You are on page 1of 16

Mobile Hyperspectral Imaging for

Material Surface Damage Detection


Sameer Aryal 1; ZhiQiang Chen, M.ASCE 2; and Shimin Tang 3

Abstract: Many machine vision–based inspection methods aim to replace human-based inspection with an automated or highly efficient
procedure. However, these machine-vision systems have not been endorsed entirely by civil engineers for deployment in practice, partially
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

due to their poor performance in detecting damage amid other complex objects on material surfaces. This work developed a mobile hyper-
spectral imaging system which captures hundreds of spectral reflectance values in a pixel in the visible and near-infrared (VNIR) portion of
the electromagnetic spectrum. To prove its potential in discriminating complex objects, a machine learning methodology was developed with
classification models that are characterized by four different feature extraction processes. Experimental validation showed that hyperspectral
pixels, when used conjunctly with dimensionality reduction, possess outstanding potential for recognizing eight different surface objects
(e.g., with an F1 score of 0.962 for crack detection), and outperform gray-valued images with a much higher spatial resolution. The authors
envision the advent of computational hyperspectral imaging for automating damage inspection for structural materials, especially when
dealing with complex scenes found in built objects in service. DOI: 10.1061/(ASCE)CP.1943-5487.0000934. © 2020 American Society
of Civil Engineers.
Author keywords: Machine vision; Hyperspectral imaging; Damage detection; Machine learning; Dimensionality reduction.

Introduction inspection, including the Manual for Bridge Element Inspection


(AASHTO 2019) and the Guidelines for Inspecting Complex Com-
Civil engineering structures are vital for a society’s prosperity and ponents of Bridges (Leshko 2002). Due to the low cost and ubiqui-
quality of life in general. To ensure the serviceability and safety of tous availability of optical sensors (i.e., digital cameras), it is no
structures, different technology-based inspection and monitoring surprise that optical imaging has become widely adopted for struc-
methods have been researched. These technologies generally can tural inspection, wherein digital images are recorded for records or
be categorized into two types. The first is nondestructive evaluation postinspection analysis (Olsen et al. 2016). Subsequently, this has
(NDE) through advanced sensing technologies (e.g., X-ray, micro- motivated the development of machine vision techniques to aid or
wave, thermal, and ultrasonic imaging) (Cawley 2018), most of automate engineering inspection for civil structures.
which aim to detect subsurface anomalies or damage. The second Machine vision is a technical field that concerns the develop-
category includes various structural health monitoring (SHM) ment of digital imaging methods and the use of image processing
methods (Sohn et al. 2003), which seek to monitor the dynamic or computer vision algorithms for the extraction of useful informa-
responses and identify the intrinsic parameters or changes in struc- tion from images (Morris 2004). With the advent of early digital
tures. Even with these technology-based inspection or monitoring cameras, researchers in the 1980s and 1990s used simple digital
methods, the reality is that, at least for transportation structures that filters, including various edge-detection methods, for image-based
are managed by DOT agencies in most states in the US, manual or damage detection for structural materials (Cheng and Miyojim
visual inspection is considered the mainstream approach (Olsen 1998; Ritchie 1987; Ritchie 1990). To further automate the process
et al. 2016). This mainly is because surface damage (e.g., cracks of image capturing, researchers also strive to develop operationally
and corrosion) on structural materials, compared with subsurface efficient methods that are expected to mitigate the human cost.
damage, is more commonplace and can be viewed as a clue to in- These methods include ground vehicle–based imaging, aerial ve-
ternal damage or possible degrading of the system integrity. For hicle–based imaging, and crowdsourcing-based imaging (e.
bridges, standard manuals and guidelines exist for practical g., Isawa et al. 2005; Kim et al. 2015; Lattanzi and Miller
2013; Ozer et al. 2015; Tung et al. 2002; Zhang and Elaksher
1
Graduate Student, Dept. of Civil and Mechanical Engineering, Univ. of 2012). For example, Ho et al. (2013) developed a system with
Missouri Kansas City, FH 5110 Rockhill Rd., Kansas City, MO 64110. three cameras attached to a cable-climbing robot to detect surface
Email: sayd8@mail.umkc.edu damage. Yeum and Dyke (2015) proposed the use of unmanned
2
Associate Professor, School of Computing and Engineering, Univ. of aerial vehicles (UAVs) for remote imaging and damage detection.
Missouri Kansas City, FH 5110 Rockhill Rd., Kansas City, MO 64110 (cor- Chen et al. (2015) proposed a mobile-cloud infrastructure enabled
responding author). ORCID: https://orcid.org/0000-0002-0793-0089. approach that exploits collaborative mobile and cloud computing to
Email: chenzhiq@umkc.edu harness crowdsourcing for damage inspection.
3
Ph.D. Candidate, Dept. of Computer Science and Electrical Engineer- With the abundance of these optical imaging platforms, one prom-
ing, Univ. of Missouri Kansas City, FH 5110 Rockhill Rd., Kansas City,
ising fact is the ease of obtaining imagery databases. Besides the ba-
MO 64110. Email: st78d@mail.umkc.edu
Note. This manuscript was submitted on March 30, 2020; approved on sic image processing methods, these databases enable the adoption of
July 15, 2020; published online on November 9, 2020. Discussion period a machine learning (ML) paradigm. To this end, many machine learn-
open until April 9, 2021; separate discussions must be submitted for indi- ing methods feature the use of supervised or nonsupervised classi-
vidual papers. This paper is part of the Journal of Computing in Civil En- fiers (Chen et al. 2011; Gavilán et al. 2011; Kaseko and Ritchie
gineering, © ASCE, ISSN 0887-3801. 1993; Liu et al. 2002; Prasanna et al. 2014; Zakeri et al. 2017).

© ASCE 04020057-1 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


In recent years, coincident with the advances in artificial intelli- (Siesler et al. 2008). In the context of image-based damage detec-
gence (AI) and particularly the development of deep learning tion, HSI provides a significant possibility of detecting and iden-
(DL) techniques, many have heralded the era of AI-enabled damage tifying the presence of material damage amid complex scenes on
inspection. A simple search through Google Scholar, using the surfaces of civil structures.
combined keywords of crack detection, convolutional neural net- The authors developed a mobile hyperspectral imaging system
work (CNN), and image returned more than 1,200 articles from for both ground and aerial vehicle–based remote sensing. With this
the period January 2016 to May 2020. Zhang et al. (2016) first used mobile HSI system and preliminary observation (e.g., by plotting
a CNN model as a feature extractor, then fed the features into a spectral profiles for different material surface objects), it is hypoth-
classification model for the detection of cracks in images. Such esized that material damage (e.g., cracks) and other complex arti-
a CNN-based machine learning approach then was adopted in facts on structural surfaces can be detected effectively with high
many other similar efforts (e.g., Alipour et al. 2019; Cha et al. 2017; performance. The ensuing research problem is whether the resulting
Ni et al. 2019). machine-learning model based on hyperspectral pixels can outper-
One may expect that by duly considering these advances in form the same model trained based on regular images with a high
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

AI-enabled image-based damage detection and the lowering cost spatial resolution (i.e., those based on panchromatic or true-color
of mobile or edge computing devices, autonomous damage inspec- imaging). The essential contribution of this study is the proven ef-
tion may become a reality. However, these machine-vision systems fectiveness of HSI for material damage detection with complex
have not been endorsed entirely by civil engineers for deploying scenes, and its superb capacity compared with high-resolution pan-
these techniques in practice. The authors argue that if the scene chromatic (gray) images. Specifically, unlike any existing image-
complexity captured in digital images is not acknowledged, the based damage detection method, the proposed machine-learning
pace of automation ultimately would be hindered. In the case of framework deals with the detection problem with many semanti-
concrete structures in service, the scenes in images often are a mix- cally rich structural-surface materials and objects, including con-
ture of characteristic material texture, possible damage, and other crete, asphalt, crack, dry vegetation, green vegetation, water, oil, and
artifacts, such as artificial marking, vegetation, moisture, oil spill, artificial markings. Such a plentiful amount of surface object cat-
discoloring, and uneven illumination (Chen and Hutchinson 2010). egories are not dealt with in the literature of image-based damage
This implies that any image-based machine learning method or detection, but are commonly found in civil structures in service. The
end-to-end deep learning method may encounter the infamous resulting inference model, although it falls in the paradigm of
issue of generalization. In other words, if such an autonomous traditional machine learning, can be implemented readily on an
image-based system is deployed in the field, its core detection onboard computer, hence providing the possibility of developing
component (i.e., a classification model), even one trained based a real-time mobile HSI system for both hyperspectral imaging and
on a relatively large data set with complex scenes, can overfit damage detection. The significant contribution also lies in the se-
the training data but cannot generalize to the multitude of different mantically labeled hyperspectral data set (Chen 2020), which pro-
and arbitrary scenes that are not in the prepared data sets for model vides an unprecedented basis for research in hyperspectral machine
training. vision and engineering inspection automation.
To resolve this challenge, one obvious solution is to continue This paper first briefly introduces the concept of HSI, and
developing much larger data sets, given the power of deep learning describes a mobile HSI system. Next, the machine learning meth-
with an architecture that potentially can accommodate any scale odology is introduced with a focus on proving the concept of
of data sizes and any complexity in field scenes. However, this HSI-based detection and its competitive performance. Performance
inevitably triggers the issue of labeling big data (e.g., pixelwise is evaluated and discussed in terms of four classification models,
labeling of cracks and other artifacts), which is expensive and and conclusions are presented.
time-consuming (Roh et al. 2019). Another approach is to resort
to transfer learning and use small data sets enhanced by effective
data augmentation techniques to obtain the notion of learning from Hyperspectral Imaging
small data using DL models. Tang and Chen (2017) developed a
novel data augmentation technique for accommodating deep trans-
Imaging Mechanism
fer learning from small data. Regardless of the potential success of
these solutions, with the use of red, green, and blue (RGB) images, The concept of HSI originated in the 1980s when Goetz and
these methods can only asymptotically match the intelligence of colleagues at the Jet Propulsion Laboratory (JPL) began develop-
trained inspectors (who provide labeled training data), although ing the seminal Airborne Visible/Infrared Imaging Spectrometer
possibly with much higher efficiency than human inspectors. In (AVIRIS) (Green et al. 1998; Plaza et al. 2009). The AVIRIS sensor
other words, there is a performance ceiling that limits the capacity measures the upwelling radiance with 224 contiguous spectral
of regular RGB images unless the machine intelligence supersedes bands at 10 nm intervals across the spectrum from 400 to 2,500 nm.
that of human beings. These radiance spectra are measured as push-broom scanning im-
An alternative solution is to break out the normal matching of ages with a swath width of 11 km and a flight length typically of
human vision. An emerging technology for nondestructive inspec- 10–100 km (both at a spatial resolution of 20 m=pixel). Similar to
tion is hyperspectral imaging (HSI). In a hyperspectral image, any common image, a hyperspectral image is formed by sampling
a pixel contains tens to thousands of digital values in different in a two-dimensional (2D) domain, generating a matrix of pixels.
spectral bands in the visible and near-infrared (VNIR) portion of Unlike gray-level or RGB images, a hyperspectral pixel consists of
the electromagnetic spectrum, at which each digital value represents a large number of intensity values sampled at various narrow spec-
either the reflectance property or the transmittance property of a tral bands that represent the contiguous spectral curve at the pixel.
material in one band. Such a high-dimensional spectral profile is A hyperspectral image is commonly called a hyperspectral cube,
not directly visible to human eyes, which, roughly speaking, re- because it is a three-dimensional (3D) matrix (two dimensions
spond to three discrete bands (RGB) (Kaiser and Boynton 1996). in spatial directions and one in the spectral direction). Using an
Scientific knowledge of hyperspectral imaging and analysis is well AVIRIS image and taking a scanning distance of 50 km, one
archived in the knowledge body of hyperspectral spectroscopy can create a hyperspectral cube with a size of 550 × 2,500 × 224.

