See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/328448349

The Role of Hyperspectral Imaging: A Literature Review

Article  in  International Journal of Advanced Computer Science and Applications · September 2018

CITATIONS READS
0 336

4 authors, including:

Muhammad Mateen Dr Nasrullah

Chongqing University Chongqing University
5 PUBLICATIONS   2 CITATIONS    8 PUBLICATIONS   18 CITATIONS   

SEE PROFILE SEE PROFILE

Muhammad Azeem Akbar

Chongqing University
14 PUBLICATIONS   22 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

AZ model of software engineering View project

Software Requirements Change management and Implementation Process in the Domain of GSD. View project

All content following this page was uploaded by Muhammad Mateen on 23 October 2018.

The user has requested enhancement of the downloaded file.

 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

The Role of Hyperspectral Imaging: A Literature

Review
Muhammad Mateen, Junhao Wen, Nasrullah, Muhammad Azeem Akbar
School of Big Data and Software Engineering,
Chongqing University,
Chongqing, 401331, China

Abstract—Optical analysis techniques are used recently to with the spectra of sunlight that is scattered by objects at or
detect and identify the objects from a large scale of images. inside the earth. Remote images are planned to capture and
Hyperspectral imaging technique is also one of them. Vision of calculate the light returned from adjoining areas on the surface
human eye is based on three basic color (red, green and blue) of earth. Hyperspectral imaging can be applied to various
bands, but spectral imaging divides the vision into many more applications including medicines, biogeochemistry,
bands. Hyperspectral remote sensors achieve imagery data in the biophysics, industrial monitoring and remote sensing to collect
form of hundreds of adjoining spectral bands. In this paper, our the information for analysis.
purpose is to illustrate the fundamental concept, hyperspectral
remote sensing, remotely sensed information, methods for Chinese Academy of Science played a vital role in the
hyperspectral imaging and applications based on hyperspectral field of hyperspectral imaging and developed two outstanding
imaging. Moreover, in the forensic context, the novel methods imagers, one of them is known as Push broom Hyperspectral
involving deep neural networks are elaborated in this paper. The Imager (PHI) and another one is Operative Modular Imaging
proposed idea can be useful for further research in the field of Spectrometer (OMIS). In 2000, another invention about
hyperspectral imaging using deep learning. imaging came out in the form of Hyperspectral Digital
Camera (HSDC)which supports limited spectral bands with
Keywords—Deep learning; electromagnetic spectrum; high quality of spectral resolution [2]. HSDC plays a flexible
hyperspectral imaging; imaging spectroscopy; multispectral
role for different observation objects and applications
imaging; remote sensing
including environmental and agricultural monitoring.
I. INTRODUCTION Analysis of Hyperspectral imager can be performed by two
In hyperspectral, the term ―hyper‖ means ―too many‖ and different ways one of them is perspective of spectral analysis
it refers to the huge amount of measured wavelength bands. and another is based on image processing. It is more important
Hyperspectral images are used to provide sufficient spectral that the data should be well organized before selection of any
information to recognize and differentiate spectrally kind of approach. In spectroscopic analysis, the spectra should
distinctive materials. Optical analysis techniques are used to be extracted by region of interest that is usually calculated by
detect and identify the objects from a large scale of images. three different ways such as threshold an image with single
Hyperspectral imaging technique is one of them. Vision of waveband, ratio or difference image. In image processing, few
human eye is based on three basic colors (red, green and blue) images are selected from the collection of images for rapid
bands, but spectral imaging divides the vision into many more computation. Selection of those images is based on the
bands. Hyperspectral remote sensors achieve imagery data in importance of their wavelength for shifting carefully. In Fig. 2
the form of hundreds of adjoining spectral bands. [3] for careful shifting, spectra provides an option for peaks
and valleys whether in [4] based on original or preprocessed
The collective data is used to obtain a constant spectrum format.
for each imagery pixel mentioned in Fig. 1. After tuning the
sensor, terrain and atmospheric belongings are applied [1]. There are several other methods to achieve the same goal,
These imagery spectrums can be analyzed with laboratory or for example, partial least square regression and principal
field reflectance spectra to know and map surface materials component analysis. Additionally, some data compression
such as specific kind of plants or indicative minerals with ore techniques such as singular value decomposition and Fourier
deposits. transform are used for the process of more images to increase
the ability of hyperspectral imaging [5]. After the achievement
Imaging spectrometers are instruments used to produce of healthy data, the next step is to make dependent
hyperspectral images. The remote imaging and spectroscopy standardized models. Before implementation of Chemometric
are two basic technologies used to develop the hyperspectral algorithm, it is compulsory to overcome the noisy data to
sensors. Spectroscopy is a field of study about light emission increase the quality of signals. Moreover, image processing
or reflection from different materials and the changes occurred includes filtering and binning, which can increase the quality
in energy with wavelength. of data. In calibration model, the same routine is followed by
In the field of optical remote sensing, spectroscopy acts spectroscopy and spectral analysis as shown in Fig. 2.

51 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

Fig. 1. Measurement of hyperspectral remote sensor.

Fig. 2. Flowchart of hyperspectral imaging processing.

In this model different techniques such as principal artificial neural network and fuzzy inference are offered [3].
component analysis, partial least square discriminate analysis, As a result prediction map can be progressed with the use of
fisher’s linear discriminate technique, support vector machine, post- processing techniques and the information is achieved.

