b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.elsevier.com/locate/issn/15375110

Research Paper

Evaluation of the performance of portable

visible-infrared instruments for the prediction of
soil properties

Jos
e M. Soriano-Disla a,b,*, Leslie J. Janik a,c, Danielle J. Allen c,

Michael J. McLaughlin a,c
a
CSIRO Environmental Contaminant Mitigation and Technologies Program, CSIRO Land and Water, Waite Campus,
Waite Road, Urrbrae, 5064, South Australia, Australia
b
Sustainable Use, Management and Reclamation of Soil and Water Research Group (GARSA), Department of
Agricultural Science and Technology, Universidad Polit
ecnica de Cartagena, Paseo Alfonso XIII 48, 30203, Cartagena,
Spain
c
Soil Science, School of Agriculture, Food and Wine, University of Adelaide, Waite Campus, Waite Road, Urrbrae,
5064, South Australia, Australia

article info
Good soil management requires large amounts of soil data which are expensive to provide
Article history: using traditional laboratory methods. Soil infrared spectroscopy including portable/mini-
Received 9 December 2016 aturized visible-infrared spectrometers offers a cost-effective solution. There is a need to
Received in revised form test and compare the performance of portable/miniaturized mid-infrared (MIR) and visible-
7 April 2017 near-infrared (vis-NIR) spectrometers for the prediction of soil properties across a range of
Accepted 20 June 2017 soils. For this assessment, 458 soil samples from Australia were scanned by four vis-NIR
and MIR portable/miniature spectrometers and partial least squares regressions (PLSR)
applied for the prediction of 17 properties in soils dried at 40
C and sieved to <2 mm. The
Keywords: performance of these instruments was tested and compared to a reference benchtop MIR/
Mid-infrared NIR instrument. Mid-infrared handheld instruments provided the best performance, the
Near-infrared vis-NIR instrument the next most successful, and the miniature NIR instrument with a
Partial least squares regression restricted spectral range (950e1650 nm) being less successful. When models using the
Portable same spectral range obtained by different instruments were compared, similar perfor-
Soil mance was achieved, thus the spectral quality provided by different instrumentation was
Spectroscopy not decisive in determining prediction accuracy. Many new portable infrared instruments
have restricted spectral ranges, thus a number of different spectral ranges in both the MIR
and vis-NIR were assessed to determine the optimal range for prediction of soil properties.
It was concluded that the range 1650e5000 nm would be ideal.
Crown Copyright © 2017 Published by Elsevier Ltd on behalf of IAgrE. All rights reserved.

* Corresponding author. Sustainable Use, Management and Reclamation of Soil and Water Research Group (GARSA), Department of

