neighborhood [17].
181
2.3.2 SVM
182
The SVM algorithm creates a line in two-dimensional planes, a plane in three-
183
dimensional spaces, or more generally a hyperplane to divide the data into several classes
184
[18]. Support vectors are the data points nearest to the hyperplane. The distance between the
185
hyperplane and the nearest data is called margin [18]. The goal of SVM is to choose a
186
hyperplane with the maximum margin. To briefly describe the SVM technique, a linear
187
classifier is introduced first and then extended to the nonlinear classifier. Next, the maximum
188
margin of a hyperplane is described.
189
Linear classifier is used to find a line or a plane (hyperplane) to separate dataset
190
{๐ฑ๐‘–, ๐‘ฆ๐‘–}๐‘–=1
๐‘›
into two classes. Here, ๐ฑ๐‘– is the ๐‘–๐‘กโ„Ž
vector in the given dataset, ๐‘ฆ๐‘– is the label
191
associated with ๐ฑ๐‘–. The hyperplane is defined as [18]:
192
๐‘“(๐ฑ) = ๐ฐ๐ฑ + ๐‘ = 0, ๐ฐ๐ฑ = โˆ‘ ๐‘ค๐‘–๐‘ฅ๐‘–
๐‘–
(2)
where w is a weight vector, and b is a bias. As illustrated in Figure 3 for the case of two-
193
dimensional plane x1x2, f(x) =0 is a line that divides the entire dataset into two classes: f(x)
194
>0 and f(x) <0.
195
196
Figure 3. A linear classifier with maximum margins that divides the data into two sets.
197
198
When the data cannot be separated by a linear classifier, they can be mapped to a higher
199
dimension and converted to linearly separable data through a projection function ๐œ‘ [18, 19].
200
The classifier then becomes:
201
๐‘“(๐ฑ) = ๐ฐ๐œ‘(๐ฑ) + ๐‘ (3)
As the high-dimensional projection function is complicated to compute, this classifier is
202
projected back to the original dimension through a transformation known as the kernel
203
function. In this case, the weight vector can be expressed into a linear combination of the
204
training data [18]:
205
๐ฐ = โˆ‘ ๐›ผ๐‘–
๐‘›
๐‘–=1
๐œ‘(๐ฑ๐‘–) (4)
where ๐›ผ๐‘– is the coefficient related to a decision boundary. The kernel function is defined as
206
[18]:
207
๐‘˜(๐ฑ๐‘–, ๐ฑ) = ๐œ‘(๐ฑ๐‘–)๐œ‘(๐ฑ) (5)
The classifier then transforms to:
208
๐‘“(๐ฑ) = โˆ‘ ๐›ผ๐‘–
๐‘›
๐‘–=1 ๐œ‘(๐ฑ๐‘–)๐œ‘(๐ฑ) + ๐‘ = โˆ‘ ๐›ผ๐‘–
๐‘›
๐‘–=1 ๐‘˜(๐ฑ๐‘–, ๐ฑ) + ๐‘ (6)
Two kernels are widely used in the literature for various applications: polynomial kernel and
209
Gaussian kernel. A polynomial kernel is defined as [18]:
210
๐‘˜(๐ฑ๐‘–, ๐ฑ) = (๐ฑ๐ฑ๐‘– + 1)๐‘‘ (7)
A Gaussian kernel is defined as [18, 19]:
211
๐‘˜(๐ฑ๐‘–, ๐ฑ) = exp (โˆ’
โ€–๐ฑโˆ’๐ฑ๐‘–โ€–๐Ÿ
2๐œŽ2
)
(8)
where d is the degree of polynomial kernel and ฯƒ is a parameter that controls the width of
212
Gaussian kernel. Both parameters control the flexibility of the classifier. When ฯƒ is increased,
213
a greater curvature is introduced to the decision boundary but overfitting will occur if ฯƒ is
214
too large.
215
SVM looks for a higher margin to get a better classification result for the testing data.
216
The margin of a hyperplane f (x) is defined as:
217
๐‘š(๐‘“) =
1
โ€–๐ฐโ€–
(9)
As indicated in Equation (9), to maximize the margin of the classifier is equivalent to
218
minimizeโ€–๐ฐโ€–2
. The maximum margins are the margins that push up against the support
219
vectors. To ensure that the linearly-separable data are classified correctly, the maximum
220
margin and its constraint are defined as [18, 19]:
221
Minimize
1
2
โ€–๐ฐโ€–2 (10)
Subject to: ๐‘ฆ๐‘–(๐ฐ๐ฑ + ๐‘) โ‰ฅ 1 ๐‘– = 1, โ€ฆ , ๐‘›. (11)
When the data are not completely separable, the constraint is relaxed and a greater margin
222
can be achieved by [18, 19]:
223
Minimize
1
2
โ€–๐ฐโ€–2
+ ๐ถ โˆ‘ ๐œ‰๐‘–
๐‘›
๐‘–=1
(12)
Subject to: ๐‘ฆ๐‘–(๐ฐ๐‘ฅ + ๐‘) โ‰ฅ 1 โˆ’ ๐œ‰๐‘– (13)
where ๐œ‰๐‘– (0โ‰ค ๐œ‰๐‘– โ‰ค 1) is the margin error that allows an example to be in the margin and C
224
is the penalty that lowers the misclassification rate. When C is increased, a smaller margin
225
error is achieved. C needs to be adjusted to ensure the maximum margin with a minimum
226
margin error [18, 19].
227
3. Results and Discussion
228
3.1. Hyperspectral information
229
Figure 4 shows the raw hyperspectral image of one mortar specimen, the image after
230
subtraction of dark reference, and the image after dark and white reference deductions. The
231
sensor current measured from the dark reference can induce perturbation and generate a noisy
232
and drifted spectrum. The white reference can rectify illumination non-uniformity and non-
233
flatness of a spectrum. Removal of the dark reference and the white reference can correct
234
the image and produce right reflectance spectra. For each type of mortar, 9 cuboid samples
235
were prepared, 4 side faces of each sample were scanned, and 30 spectra were extracted over
236
a 50 mm ร— 50 mm side surface area, totaling 1080 spectra for each scanning day. The spectra
237
were collected from the flat surface area only to avoid any non-uniform illumination from
238
uneven spots.
239
(a) (b) (c)
Figure 4. Hyperspectral image of a mortar specimen: (a) raw, (b) after subtraction of dark
240
reference, and (c) after dark and white reference deduction.
241
Figure 5 shows the average reflectance spectra of samples with three different W/C ratios
242
over a period of 14 days. Each line represents the average reflectance spectrum of 1080
243
spectra in the wavelength range of 1200 nm to 2400 nm. The average spectra can reduce
244
potential biases and are more representative of the scanning surface. As seen in Figure 5, the
245
reflectance value on the average spectra rapidly ascends from Day 1 to Day 3 and then
246
gradually increased till Day 14 of test. After 1 day of testing, the samples were scanned after
247
they were demolded and put in air at the room temperature for 1 hour. The higher moisture
248
content on the sample surface resulted in the lower reflectance value due to water absorption.
249
The rapid increase of reflectance from Day 1 to Day 3 is because the hydration process during
250
this period rapidly consumes more water compared with that at a later stage.
251
252
0.04
0.05
0.06
0.07
0.08
0.09
0.1
0.11
0.12
1200 1400 1600 1800 2000 2200 2400
Reflectance
(R)
Wavelength (nm)
Day 1 Day 3 Day 5 Day 7
Day 9 Day 11 Day 13 Day 14
(a)
253
254
Figure 5. The average reflectance spectra over a wavelength of 1200 nm to 2400 nm for
255
samples with a W/C ratio of: (a) 0.6, (b) 0.5, and (c) 0.4.
256
When shot on the surface of materials, some of the incident light leads to vibration of
257
molecules and is absorbed by the chemical bond between atoms in the molecules. In the Near
258
Infrared Region (NIR) (from 780 nm to 2500 nm), higher vibrational energy is acquired to
259
absorb the light, which stimulates the overtones and combinations of fundamental vibrations
260
[20, 21]. Basically, overtones and combinations of the vibrations of C-H, O-H, N-H, and S-
261
H chemical bonds dominate NIR spectroscopy with each chemical bond corresponding to a
262
wavelength region for light absorbance [21]. The combination of OH and H2O corresponds
263
to the region of 1900 nm to 2000 nm [21-24], which is of particular interest in this study. The
264
reflectance change in this wavelength range can be used to track the change of H2O molecules
265
due to hydration consumption in the process of mortar curing. The reflectance values over
266
0.04
0.05
0.06
0.07
0.08
0.09
0.1
0.11
0.12
1200 1400 1600 1800 2000 2200 2400
Reflectance
(R)
Wavelength (nm)
Day 1 Day 3 Day 5 Day 7
Day 9 Day 11 Day 13 Day 14
0.06
0.07
0.08
0.09
0.1
0.11
0.12
1200 1400 1600 1800 2000 2200 2400
Reflectance
(R)
Wavelength (nm)
Day 1 Day 3 Day 5 Day 7
Day 9 Day 11 Day 13 Day 14
(b)
(c)
the wavelength range of 1920 nm to 1980 nm as shown in the marked red circles in Figure 5
267
are averaged and defined as dR (1920-1980). The average reflectance dR (1920-1980) with
268
an error bar of ยฑ one standard deviation for three types of mortar over the curing period of
269
14 days are plotted in Figure 6. Logarithmic regression was conducted to fit into the
270
experimental data for the three types of mortar. R2
value for each type of mortar is higher
271
than 90%, indicating good fitting of the regression curve. The overall reflectance during the
272
14 days shows an increasing trend when the W/C ratio is reduced from 0.6 to 0.4 since the
273
mortar with lower W/C ratio has less water content. For each type of mortar, the average
274
reflectance increases because water is reacted and reduced during the hydration process. As
275
less water is left on the sample surface, less light is absorbed by water molecules and
276
reflectance of the light is increased. Therefore, the regression curve can be used to predict
277
the curing process for mortar samples.
278
Figure 6. The average reflectance dR (1920-1980) over a curing period of 14 days for
279
samples with W/C ratios of: (a) 0.6, (b) 0.5, and (c) 0.4.
280
3.2 Classification results with KNN and SVM
281
Both KNN and SVM algorithms were used to first establish classification models with
282
the training data, and then predict classification labels for the test data. In the KNN algorithm,
283
the parameter K varied from 1 to 40 in model training. Its corresponding prediction
284
y = 0.007ln(x) + 0.07
Rยฒ = 0.92
W/C=0.6
0
0.02
0.04
0.06
0.08
0.1
0.12
0 2 4 6 8 10 12 14
Average
reflectance
dA
(1920-1980)
Curing Time (Day)
y = 0.011ln(x) + 0.07
Rยฒ = 0.93
W/C=0.5
0
0.02
0.04
0.06
0.08
0.1
0.12
0 2 4 6 8 10 12 14
Average
reflectance
dA
(1920-1980)
Curing Time (Day)
y = 0.007ln(x) + 0.09
Rยฒ = 0.91
0
0.02
0.04
0.06
0.08
0.1
0.12
0 2 4 6 8 10 12 14
Average
reflectance
dA
(1920-1980)
Curing Time (Day)
(b)
(a)
(c)
accuracies after 1, 4, 7, and 13 days of testing are presented in Figure 7. In this study, accuracy
285
is defined by the relative difference in percentage between the number of correct predictions
286
and the number of actual test data. It is observed that K = 5 yields the highest prediction
287
accuracy for Day 1, Day 7, and Day 13, and K = 6 provides the highest accuracy for Day 4.
288
Overall, K = 5 is chosen for the KNN classification model. Figure 8 shows the predicted
289
classifications after 1, 7, and 13 day of testing with K = 5 and after 4 day of testing with K =
290
6. In Figure 8, โ€˜Trueโ€™ represents the test data and โ€˜Predโ€™ symbolizes the prediction data with
291
the trained KNN classification model. The predicted classification (yellow triangles) and the
292
actual classification (blue squares) are in general agreement.
293
294
Figure 7. Prediction accuracies as a function of K after 1, 4, 7, and 13 day of testing.
295
296
45
50
55
60
65
70
75
80
0 5 10 15 20 25 30 35 40
Accuracy
(%)
K value
Day 1
Day 4
Day 7
Day 13
(b)
(a)
Figure 8. Classification predictions after (a) 1, (b) 4, (c) 7, and (d) 13 day of testing.
297
To test the applicability of parameter K, K=5 is applied to predict the classifications of
298
the test data after 2, 5, 8, and 14 day of testing, and the predicted classifications are compared
299
in Figure 9 with their actual classifications. It can be seen from Figure 9 that the predictions
300
are in good agreement with the actual classifications. Specifically, the prediction accuracies
301
during the 4 days of testing range from 70% to 75% as shown in Figure 10, which falls into
302
the same range of accuracies achieved after 1, 4, 7, and 13 day of testing. This comparison
303
indicates that K = 5 is the best fit for the KNN model. Due to low prediction accuracies (<
304
75%) for the KNN model, the SVM algorithm is attempted to improve the prediction
305
accuracy for mortar classification.
306
(d)
(c)
(a) (b)
Figure 9. Classification predictions with K = 5 after (a) 2, (b) 5, (c) 8, and (d) 14 day of
307
testing.
308
309
Figure 10. Prediction accuracies with K = 5 after 2, 5, 8, and 14 day of testing.
310
In this study, the average reflectance dR(1920-1980 nm) and its corresponding Min-Max
311
normalized reflectance calculated from Equation (14) forms two features in x1-x2 plane. The
312
representative training data in the feature plane after 13 day of testing are presented in Figure
313
11. It can be seen that the three classes of training data are mainly distributed along three
314
straight lines due to the correlation of two features selected. The relatively low reflectance
315
for a w/c ratio of 0.6 can be easily separated from the other classes using a linear separator.
316
However, the two straight lines representing the feature data in other two classes are crossed
317
each other, making it impossible to separate them directly with a linear classifier. Therefore,
318
a nonlinear classifier with the Gaussian kernel and a relaxed constraint were used to obtain a
319
high prediction accuracy of the classification model.
320
0
20
40
60
80
100
Accuracy
(%)
Day 2 Day 5 Day 8 Day 14
(c) (d)
Min-Max normalized reflectance=
dR (1920โˆ’1980) โˆ’Min.
Max.โˆ’Min.
ร— 100% (14)
321
where Min. and Max. represent the minimum reflectance and the maximum reflectance of
322
the whole data set during each test day, respectively. dR (1920 โˆ’ 1980) represents the
323
average reflectance over a wavelength range of 1920 nm to 1980 nm, corresponding to each