© ASCE 04020057-2 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


With advances in HSI and especially sensors that are not for nonlinear, methods, such as the orthogonal subspace projection
orbital or airborne platforms, the acquisition of hyperspectral cubes method (Harsanyi and Chang 1994) and the locally linear embedding
can be realized with other mechanisms. In addition to the spatial method (Chen and Qian 2007), are also found in the literature. For
scanning or push-broom imaging mechanism, two other mecha- semantic segmentation–based object detection, feature description
nisms (spectral scanning and spatial-spectral scanning) were devel- needs to consider the spatial distribution of hyperspectral pixels.
oped for HSI applications in medical and biological sciences (Lu The most straightforward treatment is to extract feature vectors based
and Fei 2014). These three scanning techniques require complex on the dimensionality-reduced hyperspectral pixels, as in extracting
postprocessing steps to achieve the end product, a data cube. The features in a gray or a color image. Advanced methods in recent years
fourth mechanism is nonscanning, and is termed snapshot imaging have concerned the extraction of spatial–spectral features in three di-
(Johnson et al. 2004). Snapshot imaging acquires spectral pixels mensions (Fang et al. 2018; Hang et al. 2015; Zhang et al. 2014). For
in a 2D field of view simultaneously, without the requirement of the classification design, as the second stage of a conventional
trajectory flights or using any moving parts in the imager. There- machine-learning problem, the commonly used classical classifiers
fore, this snapshot mechanism also is referred to as real-time HSI include artificial neural network (ANN), support vector machine
(SVM), decision tree, k-nearest neighbor (KNN), and nonsupervised
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

(Bodkin et al. 2009), and can achieve much higher frame rates and
higher signal-to-noise ratios, and can provide hyperspectral cubes clustering techniques such as k-means (Goel et al. 2003; Ji et al. 2014;
immediately after every action of capturing, as in a regular digital Marpu et al. 2009; Mercier and Lennon 2003).
camera. Due to this property, real-time or snapshot HSI opens up As deep learning became a focus in the machine vision com-
significant opportunities for its use in portable or low-altitude re- munity, DL models have been developed in recent years which
mote sensing. potentially unify DR, feature description, and classification in
end-to-end architectures for hyperspectral image–based object de-
tection. Chen et al. (2014) integrated PCA, stacked autoencoder
Hyperspectral Image Computing (SAE), and logical regression into one DL architecture (Chen
A hyperspectral cube can be denoted hðx; y; sÞ, where ðx; yÞ are et al. 2014). Convolutional neural networks (CNNs), as commonly
the spatial coordinates (in terms of pixels) and s is the spectral adopted in regular machine-vision problems, were introduced in
coordinate in terms of spectral wavelength—i.e., for visible bands, subsequent years (Chen et al. 2016; Hu and Jia 2015; Li et al.
s ∈ ½400,600 nm, and for visible to near-infrared bands, s ∈ ½400; 2017). These advances imply the advent of computational hyper-
1,000 nm. At a select location of ðx; yÞ, therefore, hðx; y; sÞ rep- spectral imaging or hyperspectral machine vision, which may im-
resents a spectral profile when plotted against the spectral variable pact many engineering fields, including damage inspection in civil
s. In the remote-sensing context (not in a medical or biological engineering.
context), i.e., when the data cube is acquired in the air, the meas-
urement at the sensor is the upwelling radiance. In general, the re- Mobile Hyperspectral Imaging System
flectance property of a material at the ground nominally does not
A mobile HSI system for ground-level and low-altitude remote
vary with solar illumination or atmospheric disturbance. Therefore,
sensing was developed by the authors. The imaging system consists
the reflectance profile of a pixel reflects the characteristic or sig-
of a Cubert S185 FireflEYE snapshot camera (Ulm, Germany) and
nature of the material at that pixel. Therefore, a raw-radiance data
a mini-PC server for onboard computing and data communication
cube needs to be corrected to generate a reflectance cube, consid-
(Cubert 2018). For ground-based imaging, the system is mounted
ering the environmental lighting and the atmospheric distortion.
to a DJI gimbal (Shenzhen, China) that provides two 15-W,
This process is called atmospheric correction (Adler-Golden et al.
1580-mA·h batteries for powering both the imaging payload and
1998). For this, many physics-based algorithms exist, including the
the operation of the gimbal. Fig. 1(a) shows the gimbaled imaging
widely used fast line-of-sight atmospheric analysis of spectral
system ready for handheld or other ground-based HSI. To enable
hypercubes (FLAASH) (Cooley et al. 2002).
low-altitude remote sensing, an unmanned aerial vehicle is used,
The most frequent use of hðx; y; sÞ as a reflectance cube is
and the gimbaled system can be installed easily on the UAV for
material detection. Compared with single-band or three-band pixels
remote sensing [Fig. 1(b)].
in common images, the distinct property of hyperspectral images
The Cubert camera has a working wavelength range of 450–
is the high dimensionality of the pixels. This high-dimensionality
950 nm covering most of the VNIR wavelengths, spanning 139
is the foundation of spectroscopy for material identification. To
bands, with a spectral resolution of 8 nm in each band. The raw
achieve such identification, various unmixing methods have been
sensor has a snapshot resolution of 50 × 50 spectral channels.
developed to distinguish the statistically significant material types
Natively, the imager captures radiance images; with a proper cal-
and their abundance levels underlying the spectral profile of a pixel
ibration procedure and the internal processing in the onboard com-
(Bioucas-Dias et al. 2012; Heylen et al. 2014). If the concern is not
puter, the camera can output reflectance images directly. To do so, a
a specific material (that consists of single molecules or atoms), but
calibration process starts with the use of a standard white reference
an object is of interest, extracting and selecting descriptive features
board, achieving a data cube for the standard whiteboard (denoted
at hyperspectral pixels becomes the first stage. This stage, com-
hW ). Then a perfect dark cube is obtained (by covering the lens
monly called feature description, is as common as it is in a typical
tightly with a black cap), denoted hB . The relative reflectance im-
machine-learning process. If pixel-based object detection is con-
age, h, is calculated given a radiance cube hR
cerned, the dimensionality of hyperspectral pixels poses the first
challenge. To avoid the so-called curse of dimensionality when hR − hw
h¼ ð1Þ
using machine learning methods, early efforts emphasized the hw − hB
development of dimensionality reduction (DR) methods for select-
ing representative feature values across spectral bands. The most Following Eq. (1), a reflectance image can be produced by the
commonly used DR methods are mathematically linear, including camera system directly or can be postprocessed from the generated
principal component analysis (PCA) and linear discriminant analy- radiance image.
sis (LDA) (e.g., Bandos et al. 2009; Chen and Qian 2008; Hang One unique feature of the Cubert HSI system is its dual acquis-
et al. 2017; Palsson et al. 2015). Other advanced, and usually ition of hyperspectral cubes and a companion gray level–intensity

© ASCE 04020057-3 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

Fig. 1. Mobile hyperspectral imaging system: (a) gimbaled setup for ground-based imaging; and (b) ready-to-fly setup with a DJI UAV.
(Images by ZhiQiang Chen.)