cnica de Cartagena, Paseo Alfonso XIII 48, 30203, Cartagena, Spain
Agricultural Science and Technology, Universidad Polite
E-mail address: martin.soriano@upct.es (J.M. Soriano-Disla).
http://dx.doi.org/10.1016/j.biosystemseng.2017.06.017
1537-5110/Crown Copyright © 2017 Published by Elsevier Ltd on behalf of IAgrE. All rights reserved.
 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6
1. Introduction
The need to increase food production globally over the next 30
years requires that we have better information on soil condi-
tion, in both developed and developing economies. Sustain-
able intensification of agricultural production requires that we
develop better management strategies for soils. In order to
support the development and application of optimum man-
agement strategies, there is a need to increase the spatial
density of soil analytical data (Nocita et al., 2015; Palm,
Sanchez, Ahamed, & Awiti, 2007; Viscarra Rossel et al., 2016).
This demand for large amounts of soil data is enhanced by the
fact that soils are highly heterogeneous and diverse tempo-
rally, spatially and to depth (Palm et al., 2007; Reimann, Birke,
Demetriades, Filzmoser, & O'Connor, 2014; Viscarra Rossel
et al., 2016), requiring the adoption of soil management de-
cisions for each specific situation.
Traditional laboratory methods of soil analysis are unable
to satisfy such demands because they are expensive and time-
consuming. Infrared spectroscopy, using the mid-infrared (MIR:
4000e400 cm1, 2500e25,000 nm), the near-infrared (NIR:
700e2500 nm, 14,286e4000 cm1) and/or the visible (vis:
400e700 nm, 14,286e25,000 cm1) regions, in combination
with multivariate regression methods, offers a practical
alternative (Nocita et al., 2015; Shepherd & Walsh, 2007;
Soriano-Disla, Janik, Viscarra Rossel, MacDonald, &
McLaughlin, 2014). Infrared soil analysis techniques,
described extensively in previous publications (e.g. Janik,
Merry, & Skjemstad, 1998; Stenberg, Viscarra Rossel,
Mouazen, & Wetterlind, 2010), are non-destructive and rela-
tively cost-effective, rapid, and precise. This allows the pre-
diction of many soil properties through the use of multivariate
regression. Thus, the technique is well suited to provide an
assessment of key soil attributes such as texture, mineralogy,
cation exchange capacity (CEC) and organic C (OC) and N
contents (Palm et al., 2007).
Recent improvements in instrumentation have seen the
field of soil spectroscopy to develop rapidly, opening up the
opportunity to take the technique from the laboratory to the
field using portable, handheld and now miniature spectrom-
eters (Alcala et al., 2013; Knadel, Stenberg, Deng, Thomsen, &
Greve, 2013; Kuang et al., 2012; Reeves, 2010; Yang & Mouazen,
2012). However, the performance of portable/miniature in-
struments can be compromised by technical limitations. As
discussed by Mouazen, Saeys, Xing, De Baerdemaeker, and
Ramon (2005) and Soriano-Disla et al. (2014), instrument per-
formance is dependent on technical specifications such as
type of energy source and detector, resolution, sampling ac-
cessories, instrument and energy intensity. Thus, for the
same set of soils and reference soil properties, different in-
struments could have different performance, which could be
detrimental for the further development of predictive regres-
sion models (Mouazen et al., 2005; Reeves, 2010).
Instrument performance is also related to the spectral
range used (Kuang et al., 2012; Mouazen et al., 2005; Reeves,
2010) where more accurate predictive models have generally
been obtained by using the MIR range as compared with NIR
(e.g. as summarised by Reeves, 2010; Soriano-Disla et al., 2014;
Viscarra Rossel, Walvoort, McBratney, Janik, & Skjemstad,
25
2006). This has been related to the fact that more information
is provided by MIR spectra. Conversely, the NIR peaks are
much less intense, being the result of overtones and combi-
nation bands of fundamental vibrations in the MIR region
(Nguyen, Janik, & Raupach, 1991; Stenberg et al., 2010).
However, there are some soil properties for which the vis-
NIR region seems to perform better than MIR (e.g. soil bio-
logical properties or cation exchange related properties;
Soriano-Disla et al., 2014). In addition, the MIR region is more
susceptible to variable water content in field samples (Kuang
et al., 2012; Reeves, 2010) and, in some cases, the absorbance
from some soil compounds (e.g. quartz) is so intense that it
overlaps other relevant soil information (Nguyen et al., 1991).
This can be detrimental for the accuracy of the predictions
(Rabenarivo et al., 2013).
In terms of the performance of portable/miniature in-
struments, there are very few studies which evaluate the
performance of portable MIR instruments (Dhawale et al.,
2015; Forrester et al., 2015; Ji et al., 2016). This is mainly
attributed to the fact that MIR instruments have been gener-
ally restricted to the laboratory until relatively recently
(Dhawale et al., 2015; Forrester et al., 2015; Reeves, McCarty, &
Hively, 2010). Comprehensive studies comparing the potential
of different portable infrared spectrometers covering vis-NIR
and MIR ranges for the prediction of soil properties in a
range of soils are lacking.
Hence, the main objective of this study was to provide an
assessment of the performance of portable and miniature MIR
and vis-NIR spectrometers to predict soil properties. The
assessment included an analysis of the impact of the quality of
the spectra on the performance of predictions and the optimal
region across MIR and vis/NIR ranges for such predictions.
2. Material and methods
2.1. Soil sampling
The samples used in this study were obtained from the
Australian CSIRO National Soil Archive (CNSA, http://www.clw.
csiro.au/aclep/archive/) and can be found in http://www.asris.
csiro.au/(APSRU reference sites). The final selection
comprised 80 soil profiles (n ¼ 458 samples) from South
Australia (66%) and New South Wales (34%) (Supplementary
Fig. 1). Soil samples were dried at 40
C and sieved <2 mm.
Samples were sourced from variable depths, most of the sam-
ples (n ¼ 315) represented the first 100 cm, the rest (n ¼ 143)
representing depths from 100 to 180 cm. Samples represented 9
soil orders which are commonly used for cropping in Australia,
mostly Calcarosols, Chromosols, Dermosols, Sodosols and
Vertosols. Minor contributions of Dermosols, Kandosols, Kur-
osols, Ferrosols and Tenosols were observed.
2.2. Soil laboratory analysis
The following properties (Table 1) were determined by the
methodology described in Rayment and Higginson (1992) and
Rayment and Lyons (2011): Exchangeable bases calcium (Ca2þ),
magnesium (Mg2þ), potassium (Kþ) and (Naþ), alcoholic 1M
ammonium chloride at pH 8.5, pre-treatment for soluble salts,
26 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6

