Microchemical Journal 193 (2023) 109129

Contents lists available at ScienceDirect

Microchemical Journal
journal homepage: www.elsevier.com/locate/microc

An overview of pre-processing methods available for hyperspectral

imaging applications
D. Cozzolino a, *, P.J. Williams b, L.C. Hoffman a
a
The University of Queensland, Centre for Nutrition and Food Sciences, Queensland Alliance for Agriculture and Food Innovation, St. Lucia, Brisbane, QLD 4072,
Australia
b
Department of Food Science, Stellenbosch University, Private Bag X1, Matieland (Stellenbosch) 7602, South Africa

A R T I C L E I N F O A B S T R A C T

Keywords: Hyperspectral imaging techniques have emerged as a powerful tool, combining the benefits of vibrational
Hyperspectral imaging spectroscopy and imaging into a single system. By capturing both spatial and spectral information, hyperspectral
Preprocessing methods imaging offers valuable insights into the characteristics of a sample. However, prior to the application of
Spectral data analysis
hyperspectral imaging techniques, it is crucial to perform pre-processing on the acquired images. The integration
Data preprocessing
Data transformation
of vibrational spectroscopy and imaging in hyperspectral imaging enables researchers to obtain detailed infor­
Noise reduction mation about the chemical composition and spatial distribution of samples. With this technology, researchers
have delved into various applications, ranging from pharmaceutical analysis, food, and agricultural assessment
to environmental monitoring. However, to ensure the accuracy and reliability of the results, appropriate pre-
processing techniques are essential. Pre-processing methods play a vital role in reducing or eliminating in­
terferences that may arise during image acquisition and subsequent analysis. These interferences could be
attributed to various factors, such as noise, uneven illumination, or unwanted artifacts. By applying suitable pre-
processing techniques, researchers can enhance the quality of hyperspectral images, ensuring more accurate and
reliable data for further analysis. This review aims to provide an overview of the pre-processing techniques
employed in the analysis of hyperspectral images.

1. Introduction the main benefit of this system is that it provides both spatial and
spectral information from the sample [1–4].
Vibrational spectroscopy (e.g. NIR, MIR and Raman) was first During the implementation and utilization of this technique, a
embraced by the agricultural and food sectors as analytical techniques to hyperspectral image is collected where a hypercube (Fig. 1), a three-
assess and evaluate the proximate composition of foods and raw in­ dimensional dataset consisting of one spectral and two spatial di­
gredients [1–4]. Both NIR and MIR spectroscopy techniques are used mensions, is constructed [1–4]. Hyperspectral imaging can also be
during the routine analysis of different types of samples in the labora­ achieved by combining the visible (VIS) and NIR wavelength ranges, and
tory, through processing and manufacturing (e.g. on–, in- and at -line), by combining Raman spectroscopy and microscopy [1–4]. Overall,
as quality control methods and for on farm applications [1–4]. hyperspectral imaging has been widely applied to quantify different
The choice between NIR and MIR spectroscopy depends on the attributes in a wide range of samples (e.g. foods, biological or envi­
purpose of the specific application. The utilization of classical NIR and ronmental samples, among others) [1–4].
MIR spectroscopy provide point based spectral information, where no The utilization of hyperspectral techniques usually acquire data in
spatial information is collected during the analysis [1–4]. Hyperspectral narrow wavelength bands [1–4]. The so-called “hyper” in hyperspectral
imaging, first used for remote sensing [5] combines digital imaging and refers to the large number of wavelength bands collected during image
spectroscopy. It is considered as the blend between the main charac­ acquisition [1–12]. The distinctive characteristic of hyperspectral sen­
teristics of point based spectral and spatial information into one system sors is that the wavelength range utilised is achieved using a detector
and is well-suited for food quality assessment and safety inspection using operating in narrow and contiguous wavelength bands [1–4]. Therefore,
an array of modalities encompassing vibrational spectroscopy [6]. Thus, hyperspectral images are highly correlated between adjacent bands

* Corresponding author.
E-mail address: d.cozzolino@uq.edu.au (D. Cozzolino).