52 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

The rest of the paper is organized as: related work of the was a motivation for the development of digital image
study is discussed in Section II. Study findings are reported in processing methodologies by multispectral data analysis [11].
Section III, and Section IV covers the future work of the In the early stage, Goetz et al. [6] defined Hyperspectral
study. The applications of hyperspectral imaging are explained remote sensing as ―The collection of images in dozens to
in Section V. Finally, the findings are concluded in hundreds of adjacent registered spectral bands so that for
Section VI. every cell of an image a glowing spectrum can be calculated.‖
According to this definition, spectral regions of VIS (visible),
II. RELATED WORK NIR (Near Infrared), SWIR (Shortwave Infrared), MWIR
Light emitted by sun strikes on the molecules, causing in (Midwave Infrared), LWIR (Longwave Infrared) and UV
absorption and reflection of light which depends on the (Ultraviolet) are covered.
structure or architecture of molecules. The wavelength of A. Hyperspectral Remote Sensing
absorption or reflection of atomic bonds and molecule makes
detection or identification of a specific object. To collect the The hyperspectral sensor is used to study soil science,
amount of light reflected by a specific object or earth surface geology, mining, land use, and hydrology to map and identify
is used for identification of that object. There are different geometric and chemical patterns of land. Information obtained
types of sensors used to collect the scattered data including by hyperspectral sensor is used to identify valuable minerals
truck-mounted sensor, airborne sensor and satellite-based and petroleum. Imaging spectroscopy is a technique used to
sensor. With the help of fast computers and development of absorb specific features of chemical bonds in minerals.
sensors, the immense research [6] is being carried out in the Material mapping includes water, ice, snow, mineral mixtures,
field of ―Hyperspectral remote sensing‖. vegetation, environmental materials, atmospheric gases, and
man-made materials. Hyperspectral airborne scanners were
In [7], Sun, L et al. introduced a novel approach based on used in 1998 to identify rocks and soils containing
supervised classification for hyperspectral imaging. hydrocarbon [12]. This experiment was named as ―Pro Smart
Supervised classification model includes a spatially weighted Experiment‖ managed by German Aerospace Center to test
random fields and spectral data reliability. To progress the the Hyperspectral ―HyMap‖ designed by the Australian
quality of classification, the real labels of training data Company Integrated Spectronics Ltd. In Hyperspectral data
samples are fixed as a constraint in the model of proposed analysis, several factors are involved to make it complex for
approach. In [8], Sun, L. et al. introduced noise (deadlines, refined techniques and algorithms.
strip noises, impulse and Gaussian) reduction technique for
hyperspectral imaging with the use of low-rank representation To demonstrate the hyperspectral data analysis, themes are
on the spectral image. According to the proposed technique managed in six main areas including un-mixing, data fusion,
spectral space of hyperspectral image recline in the low-rank target detection, physical parameter retrieval, classification
subspace. The low rank representation based on spectral space and fast computing based on image processing, machine
was oppressed by nuclear norm of image with the spectral learning, and signal processing. According to [13], researchers
dimension. It causes potential for the removal of deadlines, described all the areas one by one in the paper. In the data
strips, sparse noise and Gaussian noise on the level of each fusion, various strategies are defined including restoration of
band. the signal to noise ratio (SNR) and spatialresolution, spectral
data fusion to overcome high spectral redundancy, spatial data
In the field of image classification, high dimensionality of fusion used to enhance the resolution of imaging sensor at the
imaging data is a challenge. Hyperspectral imaging data also sub-pixel level. There is another sensor named as ―Dual push
contains high dimensionality, so the spatial-spectral broom hyperspectral sensor‖ which contains two various
classification methods are demands for computational view devices (covers different wavelengths) that are normally
point. In [9], Wu, Z. et al. introduced a novel technique for attached to the similar bench. ―Co-registration of pixels‖ and
parallel implementation of spatial-spectral classification on the ―Peer of the field of view‖ are problems lead by this
basis of adaptive random fields. In this technique, logistic configuration called Bore sight Effect. To observe and detect
regression classifier is used for spectral information. In this spectral phenomena, A. Brook et al. [14] proposed a prior to
research, graphics processing units are used for parallel its correction with the use of Bore sight Effect that is used to
implementation. GPU sets the work allocation, input and provide supplementary information not located in the
output between the CPU and the graphics processing unit corrected images.
having fully computational control of GPUs with low latency
and high bandwidth of shared memory. B. Remotely Sensed Data
Hyperspectral sensors collect the data from the surface of
In [10], Sun, L. et al. introduced a novel approach for
the earth with the help of different parameters. It is important
noise reduction in the hyperspectral images. Noise reduction
to measure the accurate values of identified objects. There are
technique is based on the super pixel-based low level rank of
a lot of methods and techniques, developed to measure the
representation for hyperspectral imaging. Under the model of
values of vegetation. Some techniques show that the results
a linear mixture, the hyperspectral cube is considered as a low
can be obtained in highly accurate form but few techniques
rank in the spectral area, which is divided HSI data into the
show the results in moderately accurate form.
sub-matrices of lower ranks.
In [15], Mutanga et al. studied on a small number of
In 1985, a technique named as ―Imaging spectrometry‖
chemicals, for example, chlorophyll and another kind of
was used for the remote sensing of the earth. In addition, there
pigments. The pigments including carotenes and chlorophyll