324
datum used in the dataset.
325
326
Figure 11. Training data in the feature plane after 13 day of testing.
327
To improve the accuracy of classification predication, the hyperparameters C and ฯƒ in
328
the SVM model were first tuned. Figure 12 shows the prediction accuracies after 1, 4, 7, and
329
13 day of testing when C is examined from 10-2
to 103
and ฯƒ is tested from 10-3
to 10. It can
330
be observed from Figure 12 that C=103
and ฯƒ=10 yield the highest prediction accuracies of
331
94.5%, 89.9%, 95%, and 87.4% after 1, 4, 7, and 13 day of testing, respectively. The
332
predicted classifications from the SVM model are compared in Figure 13 with the actual
333
classifications displayed when C = 103
and ฯƒ = 10. Compared with the classification results
334
in Figures 8 and 9 predicted from the KNN model, the overlapped range between the
335
predicted data and test data in Figure 12 is enlarged with a higher accuracy. Therefore, C =
336
103
and ฯƒ = 10 are chosen for the SVM classification model.
337
Figure 12. Prediction accuracies as a function of C and ฯƒ in the logarithmic format after (a)
338
1, (b) 4, (c) 7, and (d) 13 day of testing.
339
340
(a) (b)
(c) (d)
(a) (b)
Figure 13. Classification predictions of the test data after (a) 1 , (b) 4 , (c) 7 , and (d) 13 day
341
of testing with C = 103
and ฯƒ = 10.
342
343
To evaluate their applicability, the two hyperparameters C = 103
and ฯƒ = 10 are applied
344
to the test data collected after 2, 5, 8, and 14 day of testing. The predicted results are compared
345
in Figure 14 with their actual classifications. Figure 15 shows 93.4%, 92.9%, 93.0%, and
346
92.5% prediction accuracies after 2, 5, 8, and 14 day of testing, respectively. Therefore, C =
347
103
and ฯƒ = 10 in the SVM model are appropriate for the classification of mortar.
348
(c)
(d)
(a) (b)
Figure 14. Classification predictions with C = 103
and ฯƒ = 10 after (a) 2, (b) 5, (c) 8, and (d)
349
14 day of testing.
350
351
Figure 15. Prediction accuracies with C = 103
and ฯƒ = 10 after 2, 5, 8, and 14 day of testing.
352
3.3 Relation between average reflectance and compressive strength of mortar samples
353
For each W/C ratio, five mortar samples were tested in compression after 1, 3, 5, 7, 9,
354
11, 13, and 14 day of curing, following the ASTM C39 Guideline. The average and standard
355
deviation of the five compressive strengths of mortar for each testing day were determined.
356
The average strength with an error bar of ยฑ one standard deviation is displayed in Figure 16
357
as a function of the curing day. Overall, the compressive strength of mortar decreases with
358
the increase of W/C ratio. For each W/C ratio, the compressive strength increases
359
logarithmically with the curing time. The short error bars in Figure 16 indicate consistent
360
data obtained from the compressive tests.
361
0
20
40
60
80
100
Accuracy
(%)
Day 2 Day 5 Day 8 Day 14
(c) (d)
362
Figure 16. Compressive strengths of mortar samples with W/C ratios of 0.6, 0.5, and 0.4.
363
After 1, 3, 5, 7, 9, 11, 13, and 14 day of curing, both the compressive strength and the
364
average reflectance dA (1920-1980) were obtained. Their correlations for W/C ratios of 0.6,
365
0.5, and 0.4 are drawn in Figure 17. Exponential regression of the test data was conducted
366
for each W/C ratio with R2
larger than 90%. These relations indicate that the average
367
reflectance can be used to predict the compressive strength of mortar. Such relationships can
368
potentially be applied to predict the strength of early-age concrete pavement in practice.
369
y = 10.1ln(x) + 4.7
Rยฒ = 0.98
y = 11.2ln(x) + 6.6
Rยฒ = 0.99
y = 13.4ln(x) + 10.1
Rยฒ = 0.99
0
10
20
30
40
50
0 2 4 6 8 10 12 14
Compression
strength
(MPa)
Curing time (Day)
0.6
0.5
0.4
y = 0.041e79x
Rยฒ = 0.99
W/C=0.6
0
10
20
30
40
0.04 0.08 0.12
Compression
Stregnth
(MPa)
Average Reflenctance
dA (1920-1980)
y = 0.214e51x
Rยฒ = 0.98
W/C=0.5
0
10
20
30
40
0.04 0.08 0.12
Compression
Stregnth
(MPa)
Average Reflenctance
dA (1920-1980)
(a) (b)
Figure 17. Average reflectance dA (1920-1980) versus the compressive strength of mortar
370
with a W/C ratio of: (a) 0.6, (b) 0.5, and (c) 0.4 over 14 days.
371
3.4 Further work
372
Concrete samples with a wide range of mixture designs will be scanned with
373
hyperspectral camera. The SVM classification model will be kept updated and improved with
374
the reflectance data extracted from newly scanned concrete samples. The reflectance dataset
375
needs to be enlarged to include as many types of concrete as possible. Field test will be
376
conducted with hyperspectral camera installed on UAV to establish a similar correlation
377
between reflectance and compressive strength under actual in-situ conditions.
378
4. Conclusions
379
This study utilized a hyperspectral camera to scan the surface of mortar samples and
380
extract the reflectance data for the classification of different types of mortar and the
381
prediction of their compressive strength. Based on experimental data, regression analysis,
382
and classification results, the following conclusions can be drawn:
383
โ€ข The reflectance spectra over a wavelength range of 1200 nm to 2400 nm are
384
consistently shifted upward as mortar samples are cured and hardened over time. In
385
particular, the average reflectance in 1920-1980 nm wavelength increases
386
logarithmically with time because water molecules on each sample surface are
387
gradually reacted during the hydration process. As the water content of mortar is
388
reduced, less light is absorbed and more light is reflected. The average reflectance
389
can be accurately predicted from the curing time with their correlation coefficient
390
of over 0.9.
391
โ€ข The KNN classification model with K=5 represents the best fit to the test data taken
392
from mortar samples with a W/C ratio of 0.4 to 0.6. The classification (W/C ratio)
393
y = 0.045e66x
Rยฒ = 0.99
W/C=0.4
0
10
20
30
40
50
0.04 0.08 0.12
Compression
Stregnth
(MPa)
Average Reflenctance
dA (1920-1980)
(c)
prediction accuracies are in the range of 70% to 75%. The SVM classification model
394
with C=103
and ฯƒ=10 are the best fit to the test data with a prediction accuracy of
395
approximately 90%. Thus, the SVM classification model is recommended to
396
classify various types of mortar.
397
โ€ข The compressive strength of mortar can be exponentially related to the average
398
reflectance in 1920-1980 nm wavelength with a correlation coefficient of over 0.95.
399
With this strong exponential relation, hyperspectral imaging can be used as a rapid
400
and nondestructive evaluation tool to predict the compressive strength of mortar.
401
With further studies on concrete specimens, hyperspectral imaging is promising for
402
the prediction of early-age strength of concrete pavement in practical applications.
403
Author Contributions: Conceptualization, L.F., G. C. and H. M.; methodology, L.F. and
404
M.F.; validation, L.F. M.F. and G. C.; formal analysis, L.F. and M.F.; investigation, L.F.,
405
M.F. and A.A.; resources, G. C.; data curation, L.F. and M.F.; writingโ€”original draft
406
preparation, L.F. and M.F.; writingโ€”review and editing, L.F., M.F. and G.C.; visualization,
407
L.F., M.F. and G.C.; supervision, G.C.; project administration, G.C.; funding acquisition,
408
G.C. In general, the first two authors contributed equally to the preparation of this paper.
409
Funding: Financial support to complete this study was provided by the U.S. Department of
410
Transportation, Office of the Assistant Secretary for Research and Technology (OST-R)
411
under the Auspices of the INSPIRE University Transportation Center under Grant No.
412
69A3551747126 at Missouri University of Science and Technology. The findings and
413
opinions expressed in this paper are solely those of the authors and do not represent the
414
official policy or position of the USDOT/OST-R, or any State or other entity.
415
Conflicts of Interest: The authors declare no conflict of interest.
416
References
417
1. Bullard, J.W.; Jennings, H.M.; Livingston, R.A.; Nonat, A.; Scherer, G.W.; Schweitzer,
418
J.S.; Scrivener, K.L.; Thomas, J.J. Mechanisms of cement hydration. Cement and
419
concrete research 2011, 41, pp.1208-1223.
420
2. ACI. Specifications for structural concrete for buildings. American Concrete Institute
421
(ACI) Committee 301-72, 1972.
422
3. ASTM. Standard test method for obtaining and testing drilled cores and sawed beams of
423
concrete. American Society for Testing and Materials (ASTMs) C42/C42M-18a, 2018.
424
4. ASTM. Standard test method for tensile strength of concrete surfaces and the bond
425
strength or tensile strength of concrete repair and overlay materials by direct tension
426
(pull-off method). American Society for Testing and Materials (ASTMs)
427
C1583/C1583M-13, 2013.
428
5. ASTM. Standard Practice for Estimating Concrete Strength by the Maturity Method.
429
American Society for Testing and Materials (ASTMs) C1074-19, 2019.
430
6. Kewalramani, M. A.; Gupta, R. Concrete compressive strength prediction using
431
ultrasonic pulse velocity through artificial neural networks. Automation in Construction
432
2006, 15, pp.374-379.
433
7. Eismann, M. T. Hyperspectral remote sensing, SPIE Press, Bellingham, Washington
434
USA, 2012; pp.1-20.
435
8. Zaini, N.; Meer, F. V. D.; Ruitenbeek, F.V.; Smeth, B. D.; Amri, F.; Lievens. C. An
436
alternative quality control technique for mineral chemistry analysis of Portland cement-
437
grade limestone using shortwave infrared spectroscopy. Remote sensing 2016, 8, pp.950-
438
966.
439
9. Arita, J.; Sasaki, K.; Endo, T.; Yasuoka, Y. Assessment of concrete degradation with
440
hyper-spectral remote sensing. The 22nd Asian Conference on Remote Sensing,
441
Singapore, 2001, pp. 5-9.
442
10. Kohri, M.; Ueda, T.; Mizuguchi, H. Application of a near-infrared spectroscopic
443
technique to estimate the chloride ion content in mortar deteriorated by chloride attack
444
and carbonation. Journal of Advanced Concrete Technology 2010, 8, pp.15-25.
445
11. Brook, A.; Ben-Dor, E. Reflectance spectroscopy as a tool to assess the quality of
446
concrete in situ. Journal of Civil Engineering and Construction Technology 2011, 8,
447
pp.169-188.
448
12. Brook, A.; Ben-Dor, E. Reflectance spectroscopy as a tool to assess the strength of high-
449
performance concrete in situ. Journal of Civil Engineering and Construction Technology
450
2012, 7, pp.195-203.
451
13. Lee, J.D.; Dewitt, B.A.; Lee, S.S.; Bhang, K.J.; Sim, J.B. Analysis of concrete
452
reflectance characteristics using spectrometer and VNIR hyperspectral camera.
453
International Archives of the Photogrammetry, Remote Sensing and Spatial Information
454
Sciences 2012, 39, B7.
455
14. Zahiri, Z., Laefer, D.F.; Gowen, A. The feasibility of short-wave infrared spectrometry
456
in assessing water-to-cement ratio and density of hardened concrete. Construction and
457
Building Materials 2018, 185, 661-669.
458
15. ASTM. Standard test method for compressive strength of cylindrical concrete
459
specimens. American Society for Testing and Materials (ASTMs) C39/C39M-18, 2018.
460
16. Josserand, L.; Larrard, F.D. A method for concrete bleeding measurement. Materials
461
and Structures 2004, 37, 666.
462
17. Maglogiannis, I.G. Emerging artificial intelligence applications in computer
463
engineering: real word AI systems with applications in EHealth, HCI, information
464
retrieval and pervasive technologies Vol. 160. Ios Press, Amsterdam Netherlands, 2007;
465
pp.11-12.
466
18. Ben-Hur, A.; Weston, J. A userโ€™s guide to support vector machines. In Data mining
467
techniques for the life sciences; Carugo, O., Frank Eisenhaber, F.; Humana Press, New
468
York, NY USA, 2010; pp.223-239.
469
19. Amami, R.; Ayed, D. B.; Ellouze, N. Practical selection of SVM supervised parameters
470
with different feature representations for vowel recognition. arXiv preprint 2015,
471
arXiv:1507.06020.
472
20. Atkins, P.W; Paula, J. Atkinsโ€™ Physical Chemistry, 7th ed. Oxford University Press,
473
Cambridge, United Kingdom, 2002; pp. 320-352.
474
21. Okparanma, R. N.; Araka, P. P.; Ayotamuno, J. M.; Mouazen, A. Towards enhancing
475
sustainable reuse of pre-treated drill cuttings for construction purposes by near-infrared
476
analysis: A review. Journal of Civil Engineering and Construction Technology 2018, 3,
477
pp.19-39.
478
22. Walling, P. L.; Dabney, J. M. Moisture in skin by near-infrared reflectance spectroscopy.
479
Journal of the Society of Cosmetic Chemists 1989, 40, 151-171.
480
23. Kohri, M.; Ueda, T.; Mizuguchi, H. Application of a near-infrared spectroscopic
481
technique to estimate the chloride ion content in mortar deteriorated by chloride attack
482
and carbonation. Journal of Advanced Concrete Technology 2010, 1, 15-25.
483
24. Fan, L.; Alhaj, A.; Ma, H.; Chen, G. Assessing moisture content on the surface of mortar
484
samples from hyperspectral imaging. The 9th International Conference on Structural
485
Health Monitoring of Intelligent Infrastructure, Saint Louis USA, 2019; pp. 1150-1155.
486
487