Table 1 e Descriptive statisticsa for the soil properties.b

BD SAT DUL OC TC TN EC pH CEC Ca2þ Mg2þ Naþ Kþ ESP Sand Silt Clay
3 þ
g/cm mm/mm % dS/m cmol /kg %
N 458 458 458 308 457 457 455 439 455 455 455 455 455 455 451 451 451
Mean 1.5 0.42 0.30 0.45 1.63 0.08 0.35 7.9 22 11.5 6.6 2.6 1 9.5 46.7 12.7 40.5
Max 1.8 0.68 0.46 4.90 9.18 0.38 4.10 10 48 34.6 19.9 15.5 4.2 43.6 97 44 80.2
Min 0.8 0.31 0.07 0.02 0.04 0.00 0.03 4.4 1 0.6 0.2 0 0.1 0 9 0.8 2
SD 0.2 0.06 0.08 0.58 1.64 0.05 0.44 1.3 10 6.7 4.3 3.3 0.6 10.8 17.4 7.6 15.7
Var 0 0.00 0.01 0.33 2.68 0.00 0.20 1.6 109 44.4 18.5 10.9 0.3 117 303.7 57.4 246.6
Skw 0.6 0.5 0.5 3.2 1.4 1.3 4.2 0.7 0.1 1.0 0.5 1.5 1.4 1.2 0.5 1.1 0.2
Kurt 1.0 1.1 0.2 15.5 1.7 3.0 25.2 0.2 0.9 0.7 0.5 1.2 3.4 0.5 0.0 1.3 0.6
Med 1.5 0.42 0.31 0.24 1.01 0.07 0.20 8.1 22 9.9 6 0.8 0.9 4.2 44.4 11 42.5
Q1 1.4 0.38 0.25 0.11 0.44 0.04 0.10 7.2 13 6.8 3.2 0.2 0.5 1.3 35.5 8 28.3
Q3 1.6 0.45 0.36 0.53 2.26 0.10 0.40 8.9 30 14.9 9.7 4 1.3 15.6 56.6 16 52
a
Descriptive statistics: Number of samples used in the analysis (N), standard deviation (SD), variance (Var), skewness (Skw), kurtosis (Kurt),
median (Med), first quartile (Q1), third quartile (Q3).
b
Bulk density (BD), saturated moisture (SAT), drained upper limit (0.1 BAR) moisture (DUL), organic carbon (OC), total carbon (TC), total nitrogen
(TN), electrical conductivity (EC), cation exchange capacity (CEC), exchangeable calcium, magnesium, sodium and potassium (Ca2þ, Mg2þ, Kþ,
Naþ), and exchangeable sodium percentage (ESP).