53 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

are identifiable with 80 per cent accuracy in sample fields. as ―eCognition‖ is used to offer the suitable connection
Nitrogen can also be detected with accuracy. But the other between remote sensing imagery and graphical information
nutrients namely magnesium, calcium, sodium, potassium, and system. Thus object-oriented technique can provide powerful
phosphorus have less notified by Mutanga et al. in his automatic and semiautomatic evaluation to allocate practiced
research. Even though this is very expensive for projects to knowledge to GIS.
analyze with a lot of samples, the impending of remote
sensing is not totally oppressed. In the [16], research related III. DISCUSSIONS
with Geoinformation science and earth analysis for modeling Hyperspectral imagery process follows some methods to
and management of environment designed a remote sensing accomplish the proposed task. These methods contain study
data model in which ―spot vegetation‖ concerned with geo- area, which means where experiment is performed; field
referenced and ―Corine land cover data‖ concerned with sampling, meaning the sample data gathered from study area;
study area that provides qualitative and quantitative Chemical analysis, which means evaluation of chemical
information of earth surface. This research work used datasets compounds and Spectral processing, which means to perform
with some characteristics including geometric accuracy better operation on sample of spectra to overcome noisy data,
than 100 m, thematic accuracy greater than 85 percent, and statistical formulas are used for prediction of parameters. D.
spatial resolution of 30m. Liao et al. [21] proposed method to visualize hyperspectral
For the development of South African Imaging image in normal color by the coordination of HSI and high
Spectroscopy Landscape, an overview is explained in brief resolution image through multiple alignments. Manifold
about analysis techniques and challenges [17]. According to alignment identifies the matching color points and displays it
African research and application, indicators can be derived by into pair wise alignment, while the spaces between them play
the full width of the spectral response for example slopes, the role of bridges. Association of hyperspectral imaging and
integrals, and curve derivatives. The purpose is to identify RGB image generates a spatial image in the natural format.
data source from the imaging spectroscopy, in which spectral This approach has advantage of flexibility that’s why it can be
attributes are required to notify an application of interest. applied for different scenario. The same approach is also well
According to [18], object-based classification approach is used known because of user’s customization, in which user can
to classify the remotely sensed data. In 2000, Walter and scan visual resultant bands according to the specified interest.
Fritsch introduced a concept using multispectral remote Chemometric and spectroscopy areas provided methods and
sensing data to automatically revise the geographical analysis tools including ―Partial Least Squares Regression
information system databases. In object-based classification, Analysis‖ and ―Principal Component‖, which are useful for
not only single pixel is classified but the group of pixels is the process of hyperspectral image [22]. Workflow of
also classified based on geographical information system hyperspectral image processing is different from the workflow
database. This approach depends on two main steps, the first of color image processing but both data types are multivariate
step is based on supervised likelihood classification and the and multidimensional. Hyperspectral imaging methods such as
second step is based on matching of existing GIS objects with image acquisition, spectral and spatial preprocessing,
classified objects to detect the change occurred or to identify dimensionality reduction, calibration, feature extraction and
the faulty data. This approach develops and makes the better selection are used for HSI processing as shown in Fig. 2. Push
result of the classification. At the end, the result is obtained in broom line scanner is commonly used as distributed
the form of fully classified, partly classified and not found by hyperspectral imaging sensor. Push broom scans all
the use of a threshold, iteratively defined by a user. In [19], a wavelength data points with the same spatial coordinates. In
surface bidirectional reflectance model has been designed. The the hyperspectral imaging system, calibration is a method that
model is applicable for heterogeneous surfaces and follows a ensures the repeatability and accuracy of results gained by
semi-empirical approach. There are three parameters used to hyperspectral imaging data. Calibration is a procedure that
describe the nature of the surface and these parameters are connects wavelengths with band numbers. Spatial calibration
used in the effective algorithm of correction and processing is a method of measuring the correlation of each image pixel
the remote sensing data. The model based on surface such as meters. Spatial preprocessing is also a method used in
bidirectional reflectance is collection of two basic processes, hyperspectral imaging to minimize the noisy data from the
one of them is to diffuse reflection component which explains images. Spatial post-processing is considered more valuable
the geometrical structure on surface and also understands the than spatial preprocessing because prediction and
shadow effects while the second process defines the volume classification images are common images that need spatial
scattering properties obtained by isolated surfaces. Invisible manipulation, interpretation and pattern identification. Spatial
and near-infrared spectral bands, model and situ annotations sampling and Region of Interest (ROI) is a basic method of
demonstrate the better association for common investigated hyperspectral imaging used to mention the study area or
surface types. The model is good to reduce the largely location of interest for hyperspectral sensors. Spatial sampling
unwanted variations related to surface bidirectional results in can also be considered as background masking produced by
remotely sensed data. binary images [22]. Transformation to absorbance is also used
during reflectance or transmittance of data for analysis [23].
U.C. Benz et al. [20] proposed object-oriented analysis Hyperspectral imaging contains noisy data during scanning a
associated with fuzzy methods to implement information and large scale of image. To reduce noisy data, de-noising
explained workflow from remote sensing imagery to graphical algorithm is performed in the spectral domain of hyperspectral
information system. During proposed method, software named imaging [24]. Feature extraction is also essential method used

54 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

to transform the existing features to a set of new features. representative bands from the noisy data with a smart way. To
Selection of subset from input feature without transformation overcome noisy data in an efficient way, two sensors are
is called selection extraction. It is also called feature subset applied, so the experimental results show that proposed
selection. Mosaic based images [25] which are merged into a technique is better performed than Manual Band Removal
particular hyperspectral image, is used for the data analysis technique.
and classification algorithms.
To overcome the noisy data from hyperspectral images is
In [26], affinity propagation method is used for the also introduced by [28], a proposed technique of sparse
selection of bands in hyperspectral imaging. In this method representation based on noise reduction method. This
centered based clustering approach is used to classify the technique depends on a non-noisy component which can be
similar color bands. AP is applied by factor graph and then sparsely decayed over redundant dictionary rather than a noisy
operation performed by the centralized data points through component. The paper shows the correlation of spectral-spatial
message passing until an appropriate set of bands is achieved. structure of HSI by use of three dimensional blocks instead of
The presentation of band selection is analyzed by the two dimensional patches for sparse representation. Gaussian
classification of a pixel inside the specific image. and Poisson noise models are collectively used for signal-
Classification of pixels is better performed by affinity dependent and signal independent noises in hyperspectral
propagation instead of all other approaches. Unlike imagery. The proposed technique is good for virtual and real
conventional methods of clustering named as K-means, data sets of HS remote sensing.
agglomerative clustering, a proposed technique (AP) obtained
better results of band selection through message passing Naganathan et al. [29] proposed meat tenderness for the
method. Hyperspectral imaging contains massive information satisfaction of consumers. Purpose of this proposal was to
of colored bands. Collection of colored bands can have noisy build up and test a near infrared/visible hyperspectral imaging
data that is considered to remove before processing. During system to guess tenderness of cooked beef by hyperspectral
preprocessing manual band removal technique is used to images. For this purpose, a push-broom hyperspectral imaging
system associated with diffuse floodlighting system was
remove the noisy data. In [27], problem related with MBR is
identified, because during MBR an important data can also be designed and standardized. Three tenderness categories
removed. For this problem, a proposed technique is introduced including tender, intermediate and tough were used to identify
to automatically select the noisy bands instead of MBR the features of meat tenderness. Statistical textual features
selection. In the proposed technique, first wavelet reduction is obtained from Slice Shear Force (SSF) analysis are used in
applied to de-noise the bands of specified image and then ―canonical discriminate model‖ for prediction. The results
Affinity Propagation approach is used to classify the signify that hyperspectral imaging played a vital role in the
prediction of meat tenderness.

Fig. 3. Schematic diagram of the ISL hyperspectral imaging system.