More Related Content

Similar to HyperspectralImaging (1).pdf

Mathematical Relationships between the Compressive Strength and Some Other St...
Mathematical Relationships between the Compressive Strength and Some Other St...Mathematical Relationships between the Compressive Strength and Some Other St...
Mathematical Relationships between the Compressive Strength and Some Other St...IOSR Journals
ย 
kwanjai_yuk,+6.เธญ.เธญเธฃเธงเธฃเธฃเธ“-Relationship+between+Porosity+&+Compressive+Strength+...
kwanjai_yuk,+6.เธญ.เธญเธฃเธงเธฃเธฃเธ“-Relationship+between+Porosity+&+Compressive+Strength+...kwanjai_yuk,+6.เธญ.เธญเธฃเธงเธฃเธฃเธ“-Relationship+between+Porosity+&+Compressive+Strength+...
kwanjai_yuk,+6.เธญ.เธญเธฃเธงเธฃเธฃเธ“-Relationship+between+Porosity+&+Compressive+Strength+...s_p2000
ย 
Destructive and Non- Destructive Testing for Concrete in Sudan - A Comparativ...
Destructive and Non- Destructive Testing for Concrete in Sudan - A Comparativ...Destructive and Non- Destructive Testing for Concrete in Sudan - A Comparativ...
Destructive and Non- Destructive Testing for Concrete in Sudan - A Comparativ...iosrjce
ย 
Metamodel techniques to estimate the compressive strength of UHPFRC using var...
Metamodel techniques to estimate the compressive strength of UHPFRC using var...Metamodel techniques to estimate the compressive strength of UHPFRC using var...
Metamodel techniques to estimate the compressive strength of UHPFRC using var...Shakerqaidi
ย 
Influence of the proportion of materials on the rheology and mechanical stren...
Influence of the proportion of materials on the rheology and mechanical stren...Influence of the proportion of materials on the rheology and mechanical stren...
Influence of the proportion of materials on the rheology and mechanical stren...Shakerqaidi
ย 
Nondestructive material testing with ultrasonics
Nondestructive material testing with ultrasonicsNondestructive material testing with ultrasonics
Nondestructive material testing with ultrasonicsFatma Abdalla
ย 
Investigation on fine aggregate by broken tiles in concrete
Investigation on fine aggregate by broken tiles in concreteInvestigation on fine aggregate by broken tiles in concrete
Investigation on fine aggregate by broken tiles in concreteIJARIIT
ย 
A Proposed Equation for Elastic Modulus of High-Strength Concrete Using Local...
A Proposed Equation for Elastic Modulus of High-Strength Concrete Using Local...A Proposed Equation for Elastic Modulus of High-Strength Concrete Using Local...
A Proposed Equation for Elastic Modulus of High-Strength Concrete Using Local...IJERA Editor
ย 
Ultimate Behavior of Lightweight High Strength Concrete Filled Steel Tube (LW...
Ultimate Behavior of Lightweight High Strength Concrete Filled Steel Tube (LW...Ultimate Behavior of Lightweight High Strength Concrete Filled Steel Tube (LW...
Ultimate Behavior of Lightweight High Strength Concrete Filled Steel Tube (LW...IOSR Journals
ย 
Ultra-high-performance concrete Impacts of steel fibre shape and content on f...
Ultra-high-performance concrete Impacts of steel fibre shape and content on f...Ultra-high-performance concrete Impacts of steel fibre shape and content on f...
Ultra-high-performance concrete Impacts of steel fibre shape and content on f...Shakerqaidi
ย 
CHANDU PROJECT FINAL PPT.pptx
CHANDU PROJECT FINAL PPT.pptxCHANDU PROJECT FINAL PPT.pptx
CHANDU PROJECT FINAL PPT.pptxkasarla sagar
ย 
Fatigue behavior of high volume fly ash
Fatigue behavior of high volume fly ashFatigue behavior of high volume fly ash
Fatigue behavior of high volume fly ashIAEME Publication
ย 
A STUDY ON STRESS-STRAIN BEHAVIOUR OF TIE CONFINED CONCRETE CONTAINING CERAMI...
A STUDY ON STRESS-STRAIN BEHAVIOUR OF TIE CONFINED CONCRETE CONTAINING CERAMI...A STUDY ON STRESS-STRAIN BEHAVIOUR OF TIE CONFINED CONCRETE CONTAINING CERAMI...
A STUDY ON STRESS-STRAIN BEHAVIOUR OF TIE CONFINED CONCRETE CONTAINING CERAMI...IRJET Journal
ย 
Prediction of concrete materials compressive strength using surrogate models.pdf
Prediction of concrete materials compressive strength using surrogate models.pdfPrediction of concrete materials compressive strength using surrogate models.pdf
Prediction of concrete materials compressive strength using surrogate models.pdfShakerqaidi
ย 
Effect of carbon nanotubes on the mechanical fracture parameters and microstr...
Effect of carbon nanotubes on the mechanical fracture parameters and microstr...Effect of carbon nanotubes on the mechanical fracture parameters and microstr...
Effect of carbon nanotubes on the mechanical fracture parameters and microstr...eSAT Publishing House
ย 
A fracture mechanics based method for prediction of
A fracture mechanics based method for prediction ofA fracture mechanics based method for prediction of
A fracture mechanics based method for prediction ofSAJITH GEORGE
ย 
โ€œEXPERIMENTAL STUDY ON PARTIAL REPLACEMENT OF CEMENT BY SEWAGE SLUDGE ASH AND...
โ€œEXPERIMENTAL STUDY ON PARTIAL REPLACEMENT OF CEMENT BY SEWAGE SLUDGE ASH AND...โ€œEXPERIMENTAL STUDY ON PARTIAL REPLACEMENT OF CEMENT BY SEWAGE SLUDGE ASH AND...
โ€œEXPERIMENTAL STUDY ON PARTIAL REPLACEMENT OF CEMENT BY SEWAGE SLUDGE ASH AND...IRJET Journal
ย 
An Experimental Study on Effects of Quarry Dust as Partial Replacement of San...
An Experimental Study on Effects of Quarry Dust as Partial Replacement of San...An Experimental Study on Effects of Quarry Dust as Partial Replacement of San...
An Experimental Study on Effects of Quarry Dust as Partial Replacement of San...IRJET Journal
ย 

Similar to HyperspectralImaging (1).pdf (20)