https://doi.org/10.1016/j.microc.2023.109129
Received 1 February 2023; Received in revised form 3 July 2023; Accepted 23 July 2023
Available online 25 July 2023
0026-265X/© 2023 Elsevier B.V. All rights reserved.
D. Cozzolino et al. Microchemical Journal 193 (2023) 109129

oscillate with time. As with many applications of vibrational spectros­

copy, the surface or geometry of the sample, as well as its chemical
composition and physical properties will cause the photons to behave in
a different way in different parts of the sample [1–12]. Consequently,
the amount of information (number of reflected photons) and the re­
sidual quantity of energy created is not only influenced by the relevant
chemical information, but also by the physical characteristics of the
sample surface [1–12]. When the photons arrive at the camera (in
reflectance mode), they are detected by sensors that are also subjected to
instrumental noise. Moreover, different sensors have distinct sensitiv­
ities to the photons, making the spectral signature not only dependent on
the arriving photons but also on the quality of the instrument and
camera used [1–12].
As described, the analysis hyperspectral images are influenced by the
morphological features / external attributes as well as the spectral fea­
tures of the internal component directly associated with the sample
[1–12]. In practice, because of the high correlation of information be­
tween the bands of the hyperspectral images, different factors can
interfere [1–12]. Most of the hyperspectral systems are based on diode
array detectors or tuneable filters [1–12]. Issues related with the mal­
function of one of the diodes in the detector array may lead to dead
pixels (e.g. missing or zero values). In addition, unexpected spectral
readings (extreme values) or spiked points at one specific wavelength
can also occur [1–12].
The geometry of the sample can also play an important role and have
Fig. 1. Depiction of the differences between a digital image of maize kernels a great influence on the image collected. This can include changes in the
(top left) and a near infrared (NIR) hyperspectral image (hypercube) of the surface of the samples, surface evenness and roughness of the sample.
same sample set (top right). The hypercube is composed of an array of images at However, the images generated usually have a square shape and are
wavelengths (λ) within the NIR region each with an × and y coordinate, thus composed of both the external background (sample stage/conveyor belt)
providing spectral and spatial information. This permits one to view images at a and the interesting part of the sample (region of interest) [1–12].
wavelength of interest (bottom right) and to obtain a spectrum of a pixel within Like other vibrational spectroscopic techniques, different artifacts
a sample (bottom left).
can also influence the quality of the images [13–17]. These artifacts are
characteristic to the vibrational spectroscopy method utilised as well as
[1–12]. Like conventional spectral data, hyperspectral imaging (HSI) its performance when the sample is irradiated (e.g. light scattering), or
data necessitates a preprocessing step to address potential issues and they can be caused by the influence of the sample itself when irradiated
anomalies that may arise during image acquisition [13]. These issues (e.g. the fluorescence background in Raman spectroscopy) [1–12].
have the potential to hinder model development and yield misleading
results. Therefore, it is essential to eliminate these challenges and ensure 3. Pre-processing methods
the availability of high-quality data for subsequent analysis [13]. Pre­
processing plays a critical role in mitigating these issues and preparing Several pre-processing methods can be applied to hyperspectral
the HSI data for accurate and reliable analysis. By applying appropriate images, after collecting or during processing [15]. Some the pre-
preprocessing techniques, the data can be enhanced, noise can be processing methods used can deal with issues derived from and associ­
reduced, and artifacts can be minimized, leading to improved data ated with image segmentation, image correction and image space-
quality and more meaningful interpretation [13–16]. Consequently, spectral dimensional corrections [15]. All these methods have the
preprocessing acts as a crucial foundation for successful analysis, objective to improve the quality and interpretability of the images,
enabling researchers to derive valuable insights from hyperspectral thereby improving the classification accuracy of hyperspectral images
images. This review provides an overview of the prevalent pre- [15].
processing techniques used in the analysis and of hyperspectral images.
3.1. Dimension reduction
2. The need for pre-processing
A commonly used pre-processing method is the so-called dimension
There are different issues or anomalies that can occur during the feature reduction [13–14]. In this method, a high-dimensional space is
collection of a hyperspectral image [1–10]. Therefore, prior to the uti­ almost empty and multivariate data can be represented in a lower
lization of any data analytical technique or chemometrics, some level of dimensional space [13–14]. Thus, the undesirable effects of high-
pre-processing is required [1–12]. The sources of anomalies can be dimensional geometric characteristics and the “curse of dimension­
associated with the sample, the instrument, and the resulting image ality” are reduced. This fact has led to a chain of research on dimension
[1–12]. (feature) reduction [13–14].
The light source used during the analysis can influence the amount of
information obtained [1–10]. It has been reported that the relevant in­ 3.2. Resolution enhancement
formation obtained during the analysis is associated with the quantity of
photons emitted to and received from the sample surface [1–12]. Recently, the resolution enhancement of a hyperspectral image has
Depending on the application, the amount of light is affected by changes attracted increasing attention [13; 23]. Resolution enhancement tech­
or fluctuations. For example, in remote sensing applications these fluc­ niques for hyperspectral image spectroscopy can be broadly categorized
tuations can be associated with the sun [14–17], while the lamp or en­ into four classes. They are hyperspectral super-resolution (multi image
ergy sources (halogen lamp) are the main source in bench-top or single-image), subpixel mapping (or super resolution mapping),
applications [1–12]. In both applications, the energy emitted might hyperspectral pan sharpening, and hyperspectral and multispectral data