55 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

TABLE I. DETECTION RESULTS OF SKIN TUMOR analysis standard [33]. J. Liang et al. [34], proposed random
sampling approach for spatial-spectral techniques to decrease
the overlapping between testing and training samples and offer
No. of
Real
No. of No. of False
No. of Missing more intention and precise evaluation. Random sampling
Image Detections Positives approach [35] is often used because it cares all tagged data
Tumors
evenly and every sample is picked with the similar likelihood.
1 3 3 1 0
In [36], there is a review of hyperspectral imaging analysis
2 3 3 0 0 with different techniques. In [36], Artificial Neural Network,
3 1 1 1 0 Auto-Encoder, Stack Auto-Encoder, Conventional Neural
4 3 2 1 1 Network and deep leaning is discussed for the analysis of
5 4 4 1 0
Hyperspectral Imaging. After comparison of all these
techniques and getting results, the author mentioned that deep
6 2 1 1 1 learning outperformed in the analysis of hyperspectral
7(2) 3 3 1 0 imaging among all other stated techniques.
8(2) 7 5 3 2
Furthermore in [37], there is a new model named as R-
9(2) 3 3 3 0 VCANet designed in the context of deep learning for
10 12 6 0 6 classification of hyperspectral imaging. R- VCANet model is
a combination of Rolling Guidance Filter and Vertex
Total 41 31(76%) 12(28%) 10(24%) Component Analysis Network. R-VCANet is useful when
there is limited sampling for feature extraction of
Hyperspectral Imaging system is used for detection of skin hyperspectral data. R-VCANet is based on natural
tumor [30] , in which first of all the carcass of chicken is characteristics of HIS data, spectral properties, and spatial
analyzed for the usage of spectral information as shown in information. Hence the method proposed in [37] has
Fig. 3, and results are applied for the detection of skin tumor performed better for hyperspectral image classification,
as shown in Table I. especially when the sampling labels are limited.
There was a little deficiency about detection that tumor Geophysics plays a vital role to study about graves
less than 3 millimeters in diameter could not be identified. detection after some changes are occurred in Buried graves.
This step in research opens a gate for the detection of tumors To detect the clandestine graves, local environment and
by hyperspectral imaging and can also be improved for different types of soils can be observed and data can be
computational complexity. collected which is used for analysis. So it is important to
In the field of poultry farms, hyperspectral imaging is used become familiar with the equipment and hardware that is used
to design a rapid, accurate and non-distractive method to for data acquisition and data analysis. John J. Schultz [38]
detect the embryo and fertility improvement in eggs [31]. A used ground-penetrating radar for detection of clandestine
near infrared spectral imaging system was introduced for the graves. To use the GPR, it is also important to adjust the
identification of fertility and early embryo improvement. For antenna for the best frequency of forensic work, which
the experiment, totally 174 white shell eggs were used depends on spoidosol environment. Electromagnetic induction
including 18 infertile eggs and 156 fertile eggs incubated for 4 meter is also used to locate the clandestine graves, but EMI is
days. During the inspection by hyperspectral imaging, eggs limited for forensic research, so the lack of published research
were categorized into two parts one for fertile and other for with the use of EMI meter to detect clandestinely buried
infertile and the dataset of each category is different with the bodies became deem in the forensic context. As mentioned
day of incubation. Gabor filter used to extract the image above that EMI is limited for forensic research, a new idea is
texture information of eggs. K means clustering technique is introduced to do better in the forensic field. According to that,
also applied to cluster out the data and top results were buried remains can be easily detected by hyperspectral
achieved as 84.1% at day 4, 81.8% at day 3, 74.1% at day 2 imaging. By hyperspectral image, a large scale of an image
and 100 % at day 0. The result shows that last three bands can be achieved and with the application of deep neural
were used for detection because of maximum response of networks, it can be classified in a more précised way.
spectral transmission and can be applied for real-time For un-mixing, classification and target detection of
detection system for early embryo and fertility of chicken hyperspectral images, there are frequently used methods
eggs. In the field of hyperspectral imaging, a large number of includes, sparse linear models, Gaussian mixture models,
researchers are focused to design different kinds of spectral- latent linear models, ensemble learning, directed graphical
spatial techniques while remaining are paying attention models, support vector machines, linear regression, logistic
towards experimental development [32]. To evaluate the regression, Gaussian models, clustering and deep learning.
hyperspectral imaging classification techniques, it needs the The summary of all these methods is highlighted in Table II
reliable plan of testing including suitable standardized data with referred research articles.
sets, sampling procedure for training and testing data, and fair

56 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

TABLE II. METHODS FOR HYPERSPECTRAL IMAGING

Un-mixing Classification Target Detection

Methods
Linear Non-linear Spatial-spectral Pixel-wise Target Anomaly
spatial-
spatial-spectral
Sparse Linear Models feature extraction [40] feature extraction [41] spectral
[39]
[39]
Gaussian Mixture Models [42] un-supervised [43] [44]
dimensionality reduction
Latent Linear Models [45] feature extraction [46]
[47]
Ensemble Learning [48] transfer learning [49] [50]
sub-pixel spatial-spectral
Directed Graphical Models
mapping [51] [52]
end member
extraction and
Support Vector Machines [54] [55] band selection [56] [57] [58]
sub-pixel
mapping [53]
Linear Regression [59] [60]
Logistic Regression semi-supervised [61] band selection [62]
Gaussian Models transfer learning [63] [64] [65]
Clustering un-supervised [66] [67]
un-supervised feature supervised feature
learning [68] learning [70]
Deep Learning [72]
supervised feature un-supervised feature
learning [69] learning [71]
IV. FUTURE WORK However, the suspected areas are typically very large
rendering conventional ways of grave detection such as using
To locate clandestine graves is a challenge for government forensic detection dogs, ground-penetrating radars, and
(forensic department) after any victims. There is a lot of thermal imaging inefficient, let alone the huge cost and large
traditional ways to locate and detect the clandestine graves, for human resources required in the search operations.
example with the help of forensic trained dogs, ground
penetrating radar, electromagnetic induction meter, and hyper In this case, the idea of examining the emission and
spectral sensor, in the forensic context. There are lots of biogeochemistry of gases from graves and their detection
researchers who are doing work to facilitate the detection of through remote sensing, in particular, hyper-spectral images
clandestine graves. In case of sudden disaster, the government sounds like an interesting idea. A hyperspectral imager
forensic agencies play a vital role to detect the victims. mounted on an Unmanned Airborne Vehicle (UAV) can fly
Forensic archaeologist and anthropologists face a lot of over the suspected area and acquire imagery which may
challenges in forensic context. contain the absorption features (signature) of the gases of
interest predominantly CH4, N2O, and CO2. The higher pore
One problem which Saudi emergency responders have to air concentrations in graves and emission of methane, carbon
deal with is that sometimes heavy rains result in the flood. dioxide and nitrous oxide to the atmosphere imply the
Victims of floods get buried deep in sand at ranges from 50cm existence of graves in that specific location, which can be
up to 2 meters. The soil under which the victims are buried is associated with the missing victim. In contrast with dry sandy
likely clayey loam and poorly drained, and it can be covered soil which is aerobic and therefore conducive to methane
with water for some time (few days). The emergency consumption and/or oxidation, the search area is clayey loam
responders focus their search operations in areas where and poorly drained soil, and thus susceptible to methane
victims are likely to be located such as valleys (e.g., Alhayer production. This increases the chance of detecting gases of
valley) shown in Fig. 4. interest using hyper-spectral imaging techniques.
For better understanding about hyperspectral imaging
using deep learning, we demonstrated the process of feature
extraction using the deep neural network. In Fig. 5, there is a
data patch of hyperspectral image for input layer and on the
next number of hidden layers; hyperspectral image is further
divided for feature extraction process. After the process of
hybrid feature extraction results can be achieved from the
output layer. The outcome can be in the form of 1 or 0 that
shows the presence and absence of buried remains
respectively. According to the current research trends, deep
learning has become an advanced and robust technique to
extract the features of hyperspectral image as compared to the
traditional feature extraction techniques.
Fig. 4. Heavy Rain Flood in Alhayer Valley (Saudi Arabia).