M012139397
M012139397M012139397
M012139397
ย 
Mathematical Relationships between the Compressive Strength and Some Other St...
Mathematical Relationships between the Compressive Strength and Some Other St...Mathematical Relationships between the Compressive Strength and Some Other St...
Mathematical Relationships between the Compressive Strength and Some Other St...
ย 
kwanjai_yuk,+6.เธญ.เธญเธฃเธงเธฃเธฃเธ“-Relationship+between+Porosity+&+Compressive+Strength+...
kwanjai_yuk,+6.เธญ.เธญเธฃเธงเธฃเธฃเธ“-Relationship+between+Porosity+&+Compressive+Strength+...kwanjai_yuk,+6.เธญ.เธญเธฃเธงเธฃเธฃเธ“-Relationship+between+Porosity+&+Compressive+Strength+...
kwanjai_yuk,+6.เธญ.เธญเธฃเธงเธฃเธฃเธ“-Relationship+between+Porosity+&+Compressive+Strength+...
ย 
Destructive and Non- Destructive Testing for Concrete in Sudan - A Comparativ...
Destructive and Non- Destructive Testing for Concrete in Sudan - A Comparativ...Destructive and Non- Destructive Testing for Concrete in Sudan - A Comparativ...
Destructive and Non- Destructive Testing for Concrete in Sudan - A Comparativ...
ย 
H012664448
H012664448H012664448
H012664448
ย 
Metamodel techniques to estimate the compressive strength of UHPFRC using var...
Metamodel techniques to estimate the compressive strength of UHPFRC using var...Metamodel techniques to estimate the compressive strength of UHPFRC using var...
Metamodel techniques to estimate the compressive strength of UHPFRC using var...
ย 
Influence of the proportion of materials on the rheology and mechanical stren...
Influence of the proportion of materials on the rheology and mechanical stren...Influence of the proportion of materials on the rheology and mechanical stren...
Influence of the proportion of materials on the rheology and mechanical stren...
ย 
Nondestructive material testing with ultrasonics
Nondestructive material testing with ultrasonicsNondestructive material testing with ultrasonics
Nondestructive material testing with ultrasonics
ย 
Investigation on fine aggregate by broken tiles in concrete
Investigation on fine aggregate by broken tiles in concreteInvestigation on fine aggregate by broken tiles in concrete
Investigation on fine aggregate by broken tiles in concrete
ย 
A Proposed Equation for Elastic Modulus of High-Strength Concrete Using Local...
A Proposed Equation for Elastic Modulus of High-Strength Concrete Using Local...A Proposed Equation for Elastic Modulus of High-Strength Concrete Using Local...
A Proposed Equation for Elastic Modulus of High-Strength Concrete Using Local...
ย 
Ultimate Behavior of Lightweight High Strength Concrete Filled Steel Tube (LW...
Ultimate Behavior of Lightweight High Strength Concrete Filled Steel Tube (LW...Ultimate Behavior of Lightweight High Strength Concrete Filled Steel Tube (LW...
Ultimate Behavior of Lightweight High Strength Concrete Filled Steel Tube (LW...
ย 
Ultra-high-performance concrete Impacts of steel fibre shape and content on f...
Ultra-high-performance concrete Impacts of steel fibre shape and content on f...Ultra-high-performance concrete Impacts of steel fibre shape and content on f...
Ultra-high-performance concrete Impacts of steel fibre shape and content on f...
ย 
CHANDU PROJECT FINAL PPT.pptx
CHANDU PROJECT FINAL PPT.pptxCHANDU PROJECT FINAL PPT.pptx
CHANDU PROJECT FINAL PPT.pptx
ย 
Fatigue behavior of high volume fly ash
Fatigue behavior of high volume fly ashFatigue behavior of high volume fly ash
Fatigue behavior of high volume fly ash
ย 
A STUDY ON STRESS-STRAIN BEHAVIOUR OF TIE CONFINED CONCRETE CONTAINING CERAMI...
A STUDY ON STRESS-STRAIN BEHAVIOUR OF TIE CONFINED CONCRETE CONTAINING CERAMI...A STUDY ON STRESS-STRAIN BEHAVIOUR OF TIE CONFINED CONCRETE CONTAINING CERAMI...
A STUDY ON STRESS-STRAIN BEHAVIOUR OF TIE CONFINED CONCRETE CONTAINING CERAMI...
ย 
Prediction of concrete materials compressive strength using surrogate models.pdf
Prediction of concrete materials compressive strength using surrogate models.pdfPrediction of concrete materials compressive strength using surrogate models.pdf
Prediction of concrete materials compressive strength using surrogate models.pdf
ย 
Effect of carbon nanotubes on the mechanical fracture parameters and microstr...
Effect of carbon nanotubes on the mechanical fracture parameters and microstr...Effect of carbon nanotubes on the mechanical fracture parameters and microstr...
Effect of carbon nanotubes on the mechanical fracture parameters and microstr...
ย 
A fracture mechanics based method for prediction of
A fracture mechanics based method for prediction ofA fracture mechanics based method for prediction of
A fracture mechanics based method for prediction of
ย 
โ€œEXPERIMENTAL STUDY ON PARTIAL REPLACEMENT OF CEMENT BY SEWAGE SLUDGE ASH AND...
โ€œEXPERIMENTAL STUDY ON PARTIAL REPLACEMENT OF CEMENT BY SEWAGE SLUDGE ASH AND...โ€œEXPERIMENTAL STUDY ON PARTIAL REPLACEMENT OF CEMENT BY SEWAGE SLUDGE ASH AND...
โ€œEXPERIMENTAL STUDY ON PARTIAL REPLACEMENT OF CEMENT BY SEWAGE SLUDGE ASH AND...
ย 
An Experimental Study on Effects of Quarry Dust as Partial Replacement of San...
An Experimental Study on Effects of Quarry Dust as Partial Replacement of San...An Experimental Study on Effects of Quarry Dust as Partial Replacement of San...
An Experimental Study on Effects of Quarry Dust as Partial Replacement of San...
ย 

More from PreetiKulkarni20

IPR- Orientation.pptx
IPR- Orientation.pptxIPR- Orientation.pptx
IPR- Orientation.pptxPreetiKulkarni20
ย 
Docks and Harbors.pptx
Docks and Harbors.pptxDocks and Harbors.pptx
Docks and Harbors.pptxPreetiKulkarni20
ย 
Hyperspectral Image Analysis for Mechanical and Chemical Properti.pdf
Hyperspectral Image Analysis for Mechanical and Chemical Properti.pdfHyperspectral Image Analysis for Mechanical and Chemical Properti.pdf
Hyperspectral Image Analysis for Mechanical and Chemical Properti.pdfPreetiKulkarni20
ย 
j.1747-1567.2010.00658.x.pdf
j.1747-1567.2010.00658.x.pdfj.1747-1567.2010.00658.x.pdf
j.1747-1567.2010.00658.x.pdfPreetiKulkarni20
ย 
Determination_of_Concrete_Properties_Usi.pdf
Determination_of_Concrete_Properties_Usi.pdfDetermination_of_Concrete_Properties_Usi.pdf
Determination_of_Concrete_Properties_Usi.pdfPreetiKulkarni20
ย 

More from PreetiKulkarni20 (8)

IPR- Orientation.pptx
IPR- Orientation.pptxIPR- Orientation.pptx
IPR- Orientation.pptx
ย 
TRE-L1.pptx
TRE-L1.pptxTRE-L1.pptx
TRE-L1.pptx
ย 
Docks and Harbors.pptx
Docks and Harbors.pptxDocks and Harbors.pptx
Docks and Harbors.pptx
ย 
Hyperspectral Image Analysis for Mechanical and Chemical Properti.pdf
Hyperspectral Image Analysis for Mechanical and Chemical Properti.pdfHyperspectral Image Analysis for Mechanical and Chemical Properti.pdf
Hyperspectral Image Analysis for Mechanical and Chemical Properti.pdf
ย 
Aryal.pdf
Aryal.pdfAryal.pdf
Aryal.pdf
ย 
Jang-Ahn2019.pdf
Jang-Ahn2019.pdfJang-Ahn2019.pdf
Jang-Ahn2019.pdf
ย 
j.1747-1567.2010.00658.x.pdf
j.1747-1567.2010.00658.x.pdfj.1747-1567.2010.00658.x.pdf
j.1747-1567.2010.00658.x.pdf
ย 
Determination_of_Concrete_Properties_Usi.pdf
Determination_of_Concrete_Properties_Usi.pdfDetermination_of_Concrete_Properties_Usi.pdf
Determination_of_Concrete_Properties_Usi.pdf
ย 

Recently uploaded

The Most Attractive Pune Call Girls Manchar 8250192130 Will You Miss This Cha...
The Most Attractive Pune Call Girls Manchar 8250192130 Will You Miss This Cha...The Most Attractive Pune Call Girls Manchar 8250192130 Will You Miss This Cha...
The Most Attractive Pune Call Girls Manchar 8250192130 Will You Miss This Cha...ranjana rawat
ย 
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...roncy bisnoi
ย 
Top Rated Pune Call Girls Budhwar Peth โŸŸ 6297143586 โŸŸ Call Me For Genuine Se...
Top Rated  Pune Call Girls Budhwar Peth โŸŸ 6297143586 โŸŸ Call Me For Genuine Se...Top Rated  Pune Call Girls Budhwar Peth โŸŸ 6297143586 โŸŸ Call Me For Genuine Se...
Top Rated Pune Call Girls Budhwar Peth โŸŸ 6297143586 โŸŸ Call Me For Genuine Se...Call Girls in Nagpur High Profile
ย 
Coefficient of Thermal Expansion and their Importance.pptx
Coefficient of Thermal Expansion and their Importance.pptxCoefficient of Thermal Expansion and their Importance.pptx
Coefficient of Thermal Expansion and their Importance.pptxAsutosh Ranjan
ย 
data_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdfdata_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdfJiananWang21
ย 
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756dollysharma2066
ย 
Booking open Available Pune Call Girls Koregaon Park 6297143586 Call Hot Ind...
Booking open Available Pune Call Girls Koregaon Park  6297143586 Call Hot Ind...Booking open Available Pune Call Girls Koregaon Park  6297143586 Call Hot Ind...
Booking open Available Pune Call Girls Koregaon Park 6297143586 Call Hot Ind...Call Girls in Nagpur High Profile
ย 
University management System project report..pdf
University management System project report..pdfUniversity management System project report..pdf
University management System project report..pdfKamal Acharya
ย 
NFPA 5000 2024 standard .
NFPA 5000 2024 standard                                  .NFPA 5000 2024 standard                                  .
NFPA 5000 2024 standard .DerechoLaboralIndivi
ย 
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...ranjana rawat
ย 
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...Dr.Costas Sachpazis
ย 
Call Now โ‰ฝ 9953056974 โ‰ผ๐Ÿ” Call Girls In New Ashok Nagar โ‰ผ๐Ÿ” Delhi door step de...
Call Now โ‰ฝ 9953056974 โ‰ผ๐Ÿ” Call Girls In New Ashok Nagar  โ‰ผ๐Ÿ” Delhi door step de...Call Now โ‰ฝ 9953056974 โ‰ผ๐Ÿ” Call Girls In New Ashok Nagar  โ‰ผ๐Ÿ” Delhi door step de...
Call Now โ‰ฝ 9953056974 โ‰ผ๐Ÿ” Call Girls In New Ashok Nagar โ‰ผ๐Ÿ” Delhi door step de...9953056974 Low Rate Call Girls In Saket, Delhi NCR
ย 
Extrusion Processes and Their Limitations
Extrusion Processes and Their LimitationsExtrusion Processes and Their Limitations
Extrusion Processes and Their Limitations120cr0395
ย 
Booking open Available Pune Call Girls Pargaon 6297143586 Call Hot Indian Gi...
Booking open Available Pune Call Girls Pargaon  6297143586 Call Hot Indian Gi...Booking open Available Pune Call Girls Pargaon  6297143586 Call Hot Indian Gi...
Booking open Available Pune Call Girls Pargaon 6297143586 Call Hot Indian Gi...Call Girls in Nagpur High Profile
ย 
Vivazz, Mieres Social Housing Design Spain
Vivazz, Mieres Social Housing Design SpainVivazz, Mieres Social Housing Design Spain
Vivazz, Mieres Social Housing Design Spaintimesproduction05
ย 
Double rodded leveling 1 pdf activity 01
Double rodded leveling 1 pdf activity 01Double rodded leveling 1 pdf activity 01
Double rodded leveling 1 pdf activity 01KreezheaRecto
ย 
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptxBSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptxfenichawla
ย 
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...Christo Ananth
ย 
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance BookingCall Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Bookingroncy bisnoi
ย 

Recently uploaded (20)