image. The gray-level image has an identical field of view as the manually crafted feature description and classifier model selection,
hyperspectral cube but has a much higher spatial resolution, including unsupervised DR, material and object detection, and seg-
1,000 × 1,000. Denoting this gray image as gðu; vÞ, gðu; vÞ and mentation, are still considered as the trusted and robust approach.
hðx; y; sÞ can be fused to achieve a hyperspectral cube with a Second, the authors argue that if a DL architecture is used, the feature
higher resolution; at its peak, one can obtain a cube with a size of description and classification inevitably are coupled in an end-to-end
1,000 × 1,000 × 139. This process is called pansharpening; to architecture, and the native power of hyperspectral pixels in the spec-
obtain smooth sharpening effects, many algorithms exist (Loncan tral dimension and the gray-intensity pixels in the spatial dimensions
et al. 2015). However, pansharpening, which can provide visually would not be revealed straightforwardly. This study adopted the tra-
appealing hyperspectral images (if visualized in terms of pseudo- ditional machine learning methodology; specifically, four different
color images) as a very effective data fusion technique, does not feature description schemes were designed, and a standard classifi-
offer information other than the original low-resolution hyperspec- cation model was selected. This strategy properly facilitated proving
tral cube and the high-resolution gray image. the primary hypothesis and the ensuing research problem of justify-
ing the competitiveness of hyperspectral features. Third, the resulting
methodology in this paper can be implemented readily as a robust
Machine Learning Design filtering mechanism for real-time imaging and processing. The future
development of DL and field implementation is discussed
Rationale and Overview of Methodology subsequently.
A common machine-learning task, using either gray-level or RGB To realize the primary goal of proving the hypothesis of this
color images, is extracting visual features based on the spatial study, four machine learning models were designed. The main
distribution of pixel intensities. When the features are extracted, distinction between these four machine-learning models is their
they are fed into a classifier for supervised or unsupervised clas- feature description process. Fig. 2 presents the methodology com-
sification. This two-stage process, a traditional machine learning ponents for the four machine learning models. For the classifier,
paradigm, can be replaced with an end-to-end methodology by the seminal support vector machine (SVM) was adopted, and is
adopting the latest deep learning architecture, which possibly pro- reviewed subsequently.
duces higher detection performance than the traditional process. These machine models are
This was not adopted in this work for several reasons. 1. Model 1 (M1): hyperspectral pixels with spectral values are used
First, the high-capacity nature of a DL architecture lies in its directly as feature vectors; namely, hðx; y; sÞ at ðx; yÞ is used
capacity to represent the intrinsic structure of a complex object via directly as a feature vector, where s ∈ f1; 2; : : : ; 139g. This
its usually hundreds of networked layers of neural nets. To support a model is denoted M1(HYP), where HYP represents the feature
converging fitting for such a complex model, a very large number of extraction process.
labeled data points (or ground-truth data) is demanded for a regular 2. Model 2 (M2): linear PCA is applied as an additional feature
object detection task. In previous DL-based efforts for hyperspectral selection step to reduce the dimensionality. The profile of
data learning, such large-scale data sets have existed; for example, hðx; y; sÞ at ðx; yÞ is reduced from 139 to 6 dimensions and
the most used AVIRIS Indian Pines data (Baumgardner et al. 2015). becomes h 0 ðx; y; kÞ, where k ∈ f1; 2; 3; : : : ; 6g. This model is
However, all of them are Earth observation (EO) images (Graña denoted M2(HYP_PCA).
et al. 2020). No data set was collected and labeled for damage 3. Model 3 (M3): feature vectors are extracted based on the
detection for structural materials using ground-based HSI equip- companion gray-level images, gðu; vÞ. As a result, a feature
ment. Centering around these EO images, DL methodologies only vector at a 20 × 20 neighborhood in gðu; vÞ maps to the hyper-
recently have appeared for advanced spatial–spectral and semantic spectral pixel at ðx; yÞ. To extract the gray-level features within a
segmentation. The conventional pixel-level techniques featuring sliding 20 × 20 neighborhood in gðu; vÞ, a variant of the widely

© ASCE 04020057-4 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

Fig. 2. Methodology for feature description and different machine learning models.

used gradient-based feature extractor, the histogram of gradients • Objective 4 was to evaluate the performance of the classification
(HOG) is adopted. The resulting model is denoted model of M4(HYP_PCA, GL_HOG) to conclude whether,
M3(GL_HOG). through simple data fusion, the combined hyperspectral and gray-
4. Model 4 (M4): feature vectors are based on the combined use of level features provide more-competitive detection performance.
the feature vectors used in Model 2 and Model 3. By concat-
enating the two feature vectors, imagery information from both
the hyperspectral pixel–based spectrum and the gray value– Feature Selection and Extraction Methods
based spatial distribution is fused. This model is denoted The four models essentially differ in the use of the feature descrip-
M4(GL_HOG+HYP_PCA), and GL_HOG+HYP_PCA repre- tion processes based on the original hyperspectral data instance
sents the fourth feature extraction process in this paper. (a data cube and a companion gray-level image). Two basic feature
With these models, the specific objectives of proving the hy- description methods, the principal component analysis method as a
potheses were linear feature selection method and a variant of the histogram of
• Objective 1 was to evaluate the performance of the classification oriented gradients (HOG) method as a feature extraction method,
model M1(HYP) and to conclude whether a hyperspectral pixel are involved in the four processes.
is effective in recognizing the underlying object types, including
structural damage given a complex scene. Principal Component Analysis
• Objective 2 was to evaluate and compare the performance of A hyperspectral profile at ðx; yÞ, or denoted as a set {hðx; y; sÞjs ∈
M1(HYP) and M2(HYP_PCA) to conclude whether dimension- ½1; 2; : : : ; 139}, if treated as a feature vector, gives rise to a
ality reduction is effective in terms of improving the discrimi- 139 × 1 feature vector. As mentioned previously, such high-
nation of different objects. dimensionality readily leads to poor performance when training a
• Objective 3 was to evaluate and compare the performance of classification model (particularly when the training data are few,
M2(HYP_PCA) and M3(GL_HOG) to conclude whether hyper- and the model itself cannot accommodate the high-dimensional
spectral pixels are more effective than high-resolution gray-level space). Principal component analysis is a commonly used method
images in identifying complex object types. for dimensionality reduction, which in many cases can greatly

© ASCE 04020057-5 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


improve classification performance. In general, PCA is a statistical When the data do not have linear boundaries according to their
procedure that converts a set of data points with a high dimension underlying class memberships, the finite feature-vector space can
of possibly correlated variables into a set of new data points that be transformed into a functional space, which is termed the kernel
have linearly uncorrelated variables or are mutually orthogonal, trick. The Gaussian kernel is widely used and was adopted in this
called principal components (PCs). The component scores associ- study, which is parameterized by a kernel parameter (γ). Another
ated with the orthogonal PCs construct the coordinates in a new critical parameter, often denoted C, is a regularization parameter
space spanned by the PCs. By selecting several dominant PCs (less that controls the trade-off between achieving a low training error
than the original dimension of the data point), the original data and a low testing error or that controls the generalization capability
points can be projected to a lower-dimension subspace. Therefore, of the SVM classifier. Pertinent to a model selection problem, these
the select PC scores for an originally high-dimension feature point two hyperparameters can be determined through a cross-validation
can be used as a dimension-reduced feature vector. In this study, the process during the training. In the SVM literature, 10-fold cross-
first six components were selected empirically after testing from the validation often is used to determine the hyperparameters (Foody
first three to the first seven components. In addition, a normaliza- and Mathur 2006). To accommodate multiclass classification based
tion step was conducted for all PCA features at all pixels. The PC on a binary SVM classifier, two well-known designs can be adopted:
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

scores, when visualized, can be used as a straightforward approach (1) one-versus-all design; and (2) pairwise or one-versus-one design
to inspecting the discrimination potential of the features, which oth- (Duan et al. 2007). This study adopted the one-versus-one design.
erwise are not possible if the original high-dimension hyperspectral MATLAB version R2018a and its machine-learning toolbox were
profiles are plotted. These illustrations are described subsequently. employed to implement the multiclass SVM–based learning, includ-
ing training and testing.
Histogram of Oriented Gradients
The histogram of oriented gradient feature descriptor characterizes
the contextual texture or shape of objects in images by counting the Data Collection and Preparation
occurrence of gradient orientations in a select block in an image or
With the mobile HSI system [Fig. 1(a)], a total of 68 instances of
in the whole image. It was first proven effective by Dalal and Triggs
hyperspectral images (and their companion gray-level images)
(2005) in their seminal effort for pedestrian detection in images;
were captured in the field. Among these images, 43 images were
since then, HOG has been applied extensively for different objec-
of concrete surfaces and 25 images were of asphalt surfaces. To
tive detection tasks in the literature of machine vision. HOG differs
create scene complexity, artificial markings, oil, water, and green
from other scale-invariant or histogram-based descriptors in that its
and dry vegetation were added in 34 concrete images and 16 as-
extraction is computed over a dense grid of uniformly spaced cells,
phalt images that had hairline or apparent cracks. Of the remaining
and it uses overlapping local contrast normalization for improved
18 images, 9 images each were were of concrete and asphalt pave-
performance. There are many variants of HOG descriptors for im-
ments without cracks. A proper calibration process was included
proving the robustness and the accuracy; a commonly used variant
before the imaging; therefore, the obtained data all were reflectance
is HOG-UoCTTI (Felzenszwalb et al. 2010). The original HOG
cubes. Due to the fixed exposure setting of the camera, the gray
extractor and its variants are formulated over a gray image, which
images were captured physically as relatively dark pixels, and hence
is convenient to the research in this study, in which a companion
had low visual contrast [Figs. 3(a and c)]. In Figs. 3(a and c), no
gray-image exists for every hyperspectral cube.
image enhancement (e.g., histogram equalization) is introduced;
Considering the resolution compatibility between the hyper-
nonetheless, the contrast-normalization step in a HOG descriptor
spectral cube and the companion gray-level image, the feature ex-
effectively can mitigate this issue before producing the gray value–
traction was conducted in a 20 × 20 sliding neighborhood in a gray
based feature vectors.
image. Within this neighborhood, four histograms of undirected
To create a supervised learning–ready data set, manual and se-
gradients were averaged to obtain a no -dimensional histogram
mantic labeling was carried out. Semantic labeling is the process of
(i.e., binned per their orientation into nine bins, no ¼ 9) and a sim-
labeling each pixel in an image with a corresponding class label.
ilar operation was performed for the directed gradient to obtain a
This study used an image-segmentation (or image parsing) based
2no -dimensional histogram (i.e., binned according to gradient into
labeling approach in which clustered pixels belonging to the same
18 bins). Along with both directed and undirected gradient, the
class were delineated in the image domain and rendered with a
HOG-UoCTTI also computed another four-dimensional texture-
select color. The labeling is based on the gray-level image that
energy feature. The final descriptor was obtained by stacking the
accompanies a hyperspectral cube. In this work, this process was
averaged directed histogram, averaged undirected histogram, and
conducted using an open-source image processing program, GIMP
four normalized factors of the undirected histogram. This led to
version 2.10.12. During the labeling process, a total of six different
the final descriptor of size 4 þ 3 × no (i.e., a 31 × 1 feature vector).
classes, including cracking, green vegetation, dry vegetation, water,
oil, and artificial marking, were assigned the colors black, green,
Multiclass Support Vector Machine brown, blue, red, and yellow, respectively [Figs. 3(b and d)]. The
In the traditional machine vision literature, the combined use of background materials (concrete and asphalt) were not classified in
SVM as a classifier and HOG as a feature descriptor is regarded these complex images or in the plain (concrete/asphalt) images.
as a de facto approach for visual recognition tasks or a benchmark- To advance hyperspectral imaging–based machine learning and
ing model if used to validate a novel model (Bilal and Hanif 2020; damage detection research, the authors have shared this unique data
Bristow and Lucey 2014). This fact partially supports the selection set publicly (Chen 2020).
of SVM (in lieu of many other classical models) in this effort.
A native SVM is a binary classifier that separates data into classes
Feature Calculation and Visualization
by fitting an optimal hyperplane in the feature space separating all
training data points into two classes. As described by Foody and The colored mask images, denoted mðu; vÞ, share the same spatial
Mathur (2006), the most optimal hyperplane is the one that max- domain as the underlying gray image gðu; vÞ. For the sake of sim-
imizes the margin (i.e., from this hyperplane, the total distance to plicity, based on the color coding for the mask images, the value of
the nearest data points of the two classes is maximized). mð·Þ takes an integer value of 1, 2, : : : , 6 to indicate the underlying