57 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

Fig. 5. Hyperspectral imaging using deep neural network.

Fig. 6. Flowchart of future study.

The detection of clandestine and previously unknown  High spec PC/Laptop for experiment and data analysis
burial sites is of interest to governments to start rescue
operations. The detection of clandestine graves is an emerging Techniques\tools used for Completion of the Project:
tool in hyper-spectral remote sensing. In literature, studies  Matlab 2017
have shown that it is possible to use hyperspectral remote
sensing techniques in detection of mass graves. For this  Hybrid Feature Extraction
purpose, a UAV with hyperspectral sensors covering the
 Deep Neural Network (Classification)
visible to the shortwave infrared range was used to collect the
imagery. The flow chart of the proposed technique is shown in V. IMPLICATIONS OF HYPERSPECTRAL IMAGING
Fig. 6.
Hyperspectral imaging is widely used in various
Hardware Required for Completion of Project: applications such as Agriculture, Mineralogy, Surveillance,
Physics, Astronomy, Chemical Imaging, and Environment. In
 Hyperspectral Sensor (data acquisition) the agricultural industry, diseases cause a serious loss for the
 An Unmanned Airborne Vehicle (UAV) economy. For suitable agriculture, it is important to monitor

58 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

plants and trees for diseases. To detect diseases on early stage perform the services. The hyperspectral sensor can also
can save the plants and trees from further loss, Hyperspectral provide thermal information for a short interval of time over a
sensor can assist to control the virus by organized methods large amount of area shown in Fig. 7. Brightness Temperature
such as fungicide applications, disease-specific chemical is a sensor used to detect the temperature of objects.
applications and pesticide applications [73]. The hyperspectral
imagery is used in land cover mapping. Land cover Every object has different emissivity which discharges
classification associated with the nature of land such as electromagnetic energy. The value of emissivity is nearly
grassland, forest, concrete pavement, and sand etc. On the equal to 1 and remains constant as compared with the
other hand, Land use indicates the human use of land, for temperature of the earth. Snow covered area is also the
example, industrial, residential and agricultural zone. Land use detectable area for hyperspectral sensors. During snow survey,
classification is an application used to classify the land surface snow water equivalent has been planned with the
such as Geometric correction, ground truth, and maximum approximation of snow-covered area. Height measurement can
likelihood. Land cover change detection is another application also be calculated by Hyperspectral imager with the matchup
used to detect the changes in the earth surface. It can be of stereo images. In [74], there are two techniques to measure
identified with the comparison of existing image and the the height of objects, one is already defined as stereo matching
updated one. Changes on earth have two main types. One is and other is based on analytical plotters.
seasonal change and other is annual change, in seasonal Fruits are the important part of food because it provides
change forests or plants get changes according to the season, vitamins and energy to human’s body. To gain healthy and
but in annual changes, new things take place for example fresh fruits is also a challenge. Hyperspectral imager provides
developments on the earth or deforested place. Global the applications to monitor the quality of fruits. Peng and Lu
vegetation map is another application of hyperspectral [75] developed a reflective system to identify apple firmness
imaging. According to this application global vegetation index and solids contents with the use of steady object stage. With
data contains information of normalized difference vegetation the help of optical fiber and focusing lenses, 2-dimensional
index (NDVI). NDVI is updated on the weekly basis, so it can hyperspectral images were obtained. Huang and Lu [76]
also contain some noisy data. Water is an essential part of life. introduced another quality attribute analyzed by hyperspectral
Water quality monitoring is an important application of imaging is known as mealiness. Haung and Lu analyzed the
hyperspectral imager. Quality of water can be analyzed due to association between apple mealiness and reflective
its greenish or yellowish shaded color. To measure the hyperspectral line images. The mealiness of apple was
surface temperature of sea hyperspectral imager can also calculated by the solidity and ripeness.

Fig. 7. Incident light into a sensor on the sea.