The Most Attractive Pune Call Girls Manchar 8250192130 Will You Miss This Cha...
The Most Attractive Pune Call Girls Manchar 8250192130 Will You Miss This Cha...The Most Attractive Pune Call Girls Manchar 8250192130 Will You Miss This Cha...
The Most Attractive Pune Call Girls Manchar 8250192130 Will You Miss This Cha...
ย 
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
Call Girls Pimpri Chinchwad Call Me 7737669865 Budget Friendly No Advance Boo...
ย 
Top Rated Pune Call Girls Budhwar Peth โŸŸ 6297143586 โŸŸ Call Me For Genuine Se...
Top Rated  Pune Call Girls Budhwar Peth โŸŸ 6297143586 โŸŸ Call Me For Genuine Se...Top Rated  Pune Call Girls Budhwar Peth โŸŸ 6297143586 โŸŸ Call Me For Genuine Se...
Top Rated Pune Call Girls Budhwar Peth โŸŸ 6297143586 โŸŸ Call Me For Genuine Se...
ย 
Coefficient of Thermal Expansion and their Importance.pptx
Coefficient of Thermal Expansion and their Importance.pptxCoefficient of Thermal Expansion and their Importance.pptx
Coefficient of Thermal Expansion and their Importance.pptx
ย 
data_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdfdata_management_and _data_science_cheat_sheet.pdf
data_management_and _data_science_cheat_sheet.pdf
ย 
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
FULL ENJOY Call Girls In Mahipalpur Delhi Contact Us 8377877756
ย 
(INDIRA) Call Girl Meerut Call Now 8617697112 Meerut Escorts 24x7
(INDIRA) Call Girl Meerut Call Now 8617697112 Meerut Escorts 24x7(INDIRA) Call Girl Meerut Call Now 8617697112 Meerut Escorts 24x7
(INDIRA) Call Girl Meerut Call Now 8617697112 Meerut Escorts 24x7
ย 
Booking open Available Pune Call Girls Koregaon Park 6297143586 Call Hot Ind...
Booking open Available Pune Call Girls Koregaon Park  6297143586 Call Hot Ind...Booking open Available Pune Call Girls Koregaon Park  6297143586 Call Hot Ind...
Booking open Available Pune Call Girls Koregaon Park 6297143586 Call Hot Ind...
ย 
University management System project report..pdf
University management System project report..pdfUniversity management System project report..pdf
University management System project report..pdf
ย 
NFPA 5000 2024 standard .
NFPA 5000 2024 standard                                  .NFPA 5000 2024 standard                                  .
NFPA 5000 2024 standard .
ย 
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...
The Most Attractive Pune Call Girls Budhwar Peth 8250192130 Will You Miss Thi...
ย 
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
Structural Analysis and Design of Foundations: A Comprehensive Handbook for S...
ย 
Call Now โ‰ฝ 9953056974 โ‰ผ๐Ÿ” Call Girls In New Ashok Nagar โ‰ผ๐Ÿ” Delhi door step de...
Call Now โ‰ฝ 9953056974 โ‰ผ๐Ÿ” Call Girls In New Ashok Nagar  โ‰ผ๐Ÿ” Delhi door step de...Call Now โ‰ฝ 9953056974 โ‰ผ๐Ÿ” Call Girls In New Ashok Nagar  โ‰ผ๐Ÿ” Delhi door step de...
Call Now โ‰ฝ 9953056974 โ‰ผ๐Ÿ” Call Girls In New Ashok Nagar โ‰ผ๐Ÿ” Delhi door step de...
ย 
Extrusion Processes and Their Limitations
Extrusion Processes and Their LimitationsExtrusion Processes and Their Limitations
Extrusion Processes and Their Limitations
ย 
Booking open Available Pune Call Girls Pargaon 6297143586 Call Hot Indian Gi...
Booking open Available Pune Call Girls Pargaon  6297143586 Call Hot Indian Gi...Booking open Available Pune Call Girls Pargaon  6297143586 Call Hot Indian Gi...
Booking open Available Pune Call Girls Pargaon 6297143586 Call Hot Indian Gi...
ย 
Vivazz, Mieres Social Housing Design Spain
Vivazz, Mieres Social Housing Design SpainVivazz, Mieres Social Housing Design Spain
Vivazz, Mieres Social Housing Design Spain
ย 
Double rodded leveling 1 pdf activity 01
Double rodded leveling 1 pdf activity 01Double rodded leveling 1 pdf activity 01
Double rodded leveling 1 pdf activity 01
ย 
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptxBSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
BSides Seattle 2024 - Stopping Ethan Hunt From Taking Your Data.pptx
ย 
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
Call for Papers - Educational Administration: Theory and Practice, E-ISSN: 21...
ย 
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance BookingCall Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Booking
Call Girls Walvekar Nagar Call Me 7737669865 Budget Friendly No Advance Booking
ย 