© ASCE 04020057-6 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

Fig. 3. (a) Gray image of concrete surface; (b) mask image for (a); (c) gray image of asphalt surface; and (d) mask image for (c). Colored mask images
in (b) and (d) appear in the online version.

six surface objects; in addition, the following notations are used to where FEA represents one of the feature extraction processes,
describe the resulting data set: HYP, HYP_PCA, GL_HOG, or HYP_PCA+GL_HOG. By
skipping many background pixels or pixels that do not have
D ¼ fhn ðx; y; sÞ; gn ðu; vÞ; mn ðu; vÞjn ¼ 1; 2; : : : ; 50g ð2Þ dominant labels in Step 2b, the resulting number of meaningful
pixels (with dominant class labels) is 29,546, 8,132, 6,495,
6,273, and 5,312 features were obtained for the class labels (ck )
To generate the machine-learning data for the models, the fol-
of water, oil, artificial marking, and green vegetation, respec-
lowing process was developed:
tively. The number of features for cracks (concrete and asphalt
1. Iterating with the location of (xi ; yj ), where i; j ¼ 1; 2; : : : ; 50,
cracks) was 2,377. The dry vegetation features had the smallest
the spectral profile is stored in the vector {hðxi ; yj ; sÞjs ∈
number of features, 957.
½1; 139}, and the gray values in the corresponding gray
With the data cubes and gray images for the plain concrete
image gðu; vÞ are confined in a neighborhood block of b ¼
and asphalt surfaces (9 pairs each), 2,601 features arbitrarily were
fðu 0 ; v 0 Þju 0 ∈ ½ðxi − 1Þ × 20 þ 1; xi × 20; and v 0 ∈ ½ðyj − 1Þ ×
obtained without using mask images individually for the concrete
20 þ 1; yj × 20g. In this neighborhood of b, the gray-level im-
and the asphalt labels. After adding these features into Eq. 3, the
age patch and the corresponding mask patch corresponding to
number of labeled features used in this study [Eq. (3), K] was
the hyperspectral pixel at ðx; yÞ are denoted gðbÞ and mðbÞ,
34,748. As described previously, given the smallest (957) and the
respectively.
largest (8,132) number of features, a moderate imbalance existed.
2. Given the mask patch mðbÞ, a simplified process is used to se-
Based on the data in Eq. (3) and the original hyperspectral
lect the underlying class label for the hyperspectral pixel at
pixels, three plots in Fig. 4 qualitatively illustrate the potential dis-
(xi ; yj ). By counting the number of pixels belong to different
crimination power of hyperspectral pixels. Fig. 4(a) plots the origi-
object types within the neighborhood block b
nal spectral profiles of eight underlying objects (with only five
a. if a dominant class label exists, i.e., the number of pixels that
sample profiles for each object). Although the separability is evi-
belong to a class is greater than 50% of the total pixels in the
dent, it is objectively hard to discern their potential capability of
block (namely 200=400 pixels), this class label is assigned to
prediction. Figs. 4(b and c) plot the first three primary PC scores
(xi ; yj ); and
for obtained HYP_PCA features. These plotsshow the underlying
b. if no dominant class label exists, this pixel (xi ; yj ) and the
discrimination potential between all classes. HOG-based feature
corresponding neighborhood b is skipped.
vectors, which are in the form of histograms, also can be plotted;
3. At a pixel with a dominant class label, and per the feature ex-
however, it is challenging to discern their discrimination potential
traction method (HYP, HYP_PCA, and GL_HOG)
visually; this is possible only through a classification model.
a. if HYP is used, fhðxi ; yj ; sÞjs ∈ ½1; 139g is directly used as
the feature vector with a dimension of 139 × 1;
b. If HYP_PCA is used, PCA is conducted over the vector
fhðxi ; yj ; sÞjs ∈ ½1; 139g, and the first 6 PC scores are used Experimental Tests and Performance Evaluation
to form a much lower-dimensional 6 × 1 feature vector;
c. if GL_HOG is used, the feature extraction is based on the Data Partition and Computational Cost
gray-level patch gðbÞ using the HOG-UoCTTI method, re- To proceed with the modeling and performance evaluation, a data
sulting in a 31 × 1 feature vector; and partition strategy is needed to split the data; one part is for the
d. if HYP_PCA + GL_HOG is fused, the two corresponding training, and the other part is for model validation. In the literature,
feature vectors simply are concatenated, resulting in a 37 × 1 the widely used scheme is to use 75% of the total data for training
feature vector. and the remaining 25% for testing. Indeed, if there is a sufficient
4. By iterating this procedure over all the hyperspectral pixels for amount of data, the data splitting ratio is flexible and is up to
all 50 instances of images, which include different types of the analyst. In this study, it was meaningful to examine whether
features in consideration, the following classification data set the amount of data size was sufficient, which can be determined
is obtained for each of the feature extraction methods: if the prediction performance increases with the amount of train-
ing data. Three data splitting schemes were considered for the
DFEA ¼ fðpk ; ck Þjk ¼ 1; 2; : : : ; Kg ð3Þ obtained feature data sets [Eq. (3)]. The first scheme, Test 1, after

© ASCE 04020057-7 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


12
Asphalt 5 Asphalt
Color Marking Concrete
Concrete 4 Crack
10 Crack Water

Second Principal Component


Dry Vegetation 3 Color Marking
Green Vegetation Dry Vegetation
8
Reflectance(103 )

Oil Green Vegetation


2
Water Oil

6 1

0
4
-1

2 -2
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

-3
0
0 20 40 60 80 100 120 140 -4 -2 0 2 4 6
(a) Spectral band (b) First Principal Component

Asphalt
2 Concrete
Crack
Water
1.5 Color Marking
Dry Vegetation
Third Principal Component

Green Vegetation
1 Oil

0.5

-0.5

-1

-5 -4 -3 -2 -1 0 1 2 3 4 5
(c) First Principal Component

Fig. 4. (a) Original spectral profiles as features for different classes; (b) PC scores as features from the first and second PCs; and (c) PC scores as
features from the first and third PCs.

randomly shuffling the data, considered 25% (8,147) of the data subsequently, is that M3 was the least accurate. In contrast, model
for training, and the remaining 75% (26,601) for testing purposes. M2 with the feature type HYP_PCA required the least training time
In Test 2, the total data set was divided equally (i.e., 17,374 (while achieving the best accuracy).
for training and 17,374 for testing). In Test 3, 75% (26,601) of the The average inference times of the resulting models were in-
data set was considered for training, and the remaining 25% spected, because the models potentially would be implemented
(8,147) for testing. With these three schemes, 12 different models as a predictive model. Fig. 5(b) reports the average inference times
were evaluated. of all models and test scenarios. As expected, the reported infer-
Training time and testing are important factors for assessing ence times did not vary much among the tests. The complexity did
the classification models as well as for the selection of the best not scale with the dimensionality of the feature vectors. Models
model for inference when implemented. The training times and the M2 (in which the feature HYP_PCA had a dimension of six)
inference times for prediction are presented in Fig. 5. In Fig. 5(a), and M4 (in which the joint feature had a dimension of 37) had
two trends are apparent. First, the model training scaled up with greater computational cost than the others (M1 and M3), although
the number of training data from Test 1 to Test 3. Second, the type M1’s feature vectors had a dimension of 139. As studied in the
of features significantly affects the training time; nonetheless, it literature, the computational complexity of a SVM model relies
does not scale with the dimensionality. As evident from Fig. 5(a), on the dimensionality of the input features and their realized struc-
M3 with the feature type GL_HOG required more training time tures of kernel functions, whichever dominates (Chapelle 2007).
than the other models in all tests, whereas the differences in train- Among the achieved models (M1–M4), it is arguable that the
ing of M1, M3, and M4 were less significant. The authors assert kernel structure determines the computational cost. Regardless,
that this was due to the underlying nonlinearity of the decision as elaborated subsequently, the model select for its accuracy was
boundaries defined in the feature space; hence the learning rate M2(HYP_PCA), which had an average inference time of 80 ms in
differed significantly. The corroborating evidence, as elaborated the MATLAB’s running environment.

© ASCE 04020057-8 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


(a) (b)

Fig. 5. Time cost for (a) training models; and (b) making an inference.
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

Performance Measures as the primary measure of classification capacity. The AU-PR also is
threshold-invariant; nonetheless, it may be statistically redundant if
The performance of the classification models was evaluated in
terms of quantitative measures. With a trained model iterating over the F1 score is used as well.
the testing data points, a confusion matrix can be generated; then, In summary, the performance evaluation in this paper used
several basic accuracy measures can be calculated, including the the following performance measures: precision, recall, F1 score,
overall accuracy, precision, and recall. However, it has been estab- AU-ROC, and AU-PR. Among them, the precision and recall were
lished in the literature that the overall accuracy is not robust to im- used to evaluate the consistency of the two measures, the F1
balanced data; in such a situation, misinterpretation may occur even score was used as the primary accuracy measure, the AU-ROC
given a high accuracy measurement. Both precision and recall mea- was treated as an important measure of model capacity, and the
sures emphasize the outcomes of false predictions; therefore, in any AU-PR was treated as an auxiliary (possibly redundant) measure
model, the precision competes against the recall, or vice versa. The to the F1 score and to the AU-ROC. Selectively, the ROC and
F1 score is defined as the harmonic mean of the precision and recall PR curves were adopted to observe and evaluate comprehensively
values. Therefore, the F1 score tends to better gauge the classifi- both the capacity and the stability of the underlying classification
cation accuracy with an imbalanced data set, because only when models.
both precision and recall are high does the F1 score become high
(Goutte and Gaussier 2005).
Test Results and Description
To further evaluate the capacity and stability of the classifier, two
commonly used graphical analytics are found: receiver operating Test results from the four models (M1–M4) using three data split-
characteristics (ROCs) curves and precision-recall (PR) curves. De- ting schemes (Tests 1–3) are reported herein. Tables 1–4 report
tails of ROC and PR curves were given by Hossin and Sulaiman the performance results based on the testing data per the split-
(2015). In addition to the curves, two derived statistics as lumped ting schemes. Table 1 reports the results for M1(HYP), which
measures are used widely to indicate the classification capacity. tested the hyperspectral feature vector type HYP. Tables 2–4
These are the area under the ROC curve (AU-ROC), and the area report results for models M2(HYP_PCA), M3(GL_HOG), and
under the precision-recall curve (AU-PR) for both the AU-ROC and M4(HYP_PCA+GL_HOG), respectively.
the AU-PR, a greater value (upper bound of 1) indicates a larger Some of the general trends in Tables 1–4 are summarized as
capacity. One important consensus in the literature is that ROC follows. In most of the testing, the prediction performance in-
curves lack indicative power when the training data are imbalanced. creased as more data were used in training from Test 1 to Test 2
Nonetheless, because the AU-ROC is threshold-invariant, it is used to Test 3 (Tables 1–4). In a few cases, the testing performance