2
D. Cozzolino et al. Microchemical Journal 193 (2023) 109129

fusion [23]. Hyperspectral super resolution is an extension of ordinary drawbacks of MSC is that it depends on the correct selection of a
super-resolution in computer vision to hyperspectral image analysis, reference spectrum [26]. This is difficult to achieve in hyperspectral
where a high-resolution hyperspectral image is reconstructed from images, since the samples tend to be a mixture of different compounds
multiple low resolution hyperspectral images (or a single hyperspectral with different spectra [26]. The use of derivatives is also a popular pre-
image) acquired by a single sensor [23]. Subpixel mapping is a resolu­ processing method. The Saviztky-Golay methodology is commonly used
tion enhancement technique processed at a classification level using a to calculate the derivative of the spectrum. The sub-window of points
single hyperspectral image as input [23]. Hyperspectral pan sharpening chosen is first fitted to a polynomial degree and then the derivative is
and hyperspectral and multispectral fusion are multisensor super reso­ calculated. Derivatives (first and second) are usually the most popular in
lution techniques, in which a hyperspectral image is fused with an many spectroscopy applications. The first derivative removes the addi­
auxiliary higher-resolution data source (panchromatic or multispectral tive scattering, while the second derivative removes multiplicative
images) taken over the same area at the same time (or a similar period) scattering [15,27]. Another consequence of using derivatives is that they
to create a high-resolution hyperspectral image analysis [13–15]. have the capability of emphasizing minor spectral differences [15,27].
However, special care must be taken in the choice of the derivative
3.3. Geometric correction degree and in the sub-window size, since a high derivative degree with a
small window size can create a high amount of noise, while large sub-
Another pre-processing method is the so-called geometric correction windows can eliminate informative parts of the spectra. Overall, pre-
[11–18]. This is a transformation used in hyperspectral image analysis processing methods have one important commonality: they can elimi­
that is applied to each individual channel in such a way that a distorted nate physical artifacts reflected in the signal. For example, this can be
image is translocated to a standard reference axis (e.g. projected to co­ observed when rough samples are measured, or they have a nonplanar
ordinates in maps) [15]. Geometric corrections are commonly utilised in shape (high scattering). Fig. 2 shows the effect of various pre-processing
remote sensing applications [18–20] and is applied when a reference techniques on the NIR hyperspectral images and extracted spectra of
map of the scanned land is available [18–20]. The most popular geo­ fungi, adapted from reference [28].
metric correction method is based on ground control points [18–20].
These refer to points that are common to the distorted image and the 4. Final considerations
reference map [18–20], and are considered permanent elements such as
road intersections or airport runways. During this process a point is It is worth noting that numerous commercial software packages
identified and selected at the ground level (e.g. control point), where an dedicated to hyperspectral image analysis include a range of pre-
interpolation is then performed [18–20]. processing algorithms. However, the effective utilization of a specific
In addition, resampling methods such as nearest neighbour, bilinear pre-processing method necessitates iterative evaluation after the appli­
interpolation, and the bicubic interpolations are also combined [18–23]. cation of other processing algorithms. As highlighted by previous
These resampling methods can be used to assign the nearest pixel value research, the selection and sequencing of pre-processing steps should be
to the corrected pixel (nearest neighbour); bicubic spectral pre- carefully tailored to the specific application, image type (e.g., NIR,
processing is defined as the collection of mathematical operations that Raman), and sample characteristics, among other relevant factors.