and exchangeable bases by compulsive exchange, no pre-
treatment for soluble salts; cation exchange capacity, auto-
mated determination of ammonium and chloride ions, and
compulsive exchange, no pre-treatment for soluble salts;
electrical conductivity (EC), 1:5 soil/water extract; pH, 1:5 soil/
water suspension; organic carbon, Walkley and Black; bulk
density (BD) and drained upper limit Moisture (DUL), volu-
metric APSRU (Agricultural Production System Research Unit)
in situ methodology (Burk & Dalgliesh, 2013); saturated mois-
ture (SAT), calculated from BD (Burk & Dalgliesh, 2013); particle
size distribution (clay, silt and sand), hydrometer method (Gee
& Bauder, 1986); total carbon and nitrogen (TC and TN), Leco
analyser following the method proposed by Matejovic (1997).
2.3. Infrared spectra scanning and pre-processing
2.3.1. Instrumentation and scanning
Four portable MIR and vis-NIR instruments were tested and
compared to a reference laboratory benchtop MIR/NIR instru-
ment. For each sample, four replicate scans were recorded.
i. Reference laboratory spectrometer e Perkin Elmer
Frontier
(USA) were used in this study scanning in the frequency range
of 6000e650 cm1. The optical configuration of both in-
struments is identical (i.e. a Michelson interferometer, zinc
selenide beamsplitter and thermoelectrically cooled deuter-
ated triglycine sulphate, DTGS, detector). However, the ExoS-
can is a compact instrument where all the parts are contained
within the same module, while in the case of FlexScan, the
optical and electronic/battery components are split into two,
with the optical component held in the hand and the elec-
tronics/battery worn on the belt and linked by a communi-
cations cable. The instruments have a similar weight
(approximately 3.3 kg) and spectra were obtained through a
DRIFT accessory. Resolution, scanning time and frequency of
backgrounds were the same as that used for the Frontier
laboratory instrument, but a course-grained SiC reference disk
was used for the handheld instruments.
For both instruments, samples were placed in stainless
steel cups (9 mm diameter, 3 mm depth) and, with the in-
struments facing down, lifted by compressed air against the
optical aperture a tailor-made setup allowing a high repro-
ducibility in sample presentation.
iii. Spectra evolution (SM-3500)