Table 1. Performance measurements of M1(HYP)


Test Measure Concrete Asphalt Color marking Crack Dry vegetation Green vegetation Water Oil
Test 1 Precision 0.998 1.000 0.598 0.414 0.488 0.776 0.743 0.612
Recall 0.999 0.998 0.531 0.130 0.164 0.705 0.946 0.717
F1 0.999 0.999 0.563 0.198 0.246 0.739 0.832 0.661
AU-ROC 1.000 1.000 0.889 0.783 0.915 0.947 0.965 0.909
AU-PR 1.000 1.000 0.551 0.237 0.265 0.811 0.830 0.638
Test 2 Precision 1.000 1.000 0.611 0.567 0.558 0.818 0.787 0.628
Recall 0.999 1.000 0.655 0.167 0.192 0.704 0.957 0.702
F1 1.000 1.000 0.632 0.257 0.286 0.756 0.864 0.663
AU-ROC 1.000 1.000 0.906 0.808 0.921 0.952 0.971 0.912
AU-PR 0.999 0.999 0.592 0.312 0.290 0.800 0.845 0.642
Test 3 Precision 1.000 1.000 0.501 0.918 0.402 0.706 0.572 0.553
Recall 1.000 1.000 0.629 0.057 0.213 0.733 0.962 0.724
F1 1.000 1.000 0.558 0.107 0.278 0.719 0.717 0.627
AU-ROC 1.000 1.000 0.869 0.774 0.906 0.940 0.963 0.890
AU-PR 0.999 0.999 0.490 0.495 0.234 0.745 0.685 0.586

© ASCE 04020057-9 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


Table 2. Performance measurements of M2(HYP_PCA)
Test Measure Concrete Asphalt Color marking Crack Dry vegetation Green vegetation Water Oil
Test 1 Precision 0.930 0.901 0.957 0.914 0.878 0.986 0.973 0.955
Recall 0.886 0.910 0.964 0.921 0.664 0.979 0.992 0.974
F1 0.907 0.906 0.961 0.917 0.757 0.982 0.983 0.964
AU-ROC 0.993 0.995 0.997 0.995 0.985 0.999 0.999 0.998
AU-PRC 0.957 0.960 0.977 0.960 0.839 0.993 0.985 0.982
Test 2 Precision 0.970 0.936 0.973 0.973 0.900 0.990 0.994 0.979
Recall 0.959 0.945 0.989 0.952 0.808 0.994 0.995 0.981
F1 0.964 0.941 0.981 0.962 0.851 0.992 0.994 0.980
AU-ROC 0.998 0.998 0.999 0.997 0.993 1.000 1.000 0.998
AU-PRC 0.987 0.983 0.985 0.987 0.893 0.995 0.997 0.985
Test 3 Precision 0.987 0.942 0.984 0.971 0.908 0.989 0.988 0.992
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

Recall 0.955 0.954 0.991 0.950 0.867 0.997 0.996 0.995


F1 0.971 0.948 0.988 0.960 0.887 0.993 0.992 0.993
AU-ROC 0.998 0.997 0.999 0.997 0.994 1.000 1.000 1.000
AU-PRC 0.981 0.981 0.993 0.985 0.886 0.996 0.993 0.994

Table 3. Performance measurements of M3(GL_HOG)


Test Measure Concrete Asphalt Color marking Crack Dry vegetation Green vegetation Water Oil
Test 1 Precision 0.560 0.580 0.723 0.575 0.724 0.740 0.888 0.725
Recall 0.559 0.598 0.781 0.195 0.234 0.731 0.943 0.845
F1 0.560 0.589 0.751 0.292 0.354 0.736 0.915 0.780
AU-ROC 0.940 0.950 0.926 0.774 0.898 0.929 0.985 0.944
AU-PR 0.549 0.615 0.724 0.295 0.365 0.750 0.932 0.764
Test 2 Precision 0.638 0.690 0.752 0.676 0.827 0.786 0.929 0.785
Recall 0.669 0.637 0.827 0.273 0.290 0.801 0.964 0.903
F1 0.653 0.662 0.788 0.389 0.430 0.793 0.946 0.840
AU-ROC 0.963 0.966 0.948 0.833 0.931 0.958 0.991 0.962
AU-PR 0.647 0.701 0.756 0.399 0.466 0.796 0.954 0.831
Test 3 Precision 0.665 0.690 0.772 0.644 0.886 0.801 0.882 0.802
Recall 0.639 0.679 0.847 0.301 0.325 0.809 0.964 0.906
F1 0.652 0.684 0.808 0.410 0.476 0.805 0.921 0.851
AU-ROC 0.959 0.969 0.942 0.828 0.948 0.962 0.990 0.964
AU-PR 0.666 0.737 0.739 0.424 0.529 0.816 0.921 0.850

Table 4. Performance measurements of M4(HYP_PCA+GL_HOG)


Test Measure Concrete Asphalt Color marking Crack Dry vegetation Green vegetation Water Oil
Test 1 Precision 0.520 0.842 0.904 0.785 0.489 0.905 0.956 0.890
Recall 0.965 0.881 0.921 0.848 0.341 0.927 0.679 0.873
F1 0.676 0.861 0.913 0.815 0.402 0.916 0.794 0.881
AU-ROC 0.978 0.992 0.993 0.985 0.938 0.993 0.974 0.986
AU-PR 0.695 0.933 0.963 0.898 0.391 0.964 0.930 0.930
Test 2 Precision 0.450 0.829 0.922 0.841 0.670 0.875 0.930 0.903
Recall 0.963 0.888 0.931 0.876 0.436 0.910 0.567 0.902
F1 0.613 0.857 0.927 0.858 0.528 0.892 0.704 0.902
AU-ROC 0.969 0.993 0.994 0.990 0.957 0.990 0.961 0.988
AU-PR 0.592 0.928 0.966 0.931 0.583 0.947 0.888 0.938
Test 3 Precision 0.779 0.761 0.931 0.895 0.779 0.947 0.969 0.925
Recall 0.939 0.962 0.935 0.835 0.529 0.951 0.851 0.894
F1 0.852 0.849 0.933 0.864 0.630 0.949 0.906 0.909
AU-ROC 0.995 0.995 0.994 0.989 0.962 0.998 0.997 0.989
AU-PR 0.946 0.957 0.970 0.924 0.683 0.985 0.981 0.944

decreased in Test 3 (which used 25% of the data points for testing); positive effect on improving the accuracy of classification models.
it is speculated that random shuffling occasionally may bring in The prediction of cracks and dry vegetation was the most challeng-
outlier data that are not learned from the training data. This obser- ing task. This was expected, because cracks are treated as an object
vation indicates that the size of the training data does have a type that comes from two structural scenes (concrete and asphalt

© ASCE 04020057-10 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


pavements), the underlying materials are very diversified, and over- The AU-ROC measure was adopted as a model-capacity measure
all all the reflectance intensities are darker than other objects. The because it was invariant to classification thresholds and nonredun-
dry vegetation type had the least number of instances in the data set, dant to the F1 score.
which partially contributed to the relatively poor performance com-
pared with the prediction for the green vegetation. In addition, in all
tests of all models, the precision and recall measurements were rel- Performance Evaluation
atively close for all class labels except cracks and dry vegetation. The F1 score and AU-ROC were used as two primary performance
This was particularly significant for M1(HYP) and M3(GL_HOG), measures in this study. To meet the objectives outlined previously,
for which the recall values tended to be much smaller than the pre- Figs. 7(a–d) illustrate the performance of all four models. Fig. 7(a)
cision values. This means that in the cases of these two objects, each shows the F1 scores for Test 1 data; Fig. 7(b) shows the F1 scores
instance of prediction tended to be correct; however, the models for Test 3 data; Fig. 7(c) shows the AU-ROC for Test 1 data; and
tended to miss many objects that were either cracks or dry vegetation Fig. 7(d) shows the AU-ROC for Test 3 data. The results from Test
objects. Furthermore, this result implies that a trade-off measure, 2 are not included herein because the observed trends from Test 2
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

such as the F1 score, should be used as a better accuracy measure. results are between Test 1 and Test 2. The following evaluation
To determine the coherency and the distinction of these differ- focuses on the four objectives of this study described previously.
ent measurements, scatter plots of all measures are illustrated in The F1 scores of Model M1(HYP) [Figs. 7(a and b)] indicate
Figs. 6(a–d) for models M1–M4, respectively; for the sake of brev- that the model successfully identifies (F1 > 0.7) the plain concrete,
ity, only the Test 2 data scheme is included. Figs. 6(a–d) plot the F1 asphalt, green vegetation, and water. For the ed marking and oil, the
scores against themselves (as indicated by a 1:1 line), and all other F1 score was about 0.6. However, for cracks and dry vegetation, the
measurements are plotted against F1. The AU-PR as a model F1 measurements were less than 0.3, indicating little prediction ac-
capacity measure was statistically consistent with the F1 score, be- curacy. First, this reflects the challenge in recognizing cracks, pri-
cause it followed a linear trend in each model with a high R2 value marily caused by the inherent spectral complexity. Second, it is
(R2 > 0.87). This implies that AU-PR largely was redundant as a primarily due to the smaller number of dry-vegetation data points.
capacity measure. Therefore, the F1 score can replace it as both For both objects, this presumptively is attributed to the high dimen-
a prediction accuracy measure and a model capacity measure. sionality of the feature vectors of HYP. Nonetheless, as expected

1
0.9

0.95 R2 = 0.9181
0.7
Measure value

Measure value

R2 = 0.9882 0.9
0.5

0.85
0.3

0.1 0.8
0.1 0.3 0.5 0.7 0.9 0.8 0.85 0.9 0.95 1
(a) F1 (b) F1

Preccision Recall F1 Preccision Recall F1


AU-ROC AU-PR Linear (F1) AU-ROC AU-PR Linear (F1)
Linear (AU-PR) Linear (AU-PR)

1 1

0.9
0.9
0.8
0.8
Measure value

Measure value

0.7
R2 = 0.8724
0.6 R2 = 0.9877 0.7

0.5
0.6
0.4
0.5
0.3
0.2 0.4
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1
(c) F1 (d) F1
Preccision Recall F1 Preccision Recall F1
AU-ROC AU-PR Linear (F1) AU-ROC AU-PR Linear (F1)
Linear (AU-PR) Linear (AU-PR)

Fig. 6. Performance measurement scatter plots using Test 2 data: (a) M1; (b) M2; (c) M3; and (d) M4.