help to minimize or eliminate the influence of undesirable phenomena While the outcome or output can be viewed as a trial-and-error process,
directly affecting the spectra. These issues can be associated with light prioritizing simplicity is crucial. Therefore, it is important that a chosen
scattering, particle size effects, surface roughness and detector artifacts pre-processing method is easily interpretable. Occasionally the back­
[18–23]. In addition, other common pre-processing methods used in ground is easily removed from the raw data and then the remaining data
vibrational spectroscopy are also applied to hyperspectral images. These in the region of interest (ROI) are pre-processed; at times spectral pre-
methods are used to enhance the spectral signal and to improve the processing (e.g. SNV) is needed to remove the background while
visualization of the surface information. Some of these pre-processing another spectral pre-processing technique (S-G) is required for the
methods are highlighted below. analysis of the sample. Pre-processing techniques in hyperspectral im­
aging play a crucial role in modifying the spatial and spectral informa­
3.4. Spectral correction tion of the image. However, it is important to note that improper
application of pre-processing methods can lead to the removal of
The so-called instrumental noise can be removed using smoothing informative components or introduce artifacts. Selecting the appropriate
techniques. One of the most common smoothing techniques is the use of pre-processing method is vital to avoid the loss of relevant information.
the Saviztky-Golay method [24]. This method is based on the selection In this regard, the simplicity of a pre-processing methodology is desir­
of a sub-windows around a specific point and calculating its projection able if it can achieve the desired results effectively. Spatial and spectral
onto a polynomial fitting of the points of the sub-window. This tech­ corrections in pre-processing are interconnected. For instance, smooth­
nique is simple to implement. However, special care must be taken in the ing the spectra not only eliminates spectral noise but also smooths the
selection of the spectral sub-window, since large sub-windows will corresponding image data. It is important to acknowledge that any pre-
eliminate informative peaks, while small windows might generate more processing method employed will inevitably lead to some loss of infor­
noise [24]. Another pre-processing method used is to reduce or remove mation. In many cases, it is the responsibility of the user to determine
scattering. Scattering is observed as a drift in the baseline of the spectra what should be treated as noise and retain only the analytically relevant
that can be additive or multiplicative, depending on the nature of the data.
sample and the physical interaction of the sample with the light. Two of
the most common methods are standard normal variate (SNV) [25] and 5. Conclusions
multiplicative scatter correction (MSC) [26]. In SNV the mean of the
spectra is subtracted from the original data matrix, and then divided by Pre-processing plays a crucial role in the application of hyperspectral
the standard deviation thereof. This method removes additive scattering imaging as an analytical tool, particularly in fields such as food and
without changing the shape of the original spectrum. However, it cannot biological applications. These methods are essential for mitigating in­
handle multiplicative scattering effects [26]. A method often used to terferences related to sample darkness or background issues, ensuring
remove multiplicative scatter is MSC. This technique estimates the the reliability of the data for subsequent analysis. Moreover, pre-
spectrum of pixels against a reference spectrum, where the corrected processing is indispensable in the development of accurate calibration
spectrum is the subtraction of the offset from the original spectrum models and the establishment of efficient applications in hyperspectral
divided by the slope [26]. Different authors consider that the main imaging. However, it is important for users to exercise discretion in