A Fourier-Transform infrared (FT-IR) benchtop spectrom-
eter (Frontier; Perkin Elmer, USA), equipped with a deuterated
triglycine sulphate, DTGS, detector and extended range
KBr beam-splitter, was used in the frequency range
7800e400 cm1 range, with a resolution of 8 cm1, and scan-
ning time of 15 s. A fine-grained silicon carbide (SiC) reference
disc (assumed to have a reflectance R0 ¼ 1) was used as a
background, taken after a 1 h warm-up. Samples were pre-
sented manually to the diffuse reflectance FT-IR (DRIFT)
accessory using an in-house sample holder/slide.
ii. ExoScan™ and FlexScan™
Two handheld FT-IR instruments, the ExoScan (model
4100) and the FlexScan (model 4200) manufactured by Agilent
This instrument (SM-3500; Spectra Evolution, USA) has a
crossed Czerny-Turner configuration using ruled gratings as
the dispersive element. Spectra were obtained by contact of the
sample with a high energy contact probe with external illumi-
nation (Spectra Evolution, USA) coupled to the instrument with
a fixed fibre optic cable. The reflected energy enters the spec-
trometer and is collimated before being reflected off the grat-
ings and refocused onto three detectors (range 350e2500 nm): a
512-element Si array (vis-NIR; up to 1000 nm) and two Peltier
cooled Indium Gallium Arsenide (InGaAs) arrays of 256 ele-
ments each extending detection to 2500 nm. The resolution was
3.5 nm, 10 nm and 7 nm for the 350e1000 nm, 1000e1900 nm
and 2100e2500 nm spectral ranges, respectively. A total of 30
spectra were averaged, with a white reference (Spectralon™,
Spectra Evolution, SA) scanned every 30 min.
 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6
Samples were presented as proposed by Ben-Dor, Ong, and
Lau (2013). Briefly, soil samples were prepared on a Petri dish
of 10 mm depth and 63 mm diameter. The surface of the
samples was tamped using a glass surface, and then brought
to the contact probe by using a laboratory jack.
iv. NIRScan Nano (NIRScan)
A miniature NIRScan Nano (Texas Instruments, USA) was
used (weight ~ 100 g). With this instrument, samples are pre-
sented in close contact to a sapphire window and illuminated
with two integrated and angled tungsten infrared lamps. The
diffusively reflected light is gathered and focused through the
input slit (25 mm wide by 1.69 mm tall). The light then strikes a
reflective grating which, in combination with a focussing lens,
disperses the light into its constituent wavelengths. The
focussing lenses form an image of the slit at the Digital Light
Proccessing (DLP) micro-mirror device (DMD); (0.2-inch WVGA,
854  480 orthogonal pixel, NIR optimized). The energy reflected
by the DMD is directed through the collection optics to the
single pixel InGaAs detector (900e1700 nm, 10 nm resolution).
Thirty average scans were taken, scanning the white reference
(Spectralon™, Spectra Evolution, USA) every 30 m. Samples
contained in glass vials were scanned directly by placing the
vials on the top of the instrument.
2.3.2. Pre-processing of spectra
After close inspection of the spectra, the noisy edges of the
spectra were removed. In the case of the Frontier reference
instrument, only the MIR range (4000e450 cm1) was consid-
ered because the energy of the instrument is optimised for
that region and to compare with the other MIR instruments.
Different typical pre-processing approaches were tested, with
the optimum for each instrument applied as follows: 1)
Frontier: de-trending, 4000e450 cm1; 2) ExoScan and FlexS-
can: de-trending, 4000e750 cm1; This transformation
removed baseline offset and slope from each spectrum by
fitting a polynomial to each sample spectrum, then removing
the estimated baseline curvature; 3) SM-3500 and NIRScan:
SavitzkyeGolay smoothing with first derivative and second-
order polynomial and window size of 5 points (Savitzky &
Golay, 1964), 400e2450 nm. These pre-processing approaches
are commonly used for the correction of NIR, vis-NIR or MIR
spectra (e.g. Shepherd & Walsh, 2002; Shi et al., 2014; Soriano-
Disla et al., 2013; Stenberg et al., 2010).
All the transformations and model development (as shown
in Section 2.4) were carried out by Unscrambler X 10.3 (Camo,
Norway).
2.4. Development of multivariate models
2.4.1. Principal component analysis
A principal component analysis (PCA) was developed to
describe the variability of samples and detect potential outliers.
The 4000e450 cm1 spectral region of the Frontier spectra
following baseline offset correction was used for this analysis,
which was developed by leave-one-out cross-validation. After
the analysis, the spectra are represented by multidimensional
scores maps related to corresponding loading vectors.
27
2.4.2. Evaluating the performance of the instruments:
development of partial least squares regression
A well-known multivariate algorithm, partial least squares
regressions (PLSR; Martens & Naes, 1989), was used for the
development of predictive models. For each instrument and
property, the following calibration approach was followed:
25% of the samples were randomly selected and set aside
(prediction set). The remaining 75% were used for model
development by leave-one-out cross-validation and selection
of the optimum number of factors. These models were then
tested on the prediction set.
The performance of the models was evaluated in terms of
the root means squares error of the prediction (RMSEP), coeffi-
cient of determination (R2), and the ratio of the standard devi-
ation of the reference values to the RMSEP of the prediction
(RPD) (Williams, 1987). In general, models with high values of R2
and RPD are preferred and indicative of prediction performance
(conversely, models with RPD and R2 values below 1.5 and 0.5
considered poor or unreliable; Sudduth & Hummel, 1996). The
number of factors, nature of the property (i.e. distribution,
concentration range), type of response in the infrared, presence
of outliers, and ability for modelling low and high concentration
samples, were also considered in the assessment.
2.4.3. Testing the performance of restricted spectral ranges
Following the same methodology that was described above,
additional PLSR predictive models were developed but using
selected specific spectral ranges from the SM-3500 and Frontier
spectra. These represent reference spectrometers, charac-
terised by their high energy and high signal to noise ratio. This
allowed us to test the potential for the prediction of soil prop-
erties using restricted spectral ranges which could be poten-
tially included in future miniature instrument development.
The following spectral ranges were tested: 1) Using SM-3500
spectra: 1650e2450 nm (range in the NIR not covered by the
NIRScan; the end 2450e2500 nm was avoided), 400e700,
1650e2450 nm (previous range including the visible),
400e1650 nm (range already included in NIRScan but with the
addition of the visible); 2) Using Frontier spectra: 3000e5000 nm
(2000e3334 cm1; range likely to be included in new instru-
ment development), 2000e5000 nm (2000e5000 cm1; same
than previous but also covering the 2500e3000 nm region and
including part of the NIR), 2500e5555 nm (4000e1800 cm1; MIR
restricted range but including more information in the low
spectral range) and 5522e11,136 nm (898e1811 cm1; range of a
prototype spectrometer tested by Dhawale et al., 2015). Models
were developed as previously described in Section 2.4.2. For the
prediction of CEC, exchangeable Ca2þ, magnesium Mg2þ, OC,
pH, sand, TC and TN.
3. Results and discussion
3.1. Sample analytical characterisation
In agreement with Gobrecht, Roger, and Bellon-Maurel (2014),
and prior to model development, it is essential to evaluate the
distributional properties of soil variables (Table 1), which were
classified in three groups accordingly:
28 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6