© ASCE 04020057-11 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)

(c) (d)

Fig. 7. Model performance: (a) F1 score with Test 1 data; (b) F1 score with Test 3 data; (c) AU-ROC with Test 1 data; and (d) AU-ROC with
Test 3 data.

from the AU-ROC measurements [Figs. 7(c and d)], there is no of M2(HYP_PCA) and M3(GL_HOG) were evaluated and com-
doubt that Model M1(HYP) is highly effective in recognizing these pared. The F1 scores and the AU-ROC measurements shown in
structural surface objects. For concrete, the average AU-ROC was Figs. 7(a–d), respectively, clearly indicate that in the prediction of
about 1. Therefore, hyperspectral pixels as feature vectors are ef- all class labels, M2(HYP_PCA) superseded M3(GL_HOG). Even
fective in recognizing most of the structural surface objects, and considering the fourth model, M4, to be evaluated, M2(HYP_PCA)
have relatively less accuracy only in the detection of cracks and was the best model in Test 3. This model provided outstanding
dry vegetation. performance; the smallest F1 score was 0.757, and the smallest
Comparing the accuracy of Model M2(HYP_PCA) with that of AU-ROC measurement was 0.985, both for predicting dry vegeta-
M1(HYP) [Figs. 7(a and b)] showed that the detection accuracy sig- tion; for crack prediction, F1 ¼ 0.917 and AU-ROC ¼ 0.995. The
nificantly increased. For the two weak prediction instances of cracks control model in this test, M3(GL_HOG), had F1 ¼ 0.354 and
and dry vegetation in M1(HYP), the F1 score increased from 0.198 AU-ROC ¼ 0.898 for dry vegetation, and AU-ROC ¼ 0.774, and
to 0.917 in Test 1 and from 0.107 to 0.96 in Test 3 with Model M2 F1 ¼ 0.292 for cracks.
(HYP_PCA). For dry vegetation, the F1 scores changed from 0.246 To further visualize the prediction capacity and the stability of
to 0.757 in Test 1, and from 0.278 to 0.887 in Test 3. For all other the two most important models in this work, M2(HYP_PCA) and
class labels, the classification accuracy still increased from their ini- M3(GL_HOG), the ROC curves are illustrated in Figs. 8(a and b)
tially high F1 values from M1 to M2. When inspecting Fig. 7(c), the for the two models, respectively, based on Test 3 data. Similarly,
AU-ROC measurements from M1 and M2 increased. Particularly, Figs. 9(a and b) report the PR curves. The two ROC and PR curves
the M2’s AU-ROC measurements were greater than 0.99 at predict- show the different prediction capacity for each class label as the
ing all classes. This again signifies that Model M2(HYP_PCA) had underlying classification threshold changed. The ROC and PR
nearly perfect capacity to detect all structural surface objects. The curves show that Model M2(HYP_PCA) had not only superb clas-
comparative tests herein provide the direct evidence that performing sification capability but also stronger stability, the latter of which is
dimensionality reduction over hyperspectral profiles (i.e., as HYP indicated by the smoothness of the curves as the underlying thresh-
feature vectors) substantially can unleash the embedded discrimina- old varies. In the case of M3(GL_HOG), the classification capabil-
tion capacity of the data that otherwise is not exploitable. ity (and accuracy) overall was much moderate; the stability also
A secondary yet important research problem in this work was was worse.
to prove the competitiveness of low spatial-resolution hyperspec- For Objective 4, the authors evaluated a simple data fusion
tral data compared with high-resolution gray-intensity images in technique, which joined the two feature types, HYP_PCA and
detecting material surface objects. For this purpose, the results GL_HOG, and observed the possible performance gain or loss.

© ASCE 04020057-12 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


1 1

0.9 0.9

0.8 1 0.8

0.7 0.98 0.7


True positive rate

True positive rate


0.6 0.96 0.6
0.94
0.5 0.5
0.92
0.4 Concrete 0.4 Concrete
0.9 Asphalt Asphalt
0.3 Color Marking 0.3 Color Marking
0 0.005 0.01 0.015 0.02 0.025 0.03 Crack Crack
0.2 Green Vegetation 0.2 Green Vegetation
Water Water
0.1 0.1
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

Oil Oil
Dry Vegetation Dry Vegetation
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) False positive rate (b) False positive rate

Fig. 8. ROC curves for (a) M2 (HYP_PCA); and (b) M3 (GL_HOG) using Test 3 data.

1 1
Concrete
Asphalt 0.9
Color Marking
0.8 Crack 0.8
Green Vegetation
0.7
Water
0.6 Oil 1 0.6
Precision

Precision

Dry Vegetation
0.95 0.5

0.4 0.4 Concrete


0.9 Asphalt
0.3 Color Marking
0.85 Crack
0.2 0.2 Green Vegetation
0.8 Water
0.1 Oil
0.9 0.95 1
Dry Vegetation
0 0
0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) Recall (b) Recall

Fig. 9. Precision-recall curves for (a) M2 (HYP_PCA); and (b) M3 (GL_HOG) using Test 3 data.

The comparative evaluation was conducted between M4(HYP_ features corresponding to the hyperspectral pixel, even though the
PCA+GL_HOG) and M2, and between M4 and M3. In terms of gray images were captured at a much higher resolution (20 times
both F1 and AU-ROC measurements, the performance of M4 was higher). Although improved gray value–based feature extraction
worse than that of M2. This implies that the gray value–based techniques can be used, hyperspectral pixels (which are captured
features (GL_HOG) compromised the classification potential of in the visible to near-infrared spectral bands with high dimension-
HYP_PCA. When M4 and M3 were assessed, for the object types ality) have undoubtful promise in the detection of complex struc-
of concrete, asphalt, crack, and green vegetation, performance tural surface objects.
increased; however, for other objects (color marking, green vegeta- Furthermore, the methodology in this study was a traditional
tion, water, and oil), the performance decreased. machine learning paradigm. As reasoned previously, this method-
ology was proper considering the hypothesis-testing nature of this
effort. In addition, the hyperspectral data set presented in this paper
Discussion was unique, and the models developed herein form a baseline for
future studies (e.g., semantic object segmentation). Moreover, the
The observed superior performance of M2(HYP_PCA) compared selected Model M2(HYP_PCA) developed herein, which is much
with that of M3(GL_HOG) implies that a single hyperspectral pixel less complex than the latest DL models yet has very high accuracy
is much more effective than a 20 × 20 neighborhood of gray values at pixel-level prediction, can be implemented readily in an edge-
in discriminating the complex structural surface objects. Hyper- computing device (e.g., the onboard computer that was natively
spectral pixels with reflectance features at both visible and infrared designed for the Cubert camera in our imaging system). The re-
bands, once preprocessed (e.g., a PCA-based feature selection sulting UAV-based imaging system, therefore, can realize high-
step), outperformed the gray values or the embedded texture/shape throughput hyperspectral imaging and onboard processing, in which

© ASCE 04020057-13 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


the ML model developed in this study can serve as a rapid filtering Data Availability Statement
tool so that any images with cracks (or other anomalies) can be
detected quickly. Given the resulting data filtered by the model pro- Some or all data, models, or code generated or used during the
posed in this paper, a DL model, i.e., a semantic spatial-spectral study are available from the corresponding author by request.
segmentation model, can be deployed to realize more detailed The training and testing data set is archived publicly (Chen 2020).
and semantics-rich damage detection.
One specific future direction is to extract spectral–spatial fea-
tures using a more integral approach. As mentioned previously, a Acknowledgments
pansharpening technique can be used to generate a high spatial-
resolution hyperspectral cube. Specifically, the hyperspectral cube This material is based partially upon work supported by the National
of 20 × 20 × 139 used in this study can be turned into a cube Science Foundation (NSF) under Award number No. IIA-1355406
with much higher spatial dimension, namely a 1,000 × 1,000 × and work supported by the United States Department of Agricul-
139 cube. With this 3D cube, and a mask image of the same size, ture’s National Institute of Food and Agriculture (USDA-NIFA)
under Award No. 2015-68007-23214. Any opinions, findings, and
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

advanced spatial-spectral features may be extracted (Fang et al.