59 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

Remote sensing of vegetation is widely used in the world [4] Wang, H., Peng, J., Xie, C., Bao, Y. and He, Y. Fruit quality evaluation
to check the quality of vegetables. The important application using spectroscopy technology: a review. Sensors, 15, 5 2015), 11889-
11927.
of hyperspectral imaging about vegetables including
[5] Bonnier, F., Bertrand, D., Rubin, S., Venteo, L., Pluot, M., Baehrel, B.,
mushrooms, onions, cherry tomato, and spinach leaves. Ariana Manfait, M. and Sockalingum, G. Detection of pathological aortic
and Lu [77] designed a VIS-NIR hyperspectral imaging tissues by infrared multispectral imaging and chemometrics. Analyst,
system joining reflective mode and transmitted mode 133, 6 2008), 784-790.
collectively, while with the use of moving transport proposal. [6] Lanaras, C., Baltsavias, E. and Schindler, K. ADVANCES IN
The system was able to identify internal issues of pickles and HYPERSPECTRAL AND MULTISPECTRAL IMAGE FUSION AND
cucumbers which were impossible to visualize by the human SPECTRAL UNMIXING. International Archives of the
Photogrammetry, Remote Sensing & Spatial Information Sciences,
eye. Hyperspectral sensors can also be used to monitor the 402015).
quality of meat such as pork, chicken, and beef. In the context [7] Sun, L., Wu, Z., Liu, J., Xiao, L. and Wei, Z. Supervised spectral–spatial
of pattern recognition, marbling is used to detect the quality of hyperspectral image classification with weighted Markov random fields.
meat. To extract the marbling, a proposed technique used a IEEE Transactions on Geoscience and Remote Sensing, 53, 3 2015),
wide line detector implemented by pattern recognition [78]. 1490-1503.
Results are achieved with the help of standard marbling charts [8] Sun, L., Jeon, B., Zheng, Y. and Wu, Z. Hyperspectral image restoration
and classified with the accuracy of 99 per cent. Hyperspectral using low-rank representation on spectral difference image. IEEE
Geoscience and Remote Sensing Letters, 14, 7 2017), 1151-1155.
imaging is also applied for seafood such as prawn, cod and
smoked salmon. Near Infrared (NIR) Hyperspectral System [9] Wu, Z., Shi, L., Li, J., Wang, Q., Sun, L., Wei, Z., Plaza, J. and Plaza, A.
GPU parallel implementation of spatially adaptive hyperspectral image
was designed to detect the grains diseases on the early stage classification. IEEE Journal of Selected Topics in Applied Earth
with the use of mathematical modeling framework. This Observations and Remote Sensing, 11, 4 2018), 1131-1143.
system was based on supervised classification [79]. In [80], [10] Sun, L., Jeon, B., Soomro, B. N., Zheng, Y., Wu, Z. and Xiao, L. Fast
Khan et al. studied on advanced developments in the field of Superpixel Based Subspace Low Rank Learning Method for
HIS Analysis. Details of this review based on fundamentals of Hyperspectral Denoising. IEEE Access, 62018), 12031-12043.
Hyperspectral Imaging including spatial resolution, spectral [11] Goetz, A. F., Vane, G., Solomon, J. E. and Rock, B. N. Imaging
resolution, temporal resolution, feature extraction. Secondly, spectrometry for earth remote sensing. Science, 228, 4704 1985), 1147-
1153.
authors explained the advance applications of hyperspectral
[12] Moser, G., Serpico, S. B. and Benediktsson, J. A. Land-cover mapping
imaging such as food quality, safety, security, remote sensing, by Markov modeling of spatial–contextual information in very-high-
especially in the field of forensic documents examination to resolution remote sensing images. Proceedings of the IEEE, 101, 3
study about intrinsic and extrinsic elements. After a long 2013), 631-651.
discussion about hyperspectral imaging, we have come to [13] Bioucas-Dias, J. M., Plaza, A., Camps-Valls, G., Scheunders, P.,
know that hyperspectral imaging is an active area for research, Nasrabadi, N. and Chanussot, J. Hyperspectral remote sensing data
through which new ideas can be implemented and the world analysis and future challenges. IEEE Geoscience and remote sensing
magazine, 1, 2 2013), 6-36.
can be changed.
[14] Brook, A. and Ben-Dor, E. Advantages of the boresight effect in
VI. CONCLUSION hyperspectral data analysis. Remote Sensing, 3, 3 2011), 484-502.
[15] Mutanga, O. and Skidmore, A. K. Continuum-removed absorption
In this paper, we described the fundamental concept, features estimate tropical savanna grass quality in situ. City, 2003.
hyperspectral remote sensing, remotely sensed information, [16] D az-Delgado, R. and Pons, X. Spatial patterns of forest fires in
methods for hyperspectral imaging and applications based on Catalonia (NE of Spain) along the period 1975–1995: analysis of
hyperspectral imaging. We introduced a better approach to vegetation recovery after fire. Forest ecology and management, 147, 1
detect the buried remains with the use of deep neural networks 2001), 67-74.
for feature extraction. Finally, we focused on the use of [17] Mutanga, O., Van Aardt, J. and Kumar, L. Imaging spectroscopy
(hyperspectral remote sensing) in southern Africa: an overview. South
hyperspectral imaging in different applications. However, African Journal of Science, 105, 5-6 2009), 193-198.
hyperspectral imaging technology using deep learning is
[18] Walter, V. Object-based classification of remote sensing data for change
gradually becoming a great opportunity for researchers, in the detection. ISPRS Journal of photogrammetry and remote sensing, 58, 3-
field of biomedical, vegetation, especially in the forensic 4 2004), 225-238.
context. [19] Roujean, J. L., Leroy, M. and Deschamps, P. Y. A bidirectional
reflectance model of the Earth's surface for the correction of remote
ACKNOWLEDGMENT sensing data. Journal of Geophysical Research: Atmospheres, 97, D18
1992), 20455-20468.
This research was supported by the Basic and Advanced
[20] Benz, U. C., Hofmann, P., Willhauck, G., Lingenfelder, I. and Heynen,
Research Projects in Chongqing, China under Grant M. Multi-resolution, object-oriented fuzzy analysis of remote sensing
No.61672117. data for GIS-ready information. ISPRS Journal of photogrammetry and
remote sensing, 58, 3-4 2004), 239-258.
REFERENCES
[21] Liao, D., Qian, Y., Zhou, J. and Tang, Y. Y. A manifold alignment
[1] Ahmadi, S. B. B., Nanehkaran, Y. A. and Layazali, S. Review on hyper-
approach for hyperspectral image visualization with natural color. IEEE
spectral imaging system. Int. J. Sci. Eng. Res., 4, 5 2013), 253-258.
Transactions on Geoscience and Remote Sensing, 54, 6 2016), 3151-
[2] Tong, Q., Zhang, B. and Zheng, L. S. Hyperspectral remote sensing 3162.
technology and applications in China2004).
[22] Park, B., Yoon, S.-C., Windham, W. R., Lawrence, K. C., Kim, M. S.
[3] Feng, Y.-Z. and Sun, D.-W. Application of hyperspectral imaging in and Chao, K. Line-scan hyperspectral imaging for real-time in-line
food safety inspection and control: a review. Critical reviews in food poultry fecal detection. Sensing and instrumentation for food quality and
science and nutrition, 52, 11 2012), 1039-1058. safety, 5, 1 2011), 25-32.