HyperspectralImaging (1).pdf

  • 1. Hyperspectral Imaging Features for Mortar Classification and Compressive Strength 1 Assessment 2 Liang Fan 1, Ming Fan 2, Abdullah Alhaj 1, Genda Chen 1,* and Hongyan Ma 1 3 1 Department of Civil, Architectural and Environmental Engineering, Missouri University 4 of Science and Technology, Rolla, MO 65401, United States; lf7h2@mst.edu; 5 ahanbc@mst.edu; mahon@mst.edu 6 2 Department of Mining and Minerals Engineering, Virginia Polytechnic Institute and 7 State University, Blacksburg, VA 24060; mingfan@vt.edu 8 * Correspondence: gchen@mst.edu; Tel.: +01-573-341-4462 9 Abstract: In this study, hyperspectral imagery with two computational algorithms are 10 proposed to classify the type of mortar and assess the in-situ strength of fresh mortar in near 11 real time. Each scanning on a mortar surface includes 30 spatial pixels selected for analysis, 12 each assigned with a light reflectance spectrum over 400 - 2500 nm. Three groups of mortar 13 samples with a water-to-cement (W/C) ratio of 0.6, 0.5 and 0.4, respectively, were cast and 14 scanned from Day 1 to 14 of curing. Reflectance data at a wavelength range of 1920 nm to 15 1980 nm, associated with the O-H chemical bond, were extracted and averaged to classify 16 the different mortar types with K-Nearest Neighbors (KNN) and Support Vector Machine 17 (SVM) algorithms and to predict their compressive strength from a regression equation. The 18 results showed that the average reflectance increased with time due to water molecules 19 reaction during curing process. The KNN classification model with K=5 had a prediction 20 accuracy of 70% to 75%, and the SVM classification model with C=1000 and ฯƒ=10 showed 21 a prediction accuracy of approximately 90%. Therefore, the SVM classification algorithm 22 is recommended for use in mortar classification. The compressive strength is well correlated 23 with the average reflectance with a coefficient of over 0.98. 24 Keywords: Hyperspectral imaging; W/C ratio; reflectance; KNN; SVM; compressive 25 strength 26 27 1. Introduction 28 Concrete is a mixture of aggregate, sand, cement and water in a certain proportion. The 29 cement and water together, referred to as cement paste, hardens through hydration reactions 30 and binds the aggregate and sand to achieve the strength of concrete in a curing process over 31 time [1]. Concrete hydration process starts immediately after concrete casting. Cement reacts 32
  • 2. with water to generate hydration products like C-S-H gel and calcium hydroxide. The 33 hydration products grow, interconnect, and bond aggregate and sand. Concrete is formless 34 and shaped to various forms of interest, when newly mixed, and durable, when cured and 35 hardened. In the construction of highway pavements, ACI Code 301-72 requires a minimum 36 of curing period for concrete pavement to ensure that the constructed roadway is safe to 37 traffic without damage [2]. In the repair and resurface of existing roadways and their 38 transportation network in an urban environment, it is imperative to determine the early-age 39 strength of concrete pavements so that the impact of roadway construction on traffic is 40 minimized. 41 Coring and pullout test are two of the conventional approaches that have been used for 42 on-site evaluation of the compressive strength of concrete. With the coring method, concrete 43 cores are acquired by drilling a concrete structure at selected locations, and tested for their 44 compressive strength [3]. During the pullout test, a metal disk is attached to the concrete 45 surface with super glue. After a short curing period, the metal disk is pulled perpendicularly 46 off the surface and the pullout force can be used to calculate the compressive strength of the 47 concrete structure [4]. The pullout force can be related to the compressive strength of 48 concrete based on a pre-determined calibration curve. Both the coring and pullout test are 49 destructive, potentially compromising the integrity of concrete structures. 50 Nondestructive approaches such as the maturity method and the ultrasonic pulse velocity 51 (UPV) have also been used to determine the compressive strength of concrete. The maturity 52 method allows the estimate of early-age compressive strength of in-place concrete in real 53 time. A maturity index as a function of curing time and temperature is determined according 54 to the ASTM C1074 Standards [5]. In applications, a reference strength-maturity curve must 55 be developed for each project-specific material in advance. With the UPV method, the 56 velocity of an ultrasonic pulse that travels through concrete is measured and converted to the 57 strength of concrete based on their pre-determined calibration curve [6]. The field application 58 of this method is limited due to the effects of voids, cracks and steel bars. 59 Hyperspectral imagery has been used to assess various conditions of concrete by imaging 60 a concrete surface and analyzing the light reflectance as a function of wavelength for each 61 pixel in an image. Such a reflectance-wavelength spectrum can be divided into many narrow 62 and continuous wavelength bands for their correlation to specific materials on the concrete 63 surface [7]. By analyzing the change of reflectance values at these prominent bands, different 64 materials can be discriminated and classified. For instance, dark gray, light gray and 65 dolomitic limestone were distinguished in the selection of Portland cement clinkers based on 66
  • 3. the reflectance variations of carbonate (CO3) and Al-OH in wavelength ranges of 2125โ€“2400 67 nm and 2170โ€“2250 nm, respectively [8]. The carbonation degradation depth of concrete was 68 estimated from reflectance values at a wavelength of 440 nm, 1500 nm, and 2340 nm [9]. 69 The total chloride content in mortar specimens was linearly related to the reflectance at a 70 wavelength of approximately 2260 nm [10]. The status of concrete (hydration, curing and 71 hardening) was determined by constructing a logistic regression model with reflectance 72 spectra [11]. 73 In the past decade, hyperspectral imaging has also been used to estimate the compressive 74 strength of concrete. For example, a partial least square regression model was developed to 75 establish the relation between concrete strength (7, 14 and 28 days) and its corresponding 76 reflectance over the entire wavelength range [12]. The reflectance spectrum of eight concrete 77 samples with various W/C ratios and curing ages moved upward with an increase of 78 compression strength [13]. In both studies, the mix designs of concrete were not introduced 79 and the relation between the compression strength and the reflectance at a characteristic 80 wavelength range was not clearly interpreted. Three groups of 28-day cured concrete 81 specimens with a W/C ratio of 0.5, 0.65 and 0.8 were differentiable by comparing absorbance 82 values (complimentary to reflectance) in a wavelength range of 1940-1970 nm [14]. In that 83 study, the relation between absorbance and compression strength was not discussed. 84 The ultimate goal of this study is to rapidly classify the type of concrete with various 85 W/C ratios in pavement construction of highways through hyperspectral imaging from an 86 unmanned aerial vehicle, and determine the early-age compressive strength of concrete 87 pavements from light reflectance spectra. The focus of this paper is to develop a dataset with 88 light reflectance and its corresponding compressive strength of mortar of various types, a 89 classification model for mortar type, and a regression curve of reflectance versus compressive 90 strength corresponding to a specific mortar type. Specifically, three groups of mortar cuboid 91 samples with a W/C ratio of 0.4, 0.5 and 0.6 were cast. For each group, five mortar samples 92 were tested for compressive strength after 1, 3, 5 7, 9, 11, 13, or 14 days of curing. Another 93 nine samples were scanned using a hyperspectral camera from Day 1 to 14. A large set of 94 reflectance data were extracted from the scanned images and used to train Nearest Neighbors 95 (KNN) and Support Vector Machine (SVM) classifiers for discrimination of three mortar 96 types. The compressive strength of each type of mortar samples was measured corresponding 97 to the hyperspectral imaging schedule and related to the light reflectance by an exponential 98 regression model developed. 99 100
  • 4. 2. Experiment Setup 101 2.1. Sample preparation 102 Three types of mortar samples were prepared and designated as C1, C2, and C3 in Table 103 1. They are a mixture of water, ordinary Portland cement and Missouri river sand with a 104 W/C/Sand weight ratio of 0.194/0.324/1.0, 0.182/0.364/1.0, and 0.165/0.415/1.0, or a W/C 105 ratio of 0.6, 0.5, and 0.4, respectively. Type I Portland cement was used as detailed in Table 106 2 for its chemical composition. The Missouri river sand used had the maximum particle size 107 of 4.75 mm, a specific gravity of 2.64, and a fineness modulus of 2.71. Freshly mixed mortar 108 was poured into standard cubic steel molds that are 50 mm ร— 50 mm ร— 50 mm in size. After 109 casting, the specimens were covered with wet burlaps and plastic sheets to prevent surface 110 cracking due to shrinkage. After 24 hours of curing, they were demolded for compressive 111 tests and hyperspectral scanning. For each type of mortar mixture, compressive tests of 40 112 samples were conducted according to the ASTM Standard C39 [15], 5 samples tested after 113 1, 3, 5, 7, 9, 11, 13, and 14 days of curing and hardening. All the samples were cured in air 114 with a temperature of 23โ€ฏยฑโ€ฏ1.7โ€ฏยฐC and a relative humidity (RH) of 50โ€ฏยฑโ€ฏ5%. Hyperspectral 115 scanning on 9 samples with each mortar mixture was conducted continuously for 13 days 116 from the end of 1st day to 14th day of curing and hardening. For each cuboid sample, only 117 the four vertical side surfaces were scanned since the horizontal top surface was relatively 118 uneven. In addition, the top surface had a thin layer of cement paste due to water bleeding 119 during mortar settlement, which made its composition different from the side surfaces [16]. 120 Table 1. Mix proportions of three types of mortar samples by weight (kg/m3 ) 121 Types of mortar samples C1 C2 C3 Water 288 270 245 Ordinary Portland cement 480 540 615 Missouri river sand 1482 1482 1482 Table 2. Mass percentage (%) of oxides in cement 122 SiO2 CaO Al2O3 Fe2O3 MgO SO3 Loss of ignition 19.8 64.2 4.5 3.2 2.7 3.4 2.6 123
  • 5. 2.2. Hyperspectral scanning 124 A wideband hyperspectral camera (Headwall Hyperspec VNIR-SWIR dual sensor) was 125 used to scan the mortar samples. The co-aligned VNIR-SWIR sensor has a broad wavelength 126 range of 400 - 2500 nm. The VNIR sensor has a spectral range of 400-1000 nm with 2.2 nm 127 in spectral resolution and the SWIR sensor has a spectral range of 900-2500 nm with 6 nm 128 in spectral resolution. Figure 1 shows the experimental setup of a cuboid mortar sample. A 129 light source (LED illumination) was set at 0.5 m away from the mortar sample and lit the 130 sample from one side (left in the photo). The hyperspectral camera was set right in front of 131 the mortar sample at 1.2 m standoff distance from the front vertical side of the mortar sample 132 for better resolution of near-distance imaging. The camera was installed on a tripod, both 133 connected to a laptop installed with Hyperspec III software to control the cameraโ€™s rotation 134 (ยฑ5ยฐ) in the horizontal plane and collect images continuously. A grey tarp was set right behind 135 the mortar sample as a reference. 136 137 138 Figure 1. Test setup of a cuboid mortar sample with illumination light, a hyperspectral 139 camera, a laptop computer, and a grey tarp. 140 Prior to each sample scanning, the hyperspectral camera was calibrated through the 141 collection and processing of dark and white reference data. Measuring electric current in the 142 camera system, the dark reference was collected with the camera lens covered, and deducted 143 from any scanned image to cleanse noise. The white reference was used to get a white balance 144 to enhance imaging quality. It was collected by aiming the camera lens at the grey tarp with 145 a reflectance of 32%. The grey tarp was chosen in this study since its color was close to that 146 of the mortar samples. Frame period and exposure time were adjusted to ensure that 60% of 147 the saturated light intensity was detected by using the grey tarp since a lack of light intensity 148
  • 6. can generate too many bad pixels to correct mathematically. The rotation angle was adjusted 149 so that the camera can scan the mortar surface area of interest at a fixed standoff distance of 150 1.2 m. The rotation speed of the camera was also adjusted until no distorted shapes or forms 151 were seen in the captured image. 152 At the completion of each sample scanning, the scanned data files were transferred from 153 the camera (480 GB solid-state drive) to the laptop computer. SpectralView software was 154 then used to extract the reflectance spectrum for each pixel in the image by: 155 Calibrated Reflectance = ๐‘…๐‘Ž๐‘คโˆ’๐ท๐‘Ž๐‘Ÿ๐‘˜ ๐‘Šโ„Ž๐‘–๐‘ก๐‘’โˆ’๐ท๐‘Ž๐‘Ÿ๐‘˜ ร— ๐‘Šโ„Ž๐‘–๐‘ก๐‘’ ๐‘…๐‘’๐‘“๐‘’๐‘Ÿ๐‘’๐‘›๐‘๐‘’ ๐‘…๐‘’๐‘“๐‘™๐‘’๐‘๐‘ก๐‘Ž๐‘›๐‘๐‘’ C๐‘Ž๐‘™๐‘–๐‘๐‘Ÿ๐‘Ž๐‘ก๐‘–๐‘œ๐‘› (1) where Raw is the raw reflectance spectrum without processing, Dark means the dark 156 reference spectrum, White means the white reference spectrum, and White Reference 157 Reflectance Calibration denotes the maximum reflectance of white reference spectrum to 158 ensure no saturation in measurement. The software SpectralView automatically calculates 159 the normalized reflectance using the dark/white reference spectra. 160 2.3 Data classification techniques 161 Two classification models with KNN and SVM algorithms were established to 162 distinguish various types (W/C ratios) of mortar samples from the reflectance dataset as 163 shown in Figure 2. The reflectance dataset is a group of data with each datum showing a 164 reflectance value and its corresponding class label (W/C ratio). In this study, 80% of the 165 reflectance data were used for training and the remaining 20% were used for testing of the 166 classification models. Both KNN and SVM algorithms were trained in Python to construct 167 the classification models. The established classification models were then used to predict the 168 label (Class A, Class B or Class C) for a W/C ratio of 0.6, 0.5, or 0.4 given a test example of 169 known reflectance value. 170 171 Figure 2. Reflectance data classification using classification algorithms. 172 2.3.1 KNN 173 The KNN algorithm computes the proximity of a test example z to K data points in the 174