conclusions or recommendations expressed in this material are those
2018; Hang et al. 2015; Zhang et al. 2014), and then a classifier
of the author(s) and do not necessarily reflect the views of NSF or
(e.g., a kernel SVM as used in this study) can be adopted. This
USDA-NIFA.
pansharpening may act as an effective data augmentation tech-
nique, such that it can support a deep learning architecture. This
promising direction is beyond the scope of this work.
References
AASHTO. 2019. Manual for bridge element inspection. Washington, DC:
Conclusion AASHTO.
Adler-Golden, S., A. Berk, L. Bernstein, S. Richtsmeier, P. Acharya, M.
A mobile hyperspectral imaging system was developed for ground- Matthew, G. Anderson, C. Allred, L. Jeong, and J. Chetwynd. 1998.
based and aerial data collection. One of its primary applications “FLAASH, a MODTRAN4 atmospheric correction package for hyper-
is to inspect surfaces of structural materials, such as concrete and spectral data retrievals and simulations.” In Proc., 7th JPL Airborne
asphalt materials, which commonly are used in transportation struc- Earth Science Workshop, 9–14. Pasadena, CA: California Institute
tures. The innovation lies in its ability to detect surface damage of Technology.
and other artifacts at the material levels due to its high-dimensional Alipour, M., D. K. Harris, and G. R. Miller. 2019. “Robust pixel-level crack
detection using deep fully convolutional neural networks.” J. Comput.
pixels with reflectance in both visible and near-infrared bands. The
Civ. Eng. 33 (6): 04019040. https://doi.org/10.1061/(ASCE)CP.1943
effectiveness of the technique was proven in composition with -5487.0000854.
regular gray-level images that are very high-resolution and com- Bandos, T. V., L. Bruzzone, and G. Camps-Valls. 2009. “Classification of
monly used in practice. Four machine-learning models character- hyperspectral images with regularized linear discriminant analysis.”
ized by different feature description processes were trained and IEEE Trans. Geosci. Remote Sens. 47 (3): 862–873. https://doi.org/10
tested. With a total of 34,748 labeled features of different types, .1109/TGRS.2008.2005729.
three data-splitting schemes were used to evaluate the effects of Baumgardner, M. F., L. L. Biehl, and D. A. Landgrebe. 2015. 220 band
data size. A multiclass support vector machine with a Gaussian AVIRIS hyperspectral image data set: June 12, 1992 Indian Pine Test
kernel was adopted in all models. State-of-the-art measures were Site 3. West Lafayette, IN: Purdue Univ. Research Repository. https://
adopted to test the models, and the issue of data imbalance was doi.org/10.4231/R7RX991C.
considered. From a comprehensive evaluation, two major conclu- Bilal, M., and M. S. Hanif. 2020. “Benchmark revision for HOG-SVM
sions were formulated: pedestrian detector through reinvigorated training and evaluation meth-
odologies.” IEEE Trans. Intell. Transp. Syst. 21 (3): 1277–1287. https://
1. The use of hyperspectral reflectance values at pixels as feature
doi.org/10.1109/TITS.2019.2906132.
vectors was effective in recognizing all eight surface objects.
Bioucas-Dias, J. M., A. Plaza, N. Dobigeon, M. Parente, Q. Du, P. Gader,
However, dimensionality reduction is essential for this effective- and J. Chanussot. 2012. “Hyperspectral unmixing overview: Geomet-
ness. The linear PCA approach was adopted; using only the first rical, statistical, and sparse regression-based approaches.” IEEE J. Sel.
six component scores, the resulting classification models were Top. Appl. Earth Obs. Remote Sens. 5 (2): 354–379. https://doi.org/10
highly accurate, high-capacity, and stable. .1109/JSTARS.2012.2194696.
2. Compared with the model based on gray-level features (GL_ Bodkin, A., A. Sheinis, A. Norton, J. Daly, S. Beaven, and J. Weinheimer.
HOG), based on a popular variant of the HOG descriptor, the 2009. “Snapshot hyperspectral imaging: The hyperpixel array camera.”
PCA-adapted hyperspectral features had much better detection In Proc., Algorithms and Technologies for Multispectral, Hyperspec-
performance. A single hyperspectral pixel with high-dimension tral, and Ultraspectral Imagery XV, 73340H. Bellingham, WA:
spectral reflectance evidentially is competitive compared with International Society for Optics and Photonics.
the corresponding high-resolution gray intensities that express Bristow, H., and S. Lucey. 2014. Why do linear SVMs trained on HOG
the shape and texture of the underlying objects. The data fusion features perform so well? Ithaca, NY: Cornell Univ.
technique in this study had a less desirable performance, which Cawley, P. 2018. “Structural health monitoring: Closing the gap between
research and industrial deployment.” Struct. Health Monit. 17 (5):
may be due to the simple concatenation technique used to
1225–1244. https://doi.org/10.1177/1475921717750047.
combine the GL_HOG and HYP_PCA features.
Cha, Y.-J., W. Choi, and O. Büyüköztürk. 2017. “Deep learning-based crack
With the experimental and machine learning–based evaluation damage detection using convolutional neural networks.” Comput.-Aided
results in this paper, the authors further envision the dawn of com- Civ. Infrastruct. Eng. 32 (5): 361–378. https://doi.org/10.1111/mice
putational hyperspectral imaging or hyperspectral machine vision .12263.
for material damage detection in civil engineering and their promise Chapelle, O. 2007. “Training a support vector machine in the primal.”
in dealing with complex scenes that are often found in civil struc- Neural Comput. 19 (5): 1155–1178. https://doi.org/10.1162/neco.2007
tures in service. .19.5.1155.

© ASCE 04020057-14 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


Chen, G., and S.-E. Qian. 2007. “Dimensionality reduction of hyperspectral European Conf. on Information Retrieval, 345–359. New York:
imagery.” J. Appl. Remote Sens. 1 (1): 013509. https://doi.org/10.1117 Springer.
/1.2723663. Graña, M., M. Veganzons, and B. Ayerdi. 2020. “Hyperspectral remote
Chen, G., and S.-E. Qian. 2008. “Simultaneous dimensionality reduction sensing scenes.” Accessed May 1, 2020. http://www.ehu.eus/ccwintco
and denoising of hyperspectral imagery using bivariate wavelet shrink- /index.php/Hyperspectral_Remote_Sensing_Scenes.
ing and principal component analysis.” Can. J. Remote Sens. 34 (5): Green, R. O., et al. 1998. “Imaging spectroscopy and the airborne visible/
447–454. https://doi.org/10.5589/m08-058. infrared imaging spectrometer (AVIRIS).” Remote Sens. Environ.
Chen, Y., H. Jiang, C. Li, X. Jia, and P. Ghamisi. 2016. “Deep feature ex- 65 (3): 227–248. https://doi.org/10.1016/S0034-4257(98)00064-9.
traction and classification of hyperspectral images based on convolu- Hang, R., Q. Liu, H. Song, and Y. Sun. 2015. “Matrix-based discriminant
tional neural networks.” IEEE Trans. Geosci. Remote Sens. 54 (10): subspace ensemble for hyperspectral image spatial–spectral feature fu-
6232–6251. https://doi.org/10.1109/TGRS.2016.2584107. sion.” IEEE Trans. Geosci. Remote Sens. 54 (2): 783–794. https://doi
Chen, Y., Z. Lin, X. Zhao, G. Wang, and Y. Gu. 2014. “Deep learning- .org/10.1109/TGRS.2015.2465899.
based classification of hyperspectral data.” IEEE J. Sel. Top. Appl. Hang, R., Q. Liu, Y. Sun, X. Yuan, H. Pei, J. Plaza, and A. Plaza. 2017.
Earth Obs. Remote Sens. 7 (6): 2094–2107. https://doi.org/10.1109 “Robust matrix discriminative analysis for feature extraction from hy-
/JSTARS.2014.2329330.
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