60 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

[23] Burns, D. A. and Ciurczak, E. W. Handbook of near-infrared analysis. [44] Tarabalka, Y., Haavardsholm, T. V., Kåsen, I. and Skauli, T. Real-time
CRC press, 2007. anomaly detection in hyperspectral images using multivariate normal
[24] Cen, H. and He, Y. Theory and application of near infrared reflectance mixture models and GPU processing. Journal of Real-Time Image
spectroscopy in determination of food quality. Trends in Food Science Processing, 4, 3 2009), 287-300.
& Technology, 18, 2 2007), 72-83. [45] Nascimento, J. M. and Dias, J. M. Does independent component analysis
[25] Yoon, J. H., Sheremata, S., Rokem, A. and Silver, M. A. Windows to play a role in unmixing hyperspectral data? IEEE Transactions on
the soul: vision science as a tool for studying biological mechanisms of Geoscience and Remote Sensing, 43, 1 2005), 175-187.
information processing deficits in schizophrenia. Frontiers in [46] Plaza, J., Plaza, A. J. and Barra, C. Multi-channel morphological profiles
psychology, 42013), 681. for classification of hyperspectral images using support vector machines.
[26] Qian, Y., Yao, F. and Jia, S. Band selection for hyperspectral imagery Sensors, 9, 1 2009), 196-218.
using affinity propagation. IET Computer Vision, 3, 4 2009), 213-222. [47] Shaw, G. and Manolakis, D. Signal processing for hyperspectral image
[27] Jia, S., Ji, Z., Qian, Y. and Shen, L. Unsupervised band selection for exploitation. IEEE Signal processing magazine, 19, 1 2002), 12-16.
hyperspectral imagery classification without manual band removal. [48] Merentitis, A., Debes, C. and Heremans, R. Ensemble learning in
IEEE Journal of Selected Topics in Applied Earth Observations and hyperspectral image classification: Toward selecting a favorable bias-
Remote Sensing, 5, 2 2012), 531-543. variance tradeoff. IEEE Journal of Selected Topics in Applied Earth
[28] Qian, Y. and Ye, M. Hyperspectral imagery restoration using nonlocal Observations and Remote Sensing, 7, 4 2014), 1089-1102.
spectral-spatial structured sparse representation with noise estimation. [49] Rajan, S., Ghosh, J. and Crawford, M. M. Exploiting class hierarchies
IEEE Journal of Selected Topics in Applied Earth Observations and for knowledge transfer in hyperspectral data. IEEE Transactions on
Remote Sensing, 6, 2 2013), 499-515. Geoscience and Remote Sensing, 44, 11 2006), 3408-3417.
[29] Naganathan, G. K., Grimes, L. M., Subbiah, J., Calkins, C. R., Samal, A. [50] Peerbhay, K. Y., Mutanga, O. and Ismail, R. Random forests
and Meyer, G. E. Visible/near-infrared hyperspectral imaging for beef unsupervised classification: The detection and mapping of
tenderness prediction. Computers and electronics in agriculture, 64, 2 solanummauritianum infestations in plantation forestry using
2008), 225-233. hyperspectral data. IEEE Journal of Selected Topics in Applied Earth
[30] Kim, I., Kim, M., Chen, Y. and Kong, S. Detection of skin tumors on Observations and Remote Sensing, 8, 6 2015), 3107-3122.
chicken carcasses using hyperspectral fluorescence imaging. [51] Zhao, J., Zhong, Y., Wu, Y., Zhang, L. and Shu, H. Sub-pixel mapping
Transactions of the ASAE, 47, 5 2004), 1785. based on conditional random fields for hyperspectral remote sensing
[31] Liu, L. and Ngadi, M. Detecting fertility and early embryo development imagery. IEEE Journal of Selected Topics in Signal Processing, 9, 6
of chicken eggs using near-infrared hyperspectral imaging. Food and 2015), 1049-1060.
Bioprocess Technology, 6, 9 2013), 2503-2513. [52] Altmann, Y., Pereyra, M. and McLaughlin, S. Bayesian nonlinear
[32] Lu, D. and Weng, Q. A survey of image classification methods and hyperspectral unmixing with spatial residual component analysis. IEEE
techniques for improving classification performance. International Transactions on Computational Imaging, 1, 3 2015), 174-185.
journal of Remote sensing, 28, 5 2007), 823-870. [53] Villa, A., Chanussot, J., Benediktsson, J. A. and Jutten, C. Spectral
[33] Friedl, M., Woodcock, C., Gopal, S., Muchoney, D., Strahler, A. and unmixing for the classification of hyperspectral images at a finer spatial
Barker-Schaaf, C. A note on procedures used for accuracy assessment in resolution. IEEE Journal of Selected Topics in Signal Processing, 5, 3
land cover maps derived from AVHRR data2000). 2011), 521-533.
[34] Liang, J., Zhou, J., Qian, Y., Wen, L., Bai, X. and Gao, Y. On the [54] Gu, Y., Wang, S. and Jia, X. Spectral unmixing in multiple-kernel
sampling strategy for evaluation of spectral-spatial methods in hilbert space for hyperspectral imagery. IEEE Transactions on
hyperspectral image classification. IEEE Transactions on Geoscience Geoscience and Remote Sensing, 51, 7 2013), 3968-3981.
and Remote Sensing, 55, 2 2017), 862-880. [55] Benediktsson, J. A., Palmason, J. A. and Sveinsson, J. R. Classification
[35] Mas, J.-F., Pérez-Vega, A., Ghilardi, A., Martínez, S., Loya-Carrillo, J. of hyperspectral data from urban areas based on extended morphological
O. and Vega, E. A suite of tools for assessing thematic map accuracy. profiles. IEEE Transactions on Geoscience and Remote Sensing, 43, 3
Geography Journal, 20142014). 2005), 480-491.
[36] Petersson, H., Gustafsson, D. and Bergstrom, D. Hyperspectral image [56] Bazi, Y. and Melgani, F. Toward an optimal SVM classification system
analysis using deep learning—a review. IEEE, City, 2016. for hyperspectral remote sensing images. IEEE Transactions on
Geoscience and Remote Sensing, 44, 11 2006), 3374-3385.
[37] Pan, B., Shi, Z. and Xu, X. R-VCANet: a new deep-learning-based
hyperspectral image classification method. IEEE Journal of Selected [57] Sakla, W., Chan, A., Ji, J. and Sakla, A. An SVDD-based algorithm for
Topics in Applied Earth Observations and Remote Sensing, 10, 5 2017), target detection in hyperspectral imagery. IEEE Geoscience and Remote
1975-1986. Sensing Letters, 8, 2 2011), 384-388.
[38] Roedl, G., Elmes, G. A. and Conley, J. Spatial technology applications. [58] Gurram, P. and Kwon, H. Support-vector-based hyperspectral anomaly
Springer, City, 2014. detection using optimized kernel parameters. IEEE Geoscience and
Remote Sensing Letters, 8, 6 2011), 1060-1064.
[39] Iordache, M.-D., Bioucas-Dias, J. M. and Plaza, A. Total variation
spatial regularization for sparse hyperspectral unmixing. IEEE [59] Heinz, D. C. Fully constrained least squares linear spectral mixture
Transactions on Geoscience and Remote Sensing, 50, 11 2012), 4484- analysis method for material quantification in hyperspectral imagery.
4502. IEEE Transactions on Geoscience and Remote Sensing, 39, 3 2001),
529-545.
[40] Du, P., Xue, Z., Li, J. and Plaza, A. Learning discriminative sparse
representations for hyperspectral image classification. IEEE Journal of [60] Heylen, R., Scheunders, P., Rangarajan, A. and Gader, P. Nonlinear
Selected Topics in Signal Processing, 9, 6 2015), 1089-1104. unmixing by using different metrics in a linear unmixing chain. IEEE
Journal of Selected Topics in Applied Earth Observations and Remote
[41] Charles, A. S., Olshausen, B. A. and Rozell, C. J. Learning sparse codes Sensing, 8, 6 2015), 2655-2664.
for hyperspectral imagery. IEEE Journal of Selected Topics in Signal
Processing, 5, 5 2011), 963. [61] Dópido, I., Li, J., Gamba, P. and Plaza, A. A new hybrid strategy
combining semisupervised classification and unmixing of hyperspectral
[42] Yang, H., Du, Q. and Ma, B. Decision fusion on supervised and data. IEEE Journal of Selected Topics in Applied Earth Observations
unsupervised classifiers for hyperspectral imagery. IEEE Geoscience and Remote Sensing, 7, 8 2014), 3619-3629.
and Remote Sensing Letters, 7, 4 2010), 875-879.
[62] Cheng, Q., Varshney, P. K. and Arora, M. K. Logistic regression for
[43] Tarabalka, Y., Benediktsson, J. A. and Chanussot, J. Spectral–spatial
feature selection and soft classification of remote sensing data. IEEE
classification of hyperspectral imagery based on partitional clustering Geoscience and Remote Sensing Letters, 3, 4 2006), 491-494.
techniques. IEEE Transactions on Geoscience and Remote Sensing, 47,
8 2009), 2973-2987. [63] Persello, C. and Bruzzone, L. Active learning for domain adaptation in
the supervised classification of remote sensing images. IEEE