3
D. Cozzolino et al. Microchemical Journal 193 (2023) 109129

Fig. 2. NIR hyperspectral image of Fusarium graminearum

M14-55 fungi in a glass petri dish on potato dextrose agar at
1942 nm. a) Top left is the unprocessed image and the
associated spectra (top right) indicated by the yellow square;
bottom right is the SNV corrected image with the corre­
sponding spectra; b) Top left is the Saviztky-Golay (1st de­
rivative, 2nd polynomial, 15 points) corrected image and the
corresponding spectra (top right) and bottom left is the
Saviztky-Golay (2nd derivative, 3rd polynomial, 15 points)
corrected image and the associated spectra. (For interpre­
tation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)

4
D. Cozzolino et al. Microchemical Journal 193 (2023) 109129

selecting the data to be eliminated or retained during the pre-processing [10] A. Baiano, Applications of hyperspectral imaging for quality assessment of liquid
based and semi-liquid food products: a review, J. Food Eng. 214 (2017) 10–15.
stage. By understanding the underlying principles and considering the
[11] D. Saha, A. Manickavasagan, Machine learning techniques for analysis of
nature of the measured matrix, users can make informed decisions hyperspectral images to determine quality of food products: a review, Curr. Res.
during pre-processing to enhance the overall quality and interpretability Food Sci. 4 (2021) 28–44.
of hyperspectral data. [12] J.M. Amigo, I. Martí, A. Gowen, in: Hyperspectral Imaging and Chemometrics: a
Perfect Combination for the Analysis of Food Structure, Composition and Quality,
Elsevier, 28, 2013, pp. 343–370.
CRediT authorship contribution statement [13] J. Burger, A. Gowen, Data handling in hyperspectral image analysis, Chemom.
Intel. Lab. Syst. 108 (1) (2011) 13–22, https://doi.org/10.1016/J.
CHEMOLAB.2011.04.001.
D. Cozzolino: Conceptualization, Writing – original draft, Writing – [14] J.L. Xu, C. Esquerre, D.W. Sun, Methods for performing dimensionality reduction in
review & editing. P.J. Williams: Writing – original draft, Writing – re­ hyperspectral image classification, Https://Doi.Org/10.1177/0967033518756175
26 (1) (2018) 61–75, https://doi.org/10.1177/0967033518756175.
view & editing. L.C. Hoffman: Writing – original draft, Writing – review
[15] M. Vidal, J.M. Amigo, Pre-processing of hyperspectral 252 images. Essential steps
& editing. before image analysis, Chemometrics and Intelligent Laboratory Systems 117
(2012) 138-148, https:// doi.org/10.1016/j.chemolab.2012.05.009.
[16] Amigo, J.M., Santos, C. (2019). Chapter 2.1 - Preprocessing of hyperspectral and
Declaration of Competing Interest multispectral images, Editor(s): José Manuel Amigo, Data Handling in Science and
Technology, Elsevier, Volume 32, 37-53.
[17] P.K. Varshney, M.K. Arora (Eds.), Advanced Image Processing Techniques for
The authors declare that they have no known competing financial Remotely Sensed Hyperspectral Data, Springer Berlin Heidelberg, Berlin,
interests or personal relationships that could have appeared to influence Heidelberg, 2004.
the work reported in this paper. [18] T. Toutin, Review article: geometric processing of remote sensing images: models,
algorithms and methods, Int. J. Remote Sens. 25 (2004) 1893–1924, https://doi.
org/10.1080/0143116031000101611.
Data availability [19] N.G. Kardoulas, A.C. Bird, A.I. Lawan, Geometric correction of SPOT and Landsat
Imagery : a comparison of M a p and GPS-derived control points, Photogramm.
Eng. Remote Sens. 62 (1996) 1173–1177.