perspectral images.” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.
Chen, Z. 2020. “Hyperspectral imagery for material surface damage.” Fig- 10 (5): 2002–2011. https://doi.org/10.1109/JSTARS.2017.2658948.
share. https://doi.org/10.6084/m9.figshare.12385994.v1. Harsanyi, J. C., and C. Chang. 1994. “Hyperspectral image classification
Chen, Z., J. Chen, F. Shen, and Y. Lee. 2015. “Collaborative mobile-cloud and dimensionality reduction: an orthogonal subspace projection ap-
computing for civil infrastructure condition inspection.” J. Comput. Civ. proach.” IEEE Trans. Geosci. Remote Sens. 32 (4): 779–785. https://doi
Eng. 29 (5): 04014066. https://doi.org/10.1061/(ASCE)CP.1943-5487 .org/10.1109/36.298007.
.0000377. Heylen, R., M. Parente, and P. Gader. 2014. “A review of nonlinear
Chen, Z., R. Derakhshani, C. Halmen, and J. T. Kevern. 2011. “A texture- hyperspectral unmixing methods.” IEEE J. Sel. Top. Appl. Earth
based method for classifying cracked concrete surfaces from digital im- Obs. Remote Sens. 7 (6): 1844–1868. https://doi.org/10.1109/JSTARS
ages using neural networks.” In Proc., 2011 Int. Joint Conf. on Neural .2014.2320576.
Networks, 2632–2637. New York: IEEE. Ho, H.-N., K.-D. Kim, Y.-S. Park, and J.-J. Lee. 2013. “An efficient image-
Chen, Z., and T. C. Hutchinson. 2010. “Image-based framework for based damage detection for cable surface in cable-stayed bridges.” NDT
concrete surface crack monitoring and quantification.” Adv. Civ. Eng. E Int. 58 (Sep): 18–23. https://doi.org/10.1016/j.ndteint.2013.04.006.
2010: 215295. https://doi.org/10.1155/2010/215295. Hossin, M., and M. N. Sulaiman. 2015. “A review on evaluation metrics
Cheng, H. D., and M. Miyojim. 1998. “Automatic pavement distress for data classification evaluations.” Int. J. Data Min. Knowl. Manage.
detection system.” Inf. Sci. 108 (1): 219–240. https://doi.org/10.1016 Process 5 (2): 1. https://doi.org/10.5121/ijdkp.2015.5201.
/S0020-0255(97)10062-7. Hu, W., and C. Jia. 2015. “A bootstrapping approach to entity linkage on
Cooley, T., et al. 2002.“FLAASH, a MODTRAN4-based atmospheric the semantic web.” J. Web Semant. 34 (Oct): 1–12. https://doi.org/10
correction algorithm, its application and validation.” In Proc., IEEE .1016/j.websem.2015.07.003.
Int. Geoscience and Remote Sensing Symp., 1414–1418. New York: Isawa, K., S. Nakayama, M. Ikeda, S. Takagi, S. Tosaka, and N. Kasai.
IEEE. 2005. “Robotic 3D SQUID imaging system for practical nondestructive
Cubert. 2018. “Hyperspectral VNIR Camera FireflEYE 185.” Accessed evaluation applications.” Physica C 432 (3–4): 182–192. https://doi.org
January 1, 2020. https://cubert-gmbh.com/product/firefleye-185 /10.1016/j.physc.2005.08.008.
-hyperspectral/. Ji, R., Y. Gao, R. Hong, Q. Liu, D. Tao, and X. Li. 2014. “Spectral-spatial
Dalal, N., and B. Triggs. 2005.“Histograms of oriented gradients for human constraint hyperspectral image classification.” IEEE Trans. Geosci.
detection.” In Vol. 881 of Proc., IEEE Computer Society Conf. on Remote Sens. 52 (3): 1811–1824. https://doi.org/10.1109/TGRS.2013
Computer Vision and Pattern Recognition (CVPR’05), 886–893. .2255297.
New York: IEEE. Johnson, W. R., D. W. Wilson, G. H. Bearman, and J. Backlund. 2004. “An
Duan, K.-B., J. C. Rajapakse, and M. N. Nguyen. 2007. “One-versus-one all-reflective computed tomography imaging spectrometer.” In Proc.,
and one-versus-all multiclass SVM-RFE for gene selection in cancer Instruments, Science, and Methods for Geospace and Planetary Remote
classification.” In Proc., Evolutionary Computation, Machine Learning Sensing, 88–97. Bellingham, WA: International Society for Optics and
and Data Mining in Bioinformatics, 47–56. Berlin: Springer. Photonics.
Fang, L., N. He, S. Li, A. J. Plaza, and J. Plaza. 2018. “A new spatial– Kaiser, P. K., and R. M. Boynton. 1996. Human color vision. Washington,
spectral feature extraction method for hyperspectral images using local DC: Optical Society of America.
covariance matrix representation.” IEEE Trans. Geosci. Remote Sens. Kaseko, M. S., and S. G. Ritchie. 1993. “A neural network-based method-
56 (6): 3534–3546. https://doi.org/10.1109/TGRS.2018.2801387. ology for pavement crack detection and classification.” Transp. Res.
Felzenszwalb, P. F., R. B. Girshick, D. McAllester, and D. Ramanan. 2010. Part C: Emerging Technol. 1 (4): 275–291. https://doi.org/10.1016
“Object detection with discriminatively trained part-based models.” /0968-090X(93)90002-W.
IEEE Trans. Pattern Anal. Mach. Intell. 32 (9): 1627–1645. https://doi Kim, H., S. Sim, and S. Cho. 2015. “Unmanned aerial vehicle (UAV)-
.org/10.1109/TPAMI.2009.167. powered concrete crack detection based on digital image processing.”
Foody, G. M., and A. Mathur. 2006. “The use of small training sets con- In Proc., Int. Conf. on Advances in Experimental Structural Engineer-
taining mixed pixels for accurate hard image classification: Training on ing. Urbana-Champaign, IL: Univ. of Illinois.
mixed spectral responses for classification by a SVM.” Remote Sens. Lattanzi, D. A., and G. Miller. 2013. “A prototype imaging and visualiza-
Environ. 103 (2): 179–189. https://doi.org/10.1016/j.rse.2006.04.001. tion system for robotic infrastructure inspection.” In Proc., Structures
Gavilán, M., D. Balcones, O. Marcos, D. F. Llorca, M. A. Sotelo, I. Parra, Congress 2013: Bridging Your Passion with Your Profession, 410–421.
M. Ocaña, P. Aliseda, P. Yarza, and A. Amírola. 2011. “Adaptive road Reston, VA: ASCE.
crack detection system by pavement classification.” Sensors 11 (10): Leshko, B. J. 2002. Guidelines for inspecting complex components of
9628–9657. https://doi.org/10.3390/s111009628. bridges. Washington, DC: AASHTO.
Goel, P. K., S. O. Prasher, R. M. Patel, J. A. Landry, R. B. Bonnell, and Li, W., G. Wu, F. Zhang, and Q. Du. 2017. “Hyperspectral image classi-
A. A. Viau. 2003. “Classification of hyperspectral data by decision trees fication using deep pixel-pair features.” IEEE Trans. Geosci. Remote
and artificial neural networks to identify weed stress and nitrogen status Sens. 55 (2): 844–853. https://doi.org/10.1109/TGRS.2016.2616355.
of corn.” Comput. Electron. Agric. 39 (2): 67–93. https://doi.org/10 Liu, Z., S. A. Suandi, T. Ohashi, and T. Ejima. 2002.“Tunnel crack detec-
.1016/S0168-1699(03)00020-6. tion and classification system based on image processing.” In Proc.,
Goutte, C., and E. Gaussier. 2005. “A probabilistic interpretation of Machine Vision Applications in Industrial Inspection X, 145–152. Bel-
precision, recall and F-score, with implication for evaluation.” In Proc., lingham, WA: International Society for Optics and Photonics.

© ASCE 04020057-15 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057


Loncan, L., L. B. De Almeida, J. M. Bioucas-Dias, X. Briottet, Ritchie, S. G. 1990. “Digital imaging concepts and applications in pave-
J. Chanussot, N. Dobigeon, S. Fabre, W. Liao, G. A. Licciardi, and ment management.” J. Transp. Eng. 116 (3): 287–298. https://doi.org
M. Simoes. 2015. “Hyperspectral pansharpening: A review.” IEEE /10.1061/(ASCE)0733-947X(1990)116:3(287).
Geosci. Remote Sens. Mag. 3 (3): 27–46. https://doi.org/10.1109/MGRS Roh, Y., G. Heo, and S. E. Whang. 2019. “A survey on data collection for
.2015.2440094. machine learning: A big data—AI integration perspective.” In Proc.,
Lu, G., and B. Fei. 2014. “Medical hyperspectral imaging: A review.” IEEE Transactions on Knowledge and Data Engineering. New York:
J. Biomed. Opt. 19 (1): 010901. https://doi.org/10.1117/1.JBO.19.1 IEEE.
.010901. Siesler, H. W., Y. Ozaki, S. Kawata, and H. M. Heise. 2008. Near-infrared
Marpu, P. R., P. Gamba, and I. Niemeyer. 2009. “Hyperspectral data spectroscopy: Principles, instruments, applications. New York: Wiley.
classification using an ensemble of class-dependent neural networks.”
Sohn, H., C. R. Farrar, F. M. Hemez, D. D. Shunk, D. W. Stinemates, B. R.
In Proc., 2009 1st Workshop on Hyperspectral Image and Signal
Nadler, and J. J. Czarnecki. 2003. A review of structural health mon-
Processing: Evolution in Remote Sensing, 1–4. New York: IEEE.
itoring literature: 1996–2001. Los Alamos, NM: Los Alamos National
Mercier, G., and M. Lennon. 2003. “Support vector machines for hyper-
spectral image classification with spectral-based kernels.” In Vol. 281 of Laboratory.
Proc., IGARSS 2003. 2003 IEEE Int. Geoscience and Remote Sensing Tang S., and Z. Chen. 2017. “Detection of complex concrete damage—A
Downloaded from ascelibrary.org by University of Liverpool on 11/14/20. Copyright ASCE. For personal use only; all rights reserved.

Symp. (IEEE Cat. No.03CH37477), 288–290. New York: IEEE. deep learning framework and performance evaluation.” In Int. Work-
Morris, T. 2004. Computer vision and image processing. Basingstoke, UK: shop on Computing for Civil Engineering (IWCCE). Seattle, WA:
Palgrave Macmillan. ASCE.
Ni, F., J. Zhang, and Z. Chen. 2019. “Pixel-level crack delineation in Tung, P.-C., Y.-R. Hwang, and M.-C. Wu. 2002. “The development of
images with convolutional feature fusion.” Struct. Control Health a mobile manipulator imaging system for bridge crack inspection.”
Monit. 26 (1): e2286. https://doi.org/10.1002/stc.2286. Autom. Constr. 11 (6): 717–729. https://doi.org/10.1016/S0926-5805
Olsen, M. J., A. Barbosa, P. Burns, A. Kashani, H. Wang, M. Veletzos, (02)00012-2.
Z. Chen, G. Roe, and K. Tabrizi. 2016. Assessing, coding, and marking Yeum, C. M., and S. J. Dyke. 2015. “Vision-based automated crack detec-
of highway structures in emergency situations, Volume 1: Research tion for bridge inspection.” Comput.-Aided Civ. Infrastruct. Eng.
overview. Washington, DC: Transportation Research Board. 30 (10): 759–770. https://doi.org/10.1111/mice.12141.
Ozer, E., M. Q. Feng, and D. Feng. 2015. “Citizen sensors for SHM: Zakeri, H., F. M. Nejad, and A. Fahimifar. 2017. “Image based techniques
Towards a crowdsourcing platform.” Sensors 15 (6): 14591–14614. for crack detection, classification and quantification in asphalt pave-
https://doi.org/10.3390/s150614591. ment: A review.” Arch. Comput. Methods Eng. 24 (4): 935–977. https://
Palsson, F., J. R. Sveinsson, M. O. Ulfarsson, and J. A. Benediktsson. 2015. doi.org/10.1007/s11831-016-9194-z.
“Model-based fusion of multi- and hyperspectral images using PCA and Zhang, C., and A. Elaksher. 2012. “An unmanned aerial vehicle-based im-
wavelets.” IEEE Trans. Geosci. Remote Sens. 53 (5): 2652–2663.
aging system for 3D measurement of unpaved road surface distresses.”
https://doi.org/10.1109/TGRS.2014.2363477.
Comput.-Aided Civ. Infrastruct. Eng. 27 (2): 118–129. https://doi.org
Plaza, A., et al. 2009. “Recent advances in techniques for hyperspectral
/10.1111/j.1467-8667.2011.00727.x.
image processing.” Remote Sens. Environ. Supplement, 113 (S1):
S110–S122. https://doi.org/10.1016/j.rse.2007.07.028. Zhang, L., F. Yang, Y. D. Zhang, and Y. J. Zhu. 2016. “Road crack
Prasanna, P., K. J. Dana, N. Gucunski, B. B. Basily, H. M. La, R. S. Lim, detection using deep convolutional neural network.” In Proc., 2016
and H. Parvardeh. 2014. “Automated crack detection on concrete IEEE Int. Conf. on Image Processing (ICIP), 3708–3712. New York:
bridges.” IEEE Trans. Autom. Sci. Eng. 13 (2): 591–599. https://doi IEEE.
.org/10.1109/TASE.2014.2354314. Zhang, Q., Y. Tian, Y. Yang, and C. Pan. 2014. “Automatic spatial–spectral
Ritchie, S. G. 1987. “Expert systems in pavement management.” Transp. feature selection for hyperspectral image via discriminative sparse mul-
Res. Part A: General 21 (2): 145–152. https://doi.org/10.1016/0191 timodal learning.” IEEE Trans. Geosci. Remote Sens. 53 (1): 261–279.
-2607(87)90007-0. https://doi.org/10.1109/TGRS.2014.2321405.

© ASCE 04020057-16 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2021, 35(1): 04020057

You might also like