  • 7. training set, which are closest to z. The test example is classified based on the majority class 175 label of its K nearest neighbors [17]. Weights are assigned to the contributions of the 176 neighbors so that the impact of data depends on their distances to the test example. Choosing 177 the right parameter K is important to ensure a better accuracy in classification. A small K can 178 result in overfitting due to noise in the training data, whereas a large K can lead to 179 misclassification because the nearest neighbors may include data that are located far away 180 from its neighborhood [17]. 181 2.3.2 SVM 182 The SVM algorithm creates a line in two-dimensional planes, a plane in three- 183 dimensional spaces, or more generally a hyperplane to divide the data into several classes 184 [18]. Support vectors are the data points nearest to the hyperplane. The distance between the 185 hyperplane and the nearest data is called margin [18]. The goal of SVM is to choose a 186 hyperplane with the maximum margin. To briefly describe the SVM technique, a linear 187 classifier is introduced first and then extended to the nonlinear classifier. Next, the maximum 188 margin of a hyperplane is described. 189 Linear classifier is used to find a line or a plane (hyperplane) to separate dataset 190 {๐ฑ๐‘–, ๐‘ฆ๐‘–}๐‘–=1 ๐‘› into two classes. Here, ๐ฑ๐‘– is the ๐‘–๐‘กโ„Ž vector in the given dataset, ๐‘ฆ๐‘– is the label 191 associated with ๐ฑ๐‘–. The hyperplane is defined as [18]: 192 ๐‘“(๐ฑ) = ๐ฐ๐ฑ + ๐‘ = 0, ๐ฐ๐ฑ = โˆ‘ ๐‘ค๐‘–๐‘ฅ๐‘– ๐‘– (2) where w is a weight vector, and b is a bias. As illustrated in Figure 3 for the case of two- 193 dimensional plane x1x2, f(x) =0 is a line that divides the entire dataset into two classes: f(x) 194 >0 and f(x) <0. 195 196 Figure 3. A linear classifier with maximum margins that divides the data into two sets. 197 198
  • 8. When the data cannot be separated by a linear classifier, they can be mapped to a higher 199 dimension and converted to linearly separable data through a projection function ๐œ‘ [18, 19]. 200 The classifier then becomes: 201 ๐‘“(๐ฑ) = ๐ฐ๐œ‘(๐ฑ) + ๐‘ (3) As the high-dimensional projection function is complicated to compute, this classifier is 202 projected back to the original dimension through a transformation known as the kernel 203 function. In this case, the weight vector can be expressed into a linear combination of the 204 training data [18]: 205 ๐ฐ = โˆ‘ ๐›ผ๐‘– ๐‘› ๐‘–=1 ๐œ‘(๐ฑ๐‘–) (4) where ๐›ผ๐‘– is the coefficient related to a decision boundary. The kernel function is defined as 206 [18]: 207 ๐‘˜(๐ฑ๐‘–, ๐ฑ) = ๐œ‘(๐ฑ๐‘–)๐œ‘(๐ฑ) (5) The classifier then transforms to: 208 ๐‘“(๐ฑ) = โˆ‘ ๐›ผ๐‘– ๐‘› ๐‘–=1 ๐œ‘(๐ฑ๐‘–)๐œ‘(๐ฑ) + ๐‘ = โˆ‘ ๐›ผ๐‘– ๐‘› ๐‘–=1 ๐‘˜(๐ฑ๐‘–, ๐ฑ) + ๐‘ (6) Two kernels are widely used in the literature for various applications: polynomial kernel and 209 Gaussian kernel. A polynomial kernel is defined as [18]: 210 ๐‘˜(๐ฑ๐‘–, ๐ฑ) = (๐ฑ๐ฑ๐‘– + 1)๐‘‘ (7) A Gaussian kernel is defined as [18, 19]: 211 ๐‘˜(๐ฑ๐‘–, ๐ฑ) = exp (โˆ’ โ€–๐ฑโˆ’๐ฑ๐‘–โ€–๐Ÿ 2๐œŽ2 ) (8) where d is the degree of polynomial kernel and ฯƒ is a parameter that controls the width of 212 Gaussian kernel. Both parameters control the flexibility of the classifier. When ฯƒ is increased, 213 a greater curvature is introduced to the decision boundary but overfitting will occur if ฯƒ is 214 too large. 215 SVM looks for a higher margin to get a better classification result for the testing data. 216 The margin of a hyperplane f (x) is defined as: 217
  • 9. ๐‘š(๐‘“) = 1 โ€–๐ฐโ€– (9) As indicated in Equation (9), to maximize the margin of the classifier is equivalent to 218 minimizeโ€–๐ฐโ€–2 . The maximum margins are the margins that push up against the support 219 vectors. To ensure that the linearly-separable data are classified correctly, the maximum 220 margin and its constraint are defined as [18, 19]: 221 Minimize 1 2 โ€–๐ฐโ€–2 (10) Subject to: ๐‘ฆ๐‘–(๐ฐ๐ฑ + ๐‘) โ‰ฅ 1 ๐‘– = 1, โ€ฆ , ๐‘›. (11) When the data are not completely separable, the constraint is relaxed and a greater margin 222 can be achieved by [18, 19]: 223 Minimize 1 2 โ€–๐ฐโ€–2 + ๐ถ โˆ‘ ๐œ‰๐‘– ๐‘› ๐‘–=1 (12) Subject to: ๐‘ฆ๐‘–(๐ฐ๐‘ฅ + ๐‘) โ‰ฅ 1 โˆ’ ๐œ‰๐‘– (13) where ๐œ‰๐‘– (0โ‰ค ๐œ‰๐‘– โ‰ค 1) is the margin error that allows an example to be in the margin and C 224 is the penalty that lowers the misclassification rate. When C is increased, a smaller margin 225 error is achieved. C needs to be adjusted to ensure the maximum margin with a minimum 226 margin error [18, 19]. 227 3. Results and Discussion 228 3.1. Hyperspectral information 229 Figure 4 shows the raw hyperspectral image of one mortar specimen, the image after 230 subtraction of dark reference, and the image after dark and white reference deductions. The 231 sensor current measured from the dark reference can induce perturbation and generate a noisy 232 and drifted spectrum. The white reference can rectify illumination non-uniformity and non- 233 flatness of a spectrum. Removal of the dark reference and the white reference can correct 234 the image and produce right reflectance spectra. For each type of mortar, 9 cuboid samples 235 were prepared, 4 side faces of each sample were scanned, and 30 spectra were extracted over 236 a 50 mm ร— 50 mm side surface area, totaling 1080 spectra for each scanning day. The spectra 237
  • 10. were collected from the flat surface area only to avoid any non-uniform illumination from 238 uneven spots. 239 (a) (b) (c) Figure 4. Hyperspectral image of a mortar specimen: (a) raw, (b) after subtraction of dark 240 reference, and (c) after dark and white reference deduction. 241 Figure 5 shows the average reflectance spectra of samples with three different W/C ratios 242 over a period of 14 days. Each line represents the average reflectance spectrum of 1080 243 spectra in the wavelength range of 1200 nm to 2400 nm. The average spectra can reduce 244 potential biases and are more representative of the scanning surface. As seen in Figure 5, the 245 reflectance value on the average spectra rapidly ascends from Day 1 to Day 3 and then 246 gradually increased till Day 14 of test. After 1 day of testing, the samples were scanned after 247 they were demolded and put in air at the room temperature for 1 hour. The higher moisture 248 content on the sample surface resulted in the lower reflectance value due to water absorption. 249 The rapid increase of reflectance from Day 1 to Day 3 is because the hydration process during 250 this period rapidly consumes more water compared with that at a later stage. 251 252 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11 0.12 1200 1400 1600 1800 2000 2200 2400 Reflectance (R) Wavelength (nm) Day 1 Day 3 Day 5 Day 7 Day 9 Day 11 Day 13 Day 14 (a)
  • 11. 253 254 Figure 5. The average reflectance spectra over a wavelength of 1200 nm to 2400 nm for 255 samples with a W/C ratio of: (a) 0.6, (b) 0.5, and (c) 0.4. 256 When shot on the surface of materials, some of the incident light leads to vibration of 257 molecules and is absorbed by the chemical bond between atoms in the molecules. In the Near 258 Infrared Region (NIR) (from 780 nm to 2500 nm), higher vibrational energy is acquired to 259 absorb the light, which stimulates the overtones and combinations of fundamental vibrations 260 [20, 21]. Basically, overtones and combinations of the vibrations of C-H, O-H, N-H, and S- 261 H chemical bonds dominate NIR spectroscopy with each chemical bond corresponding to a 262 wavelength region for light absorbance [21]. The combination of OH and H2O corresponds 263 to the region of 1900 nm to 2000 nm [21-24], which is of particular interest in this study. The 264 reflectance change in this wavelength range can be used to track the change of H2O molecules 265 due to hydration consumption in the process of mortar curing. The reflectance values over 266 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11 0.12 1200 1400 1600 1800 2000 2200 2400 Reflectance (R) Wavelength (nm) Day 1 Day 3 Day 5 Day 7 Day 9 Day 11 Day 13 Day 14 0.06 0.07 0.08 0.09 0.1 0.11 0.12 1200 1400 1600 1800 2000 2200 2400 Reflectance (R) Wavelength (nm) Day 1 Day 3 Day 5 Day 7 Day 9 Day 11 Day 13 Day 14 (b) (c)
  • 12. the wavelength range of 1920 nm to 1980 nm as shown in the marked red circles in Figure 5 267 are averaged and defined as dR (1920-1980). The average reflectance dR (1920-1980) with 268 an error bar of ยฑ one standard deviation for three types of mortar over the curing period of 269 14 days are plotted in Figure 6. Logarithmic regression was conducted to fit into the 270 experimental data for the three types of mortar. R2 value for each type of mortar is higher 271 than 90%, indicating good fitting of the regression curve. The overall reflectance during the 272 14 days shows an increasing trend when the W/C ratio is reduced from 0.6 to 0.4 since the 273 mortar with lower W/C ratio has less water content. For each type of mortar, the average 274 reflectance increases because water is reacted and reduced during the hydration process. As 275 less water is left on the sample surface, less light is absorbed by water molecules and 276 reflectance of the light is increased. Therefore, the regression curve can be used to predict 277 the curing process for mortar samples. 278 Figure 6. The average reflectance dR (1920-1980) over a curing period of 14 days for 279 samples with W/C ratios of: (a) 0.6, (b) 0.5, and (c) 0.4. 280 3.2 Classification results with KNN and SVM 281 Both KNN and SVM algorithms were used to first establish classification models with 282 the training data, and then predict classification labels for the test data. In the KNN algorithm, 283 the parameter K varied from 1 to 40 in model training. Its corresponding prediction 284 y = 0.007ln(x) + 0.07 Rยฒ = 0.92 W/C=0.6 0 0.02 0.04 0.06 0.08 0.1 0.12 0 2 4 6 8 10 12 14 Average reflectance dA (1920-1980) Curing Time (Day) y = 0.011ln(x) + 0.07 Rยฒ = 0.93 W/C=0.5 0 0.02 0.04 0.06 0.08 0.1 0.12 0 2 4 6 8 10 12 14 Average reflectance dA (1920-1980) Curing Time (Day) y = 0.007ln(x) + 0.09 Rยฒ = 0.91 0 0.02 0.04 0.06 0.08 0.1 0.12 0 2 4 6 8 10 12 14 Average reflectance dA (1920-1980) Curing Time (Day) (b) (a) (c)
  • 13. accuracies after 1, 4, 7, and 13 days of testing are presented in Figure 7. In this study, accuracy 285 is defined by the relative difference in percentage between the number of correct predictions 286 and the number of actual test data. It is observed that K = 5 yields the highest prediction 287 accuracy for Day 1, Day 7, and Day 13, and K = 6 provides the highest accuracy for Day 4. 288 Overall, K = 5 is chosen for the KNN classification model. Figure 8 shows the predicted 289 classifications after 1, 7, and 13 day of testing with K = 5 and after 4 day of testing with K = 290 6. In Figure 8, โ€˜Trueโ€™ represents the test data and โ€˜Predโ€™ symbolizes the prediction data with 291 the trained KNN classification model. The predicted classification (yellow triangles) and the 292 actual classification (blue squares) are in general agreement. 293 294 Figure 7. Prediction accuracies as a function of K after 1, 4, 7, and 13 day of testing. 295 296 45 50 55 60 65 70 75 80 0 5 10 15 20 25 30 35 40 Accuracy (%) K value Day 1 Day 4 Day 7 Day 13 (b) (a)
  • 14. Figure 8. Classification predictions after (a) 1, (b) 4, (c) 7, and (d) 13 day of testing. 297 To test the applicability of parameter K, K=5 is applied to predict the classifications of 298 the test data after 2, 5, 8, and 14 day of testing, and the predicted classifications are compared 299 in Figure 9 with their actual classifications. It can be seen from Figure 9 that the predictions 300 are in good agreement with the actual classifications. Specifically, the prediction accuracies 301 during the 4 days of testing range from 70% to 75% as shown in Figure 10, which falls into 302 the same range of accuracies achieved after 1, 4, 7, and 13 day of testing. This comparison 303 indicates that K = 5 is the best fit for the KNN model. Due to low prediction accuracies (< 304 75%) for the KNN model, the SVM algorithm is attempted to improve the prediction 305 accuracy for mortar classification. 306 (d) (c) (a) (b)
  • 15. Figure 9. Classification predictions with K = 5 after (a) 2, (b) 5, (c) 8, and (d) 14 day of 307 testing. 308 309 Figure 10. Prediction accuracies with K = 5 after 2, 5, 8, and 14 day of testing. 310 In this study, the average reflectance dR(1920-1980 nm) and its corresponding Min-Max 311 normalized reflectance calculated from Equation (14) forms two features in x1-x2 plane. The 312 representative training data in the feature plane after 13 day of testing are presented in Figure 313 11. It can be seen that the three classes of training data are mainly distributed along three 314 straight lines due to the correlation of two features selected. The relatively low reflectance 315 for a w/c ratio of 0.6 can be easily separated from the other classes using a linear separator. 316 However, the two straight lines representing the feature data in other two classes are crossed 317 each other, making it impossible to separate them directly with a linear classifier. Therefore, 318 a nonlinear classifier with the Gaussian kernel and a relaxed constraint were used to obtain a 319 high prediction accuracy of the classification model. 320 0 20 40 60 80 100 Accuracy (%) Day 2 Day 5 Day 8 Day 14 (c) (d)