1) Close to symmetrical distribution/wide concentration Frontier ExoScan FlexScan

range (pH, Ca2þ, CEC, ESP, Mg2þ, Naþ, particle size and TC):
This confirmed the wide diversity of the samples selected 1.5 Average

Absorbance
for this study which covered 9 soil orders, and included
both topsoils and subsoils. 1

0.5
The skewness of TC was higher than the other properties of
this category, which was attributed to the presence of a few 0
7400 6400 5400 4400 3400 2400 1400 400
samples with high concentrations. The distribution was more
Wavenumber (cm-1)
symmetrical when these samples (TC > 6.2%; most of them
corresponding with highly calcareous samples) were
Frontier ExoScan FlexScan
removed.
0.39
Standard deviation

Absorbance
2) Close to symmetrical distribution/narrow concentration
0.29
range and low concentrations (BD, DUL, SAT, TN): limited
variability and having most of the samples with concen- 0.19
trations close to 0 could be an issue for predictive model
development. When modelling variables with a narrow 0.09
concentration range, especially if the concentration is also 7400 6400 5400 4400 3400 2400 1400 400
Wavenumber (cm-1)
low, a high relative contribution of the reference analytical
error is expected, which is detrimental for model perfor-
Fig. 1 e Average and standard deviation spectra of the soils
mance. This needs to be considered when evaluating the
in this study scanned by Frontier, ExoScan and FlexScan
performance of infrared models and instruments.
(spectra displayed following baseline transformation).
Highly noisy spectra regions from the ExoScan/FlexScan
The distribution of TN was slightly asymmetrical,
were excluded.
becoming almost symmetrical when only one sample with
concentration higher than the rest (TN ¼ 0.38%) was removed.

similar. However, higher variation was observed in the

3) Skewed distribution/narrow concentration range (EC, Kþ,
FlexScan/ExoScan spectra, especially in the 3100e3700 cm1
OC): these properties were characterised by having a narrow
region. In addition, further variation was observed for the
concentration range and a few samples with high concen-
ExoScan spectra, probably as a result of the instrument
trations, which influenced the asymmetrical nature of their
being older as compared with the FlexScan and Frontier that
distributions. However, it seemed sufficient to remove few
were almost new. However, this was a consistent baseline
high concentration outliers (i.e. one sample for Kþ and OC,
and four for EC), to get almost symmetrical distributions.
NIRScan SM-3500

3.2. Spectra analysis

0.9 Average
Absorbance

Soil composition, main chemistry and variability can be 0.6

interpreted from the mean and standard deviation spectra
0.3
(Figs. 1 and 2). The MIR spectra were used to describe
composition and depicted a mixture of 1:1 and 2:1 (e.g. 0
kaolinite and smectite clay minerals), with a medium-low 2350 1850 1350 850 350

content of calcium carbonate, low concentration of OC, Wavelength (nm)

and soils being non sandy. Main variations, as represented

by the standard deviation spectrum, seemed to be due to the NIRScan SM-3500
carbonate content (about 2500e2530 cm1), particle size
(peaks about 1700e2000 cm1 and 1000e1300 cm1) and clay 0.2 Standard deviation
Absorbance

mineralogy (peaks at about 3695e3620 cm1 and

3450e3200 cm1). 0.1
For the common range scanned by ExoScan/FlexScan
instruments, spectra were almost identical and similar to
0
the reference laboratory instrument (Frontier), apart for 2350 1850 1350 850 350
variations in the low range attributed to the use of different Wavelength (nm)
backgrounds (dark for ExoScan and FlexScan to make sure
that the detector was not overloaded after setting the gain Fig. 2 e Average and standard deviation spectra of all the
manually to the maximum, and light for Frontier which soils in this study scanned by SM-3500 and NIRScan
automatically adjusts the gain). Main sources of variation, as (untransformed spectra). Highly noisy spectral regions
represented by the standard deviation spectrum, were were excluded.
 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6
Scores
15