61 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 9, No. 8, 2018

Transactions on Geoscience and Remote Sensing, 50, 11 2012), 4468- [72] Li, W., Wu, G. and Du, Q. Transferred Deep Learning for Anomaly
4483. Detection in Hyperspectral Imagery. IEEE Geosci. Remote Sensing
[64] Manolakis, D., Marden, D. and Shaw, G. A. Hyperspectral image Lett., 14, 5 2017), 597-601.
processing for automatic target detection applications. Lincoln [73] Sankaran, S., Mishra, A., Ehsani, R. and Davis, C. A review of advanced
laboratory journal, 14, 1 2003), 79-116. techniques for detecting plant diseases. Computers and electronics in
[65] Chang, C.-I. and Chiang, S.-S. Anomaly detection and classification for agriculture, 72, 1 2010), 1-13.
hyperspectral imagery. IEEE Transactions on Geoscience and Remote [74] Duvvuru, R., Rao, G. N., Bendalam, S., Gemechu, R., Chundi, A.,
Sensing, 40, 6 2002), 1314-1325. Naidu, D. S., Rao, C. R. and Rao, P. J. A Case Study On Child Labour-
[66] Villa, A., Chanussot, J., Benediktsson, J. A., Jutten, C. and Dambreville, GIS Approach.
R. Unsupervised methods for the classification of hyperspectral images [75] Mendoza, F., Lu, R., Ariana, D., Cen, H. and Bailey, B. Integrated
with low spatial resolution. Pattern Recognition, 46, 6 2013), 1556- spectral and image analysis of hyperspectral scattering data for
1568. prediction of apple fruit firmness and soluble solids content. Postharvest
[67] Balas, C., Epitropou, G. and Pappas, C. Multi/hyper-spectral imaging. Biology and Technology, 62, 2 2011), 149-160.
Handbook of Biomedical Optics2011), 131-164. [76] Huang, M. and Lu, R. Apple mealiness detection using hyperspectral
[68] Zhao, W., Guo, Z., Yue, J., Zhang, X. and Luo, L. On combining scattering technique. Postharvest Biology and Technology, 58, 3 2010),
multiscale deep learning features for the classification of hyperspectral 168-175.
remote sensing imagery. International journal of Remote sensing, 36, 13 [77] Ariana, D. P. and Lu, R. Evaluation of internal defect and surface color
2015), 3368-3379. of whole pickles using hyperspectral imaging. Journal of Food
[69] Hu, W., Huang, Y., Wei, L., Zhang, F. and Li, H. Deep convolutional Engineering, 96, 4 2010), 583-590.
neural networks for hyperspectral image classification. Journal of [78] Liu, L., Ngadi, M., Prasher, S. and Gariépy, C. Objective determination
Sensors, 20152015). of pork marbling scores using the wide line detector. Journal of Food
[70] Zhu, X. X., Tuia, D., Mou, L., Xia, G.-S., Zhang, L., Xu, F. and Engineering, 110, 3 2012), 497-504.
Fraundorfer, F. Deep learning in remote sensing: a comprehensive [79] Arngren, M., Hansen, P. W., Eriksen, B., Larsen, J. and Larsen, R.
review and list of resources. IEEE Geoscience and remote sensing Analysis of pregerminated barley using hyperspectral image analysis.
magazine, 5, 4 2017), 8-36. Journal of agricultural and food chemistry, 59, 21 2011), 11385-11394.
[71] Chen, Y., Zhao, X. and Jia, X. Spectral–spatial classification of [80] Khan, M. J., Khan, H. S., Yousaf, A., Khurshid, K. and Abbas, A.
hyperspectral data based on deep belief network. IEEE Journal of Modern trends in hyperspectral image analysis: a review. IEEE Access,
Selected Topics in Applied Earth Observations and Remote Sensing, 8, 62018), 14118-14129.
6 2015), 2381-2392.

62 | P a g e
www.ijacsa.thesai.org
View publication stats