  • 16. Min-Max normalized reflectance= dR (1920โˆ’1980) โˆ’Min. Max.โˆ’Min. ร— 100% (14) 321 where Min. and Max. represent the minimum reflectance and the maximum reflectance of 322 the whole data set during each test day, respectively. dR (1920 โˆ’ 1980) represents the 323 average reflectance over a wavelength range of 1920 nm to 1980 nm, corresponding to each 324 datum used in the dataset. 325 326 Figure 11. Training data in the feature plane after 13 day of testing. 327 To improve the accuracy of classification predication, the hyperparameters C and ฯƒ in 328 the SVM model were first tuned. Figure 12 shows the prediction accuracies after 1, 4, 7, and 329 13 day of testing when C is examined from 10-2 to 103 and ฯƒ is tested from 10-3 to 10. It can 330 be observed from Figure 12 that C=103 and ฯƒ=10 yield the highest prediction accuracies of 331 94.5%, 89.9%, 95%, and 87.4% after 1, 4, 7, and 13 day of testing, respectively. The 332 predicted classifications from the SVM model are compared in Figure 13 with the actual 333 classifications displayed when C = 103 and ฯƒ = 10. Compared with the classification results 334 in Figures 8 and 9 predicted from the KNN model, the overlapped range between the 335 predicted data and test data in Figure 12 is enlarged with a higher accuracy. Therefore, C = 336 103 and ฯƒ = 10 are chosen for the SVM classification model. 337
  • 17. Figure 12. Prediction accuracies as a function of C and ฯƒ in the logarithmic format after (a) 338 1, (b) 4, (c) 7, and (d) 13 day of testing. 339 340 (a) (b) (c) (d) (a) (b)
  • 18. Figure 13. Classification predictions of the test data after (a) 1 , (b) 4 , (c) 7 , and (d) 13 day 341 of testing with C = 103 and ฯƒ = 10. 342 343 To evaluate their applicability, the two hyperparameters C = 103 and ฯƒ = 10 are applied 344 to the test data collected after 2, 5, 8, and 14 day of testing. The predicted results are compared 345 in Figure 14 with their actual classifications. Figure 15 shows 93.4%, 92.9%, 93.0%, and 346 92.5% prediction accuracies after 2, 5, 8, and 14 day of testing, respectively. Therefore, C = 347 103 and ฯƒ = 10 in the SVM model are appropriate for the classification of mortar. 348 (c) (d) (a) (b)
  • 19. Figure 14. Classification predictions with C = 103 and ฯƒ = 10 after (a) 2, (b) 5, (c) 8, and (d) 349 14 day of testing. 350 351 Figure 15. Prediction accuracies with C = 103 and ฯƒ = 10 after 2, 5, 8, and 14 day of testing. 352 3.3 Relation between average reflectance and compressive strength of mortar samples 353 For each W/C ratio, five mortar samples were tested in compression after 1, 3, 5, 7, 9, 354 11, 13, and 14 day of curing, following the ASTM C39 Guideline. The average and standard 355 deviation of the five compressive strengths of mortar for each testing day were determined. 356 The average strength with an error bar of ยฑ one standard deviation is displayed in Figure 16 357 as a function of the curing day. Overall, the compressive strength of mortar decreases with 358 the increase of W/C ratio. For each W/C ratio, the compressive strength increases 359 logarithmically with the curing time. The short error bars in Figure 16 indicate consistent 360 data obtained from the compressive tests. 361 0 20 40 60 80 100 Accuracy (%) Day 2 Day 5 Day 8 Day 14 (c) (d)
  • 20. 362 Figure 16. Compressive strengths of mortar samples with W/C ratios of 0.6, 0.5, and 0.4. 363 After 1, 3, 5, 7, 9, 11, 13, and 14 day of curing, both the compressive strength and the 364 average reflectance dA (1920-1980) were obtained. Their correlations for W/C ratios of 0.6, 365 0.5, and 0.4 are drawn in Figure 17. Exponential regression of the test data was conducted 366 for each W/C ratio with R2 larger than 90%. These relations indicate that the average 367 reflectance can be used to predict the compressive strength of mortar. Such relationships can 368 potentially be applied to predict the strength of early-age concrete pavement in practice. 369 y = 10.1ln(x) + 4.7 Rยฒ = 0.98 y = 11.2ln(x) + 6.6 Rยฒ = 0.99 y = 13.4ln(x) + 10.1 Rยฒ = 0.99 0 10 20 30 40 50 0 2 4 6 8 10 12 14 Compression strength (MPa) Curing time (Day) 0.6 0.5 0.4 y = 0.041e79x Rยฒ = 0.99 W/C=0.6 0 10 20 30 40 0.04 0.08 0.12 Compression Stregnth (MPa) Average Reflenctance dA (1920-1980) y = 0.214e51x Rยฒ = 0.98 W/C=0.5 0 10 20 30 40 0.04 0.08 0.12 Compression Stregnth (MPa) Average Reflenctance dA (1920-1980) (a) (b)
  • 21. Figure 17. Average reflectance dA (1920-1980) versus the compressive strength of mortar 370 with a W/C ratio of: (a) 0.6, (b) 0.5, and (c) 0.4 over 14 days. 371 3.4 Further work 372 Concrete samples with a wide range of mixture designs will be scanned with 373 hyperspectral camera. The SVM classification model will be kept updated and improved with 374 the reflectance data extracted from newly scanned concrete samples. The reflectance dataset 375 needs to be enlarged to include as many types of concrete as possible. Field test will be 376 conducted with hyperspectral camera installed on UAV to establish a similar correlation 377 between reflectance and compressive strength under actual in-situ conditions. 378 4. Conclusions 379 This study utilized a hyperspectral camera to scan the surface of mortar samples and 380 extract the reflectance data for the classification of different types of mortar and the 381 prediction of their compressive strength. Based on experimental data, regression analysis, 382 and classification results, the following conclusions can be drawn: 383 โ€ข The reflectance spectra over a wavelength range of 1200 nm to 2400 nm are 384 consistently shifted upward as mortar samples are cured and hardened over time. In 385 particular, the average reflectance in 1920-1980 nm wavelength increases 386 logarithmically with time because water molecules on each sample surface are 387 gradually reacted during the hydration process. As the water content of mortar is 388 reduced, less light is absorbed and more light is reflected. The average reflectance 389 can be accurately predicted from the curing time with their correlation coefficient 390 of over 0.9. 391 โ€ข The KNN classification model with K=5 represents the best fit to the test data taken 392 from mortar samples with a W/C ratio of 0.4 to 0.6. The classification (W/C ratio) 393 y = 0.045e66x Rยฒ = 0.99 W/C=0.4 0 10 20 30 40 50 0.04 0.08 0.12 Compression Stregnth (MPa) Average Reflenctance dA (1920-1980) (c)
  • 22. prediction accuracies are in the range of 70% to 75%. The SVM classification model 394 with C=103 and ฯƒ=10 are the best fit to the test data with a prediction accuracy of 395 approximately 90%. Thus, the SVM classification model is recommended to 396 classify various types of mortar. 397 โ€ข The compressive strength of mortar can be exponentially related to the average 398 reflectance in 1920-1980 nm wavelength with a correlation coefficient of over 0.95. 399 With this strong exponential relation, hyperspectral imaging can be used as a rapid 400 and nondestructive evaluation tool to predict the compressive strength of mortar. 401 With further studies on concrete specimens, hyperspectral imaging is promising for 402 the prediction of early-age strength of concrete pavement in practical applications. 403 Author Contributions: Conceptualization, L.F., G. C. and H. M.; methodology, L.F. and 404 M.F.; validation, L.F. M.F. and G. C.; formal analysis, L.F. and M.F.; investigation, L.F., 405 M.F. and A.A.; resources, G. C.; data curation, L.F. and M.F.; writingโ€”original draft 406 preparation, L.F. and M.F.; writingโ€”review and editing, L.F., M.F. and G.C.; visualization, 407 L.F., M.F. and G.C.; supervision, G.C.; project administration, G.C.; funding acquisition, 408 G.C. In general, the first two authors contributed equally to the preparation of this paper. 409 Funding: Financial support to complete this study was provided by the U.S. Department of 410 Transportation, Office of the Assistant Secretary for Research and Technology (OST-R) 411 under the Auspices of the INSPIRE University Transportation Center under Grant No. 412 69A3551747126 at Missouri University of Science and Technology. The findings and 413 opinions expressed in this paper are solely those of the authors and do not represent the 414 official policy or position of the USDOT/OST-R, or any State or other entity. 415 Conflicts of Interest: The authors declare no conflict of interest. 416 References 417 1. Bullard, J.W.; Jennings, H.M.; Livingston, R.A.; Nonat, A.; Scherer, G.W.; Schweitzer, 418 J.S.; Scrivener, K.L.; Thomas, J.J. Mechanisms of cement hydration. Cement and 419 concrete research 2011, 41, pp.1208-1223. 420 2. ACI. Specifications for structural concrete for buildings. American Concrete Institute 421 (ACI) Committee 301-72, 1972. 422 3. ASTM. Standard test method for obtaining and testing drilled cores and sawed beams of 423 concrete. American Society for Testing and Materials (ASTMs) C42/C42M-18a, 2018. 424
  • 23. 4. ASTM. Standard test method for tensile strength of concrete surfaces and the bond 425 strength or tensile strength of concrete repair and overlay materials by direct tension 426 (pull-off method). American Society for Testing and Materials (ASTMs) 427 C1583/C1583M-13, 2013. 428 5. ASTM. Standard Practice for Estimating Concrete Strength by the Maturity Method. 429 American Society for Testing and Materials (ASTMs) C1074-19, 2019. 430 6. Kewalramani, M. A.; Gupta, R. Concrete compressive strength prediction using 431 ultrasonic pulse velocity through artificial neural networks. Automation in Construction 432 2006, 15, pp.374-379. 433 7. Eismann, M. T. Hyperspectral remote sensing, SPIE Press, Bellingham, Washington 434 USA, 2012; pp.1-20. 435 8. Zaini, N.; Meer, F. V. D.; Ruitenbeek, F.V.; Smeth, B. D.; Amri, F.; Lievens. C. An 436 alternative quality control technique for mineral chemistry analysis of Portland cement- 437 grade limestone using shortwave infrared spectroscopy. Remote sensing 2016, 8, pp.950- 438 966. 439 9. Arita, J.; Sasaki, K.; Endo, T.; Yasuoka, Y. Assessment of concrete degradation with 440 hyper-spectral remote sensing. The 22nd Asian Conference on Remote Sensing, 441 Singapore, 2001, pp. 5-9. 442 10. Kohri, M.; Ueda, T.; Mizuguchi, H. Application of a near-infrared spectroscopic 443 technique to estimate the chloride ion content in mortar deteriorated by chloride attack 444 and carbonation. Journal of Advanced Concrete Technology 2010, 8, pp.15-25. 445 11. Brook, A.; Ben-Dor, E. Reflectance spectroscopy as a tool to assess the quality of 446 concrete in situ. Journal of Civil Engineering and Construction Technology 2011, 8, 447 pp.169-188. 448 12. Brook, A.; Ben-Dor, E. Reflectance spectroscopy as a tool to assess the strength of high- 449 performance concrete in situ. Journal of Civil Engineering and Construction Technology 450 2012, 7, pp.195-203. 451 13. Lee, J.D.; Dewitt, B.A.; Lee, S.S.; Bhang, K.J.; Sim, J.B. Analysis of concrete 452 reflectance characteristics using spectrometer and VNIR hyperspectral camera. 453 International Archives of the Photogrammetry, Remote Sensing and Spatial Information 454 Sciences 2012, 39, B7. 455 14. Zahiri, Z., Laefer, D.F.; Gowen, A. The feasibility of short-wave infrared spectrometry 456 in assessing water-to-cement ratio and density of hardened concrete. Construction and 457 Building Materials 2018, 185, 661-669. 458
  • 24. 15. ASTM. Standard test method for compressive strength of cylindrical concrete 459 specimens. American Society for Testing and Materials (ASTMs) C39/C39M-18, 2018. 460 16. Josserand, L.; Larrard, F.D. A method for concrete bleeding measurement. Materials 461 and Structures 2004, 37, 666. 462 17. Maglogiannis, I.G. Emerging artificial intelligence applications in computer 463 engineering: real word AI systems with applications in EHealth, HCI, information 464 retrieval and pervasive technologies Vol. 160. Ios Press, Amsterdam Netherlands, 2007; 465 pp.11-12. 466 18. Ben-Hur, A.; Weston, J. A userโ€™s guide to support vector machines. In Data mining 467 techniques for the life sciences; Carugo, O., Frank Eisenhaber, F.; Humana Press, New 468 York, NY USA, 2010; pp.223-239. 469 19. Amami, R.; Ayed, D. B.; Ellouze, N. Practical selection of SVM supervised parameters 470 with different feature representations for vowel recognition. arXiv preprint 2015, 471 arXiv:1507.06020. 472 20. Atkins, P.W; Paula, J. Atkinsโ€™ Physical Chemistry, 7th ed. Oxford University Press, 473 Cambridge, United Kingdom, 2002; pp. 320-352. 474 21. Okparanma, R. N.; Araka, P. P.; Ayotamuno, J. M.; Mouazen, A. Towards enhancing 475 sustainable reuse of pre-treated drill cuttings for construction purposes by near-infrared 476 analysis: A review. Journal of Civil Engineering and Construction Technology 2018, 3, 477 pp.19-39. 478 22. Walling, P. L.; Dabney, J. M. Moisture in skin by near-infrared reflectance spectroscopy. 479 Journal of the Society of Cosmetic Chemists 1989, 40, 151-171. 480 23. Kohri, M.; Ueda, T.; Mizuguchi, H. Application of a near-infrared spectroscopic 481 technique to estimate the chloride ion content in mortar deteriorated by chloride attack 482 and carbonation. Journal of Advanced Concrete Technology 2010, 1, 15-25. 483 24. Fan, L.; Alhaj, A.; Ma, H.; Chen, G. Assessing moisture content on the surface of mortar 484 samples from hyperspectral imaging. The 9th International Conference on Structural 485 Health Monitoring of Intelligent Infrastructure, Saint Louis USA, 2019; pp. 1150-1155. 486 487