No data was used for the research described in the article. [20] A.J. De Leeuw, L.M.M. Veugen, H.T.C. Van Stokkom, Geometric correction of
remotely sensed imagery using ground control points and orthogonal polynomials,
References Int. J. Remote Sens. 9 (2007) 1751–1759, https://doi.org/10.1080/
01431168808954975.
[21] L. Zhang, M.J. Henson, A practical algorithm to remove cosmic spikes in Raman
[1] J.M. Amigo, H. Babamoradi, S. Elcoroaristizabal, Hyperspectral image analysis. A imaging data for pharmaceutical applications, Appl. Spectrosc. 61 (2007)
tutorial, Anal. Chim. Acta 896 (2015) 34–51, https://doi.org/10.1016/j. 1015–1020, https://doi.org/10.1366/000370207781745847.
aca.2015.09.030. [22] Z. Nenadic, J.W. Burdick, Spike detection using the continuous wavelet transform,
[2] J. Burger, P. Geladi, Hyperspectral NIR image regression part I: calibration and IEEE Trans. Biomed. Eng. 52 (2005) 74–87, https://doi.org/10.1109/TBME.
correction, J. Chemom. 19 (2005) 355–363, https://doi.org/10.1002/cem.986. 2004.839800.
[3] J.M. Amigo, Practical issues of hyperspectral imaging analysis of solid dosage [23] L. Duponchel, P. Milanfar, C. Ruckebusch, J.-P. Huvenne, Super-resolution and
forms, Anal. Bioanal. Chem. 398 (2010) 93–109, https://doi.org/10.1007/s00216 Raman chemical imaging: from multiple low resolution images to a high resolution
218 010-3828-z. image, Anal. Chim. Acta 607 (2) (2008) 168–175.
[4] D. Cozzolino, J.J. Roberts, Applications and developments on the use of vibrational [24] A. Saviztky, M.J.E. Golay, Smoothing and differentiation of data by simplified least
spectroscopy imaging for the analysis, monitoring and characterisation of crops squares procedures, Anal. Chem. 36 (1964) 1627–1639, https://doi.org/10.1021/
and plants, Molecules 21 (2016) 755–763. ac60214a047.
[5] A.F.H. Goetz, Three decades of hyperspectral remote sensing of the Earth: a [25] R.J. Barnes, M.S. Dhanoa, S.J. Lister, Standard Normal Variate Transformation and
personal view. remote sensing of environment 113, Supplement 1 (2009) S5–S16. De-trending of Near-Infrared Diffuse Reflectance Spectra, Appl. Spectrosc. 43 (5)
[6] Y.Z. Feng, G. Elmasry, D.W. Sun, A.G. Scannell, D. Walsh, N. Morcy, Near-infrared (1989) 772–777.
hyperspectral imaging and partial least squares regression for rapid and reagentless [26] P. Geladi, D. MacDougall, H. Martens, Linearization and scatter-correction for
determination of Enterobacteriaceae on chicken fillets, Food Chem. 138 (2–3) near-infrared reflectance spectra of meat, Appl. Spectrosc. 39 (3) (1985) 491–500.
(2013 Jun 1) 1829–1836. [27] A. Rinnan, F. van den Berg, S.B. Engelsen, Review of the most common pre-
[7] M.L. Amodio, M.A. Chaudhry, G. Colelli, Spectral and hyperspectral technologies processing techniques for near-infrared spectra, Trac-Trend Anal Chem 28 (10)
as an additional tool to increase information on quality and origin of horticultural (2009) 1201–1222, https://doi.org/10.1016/j.trac.2009.07.007.
crops, Agronomy 10 (7) (2020), https://doi.org/10.3390/agronomy10010007. [28] P.J. Williams, C. Bezuidenhout, L.J. Rose, Differentiation of Maize Ear Rot
[8] Y.Z. Feng, D.W. Sun, Application of hyperspectral imaging in food safety inspection Pathogens, on Growth Media, with Near Infrared Hyperspectral Imaging, Food
and control: a review, Crit. Rev. Food Sci. Nutr. 52 (2012) 1039–1058. Anal. Methods 12 (7) (2019) 1556–1570, https://doi.org/10.1007/s12161-019-
[9] M. Manley, Near-infrared spectroscopy and hyperspectral imaging: non-destructive 01490-y.
analysis of biological materials, Chem. Soc. Rev. 43 (24) (2014) 8200–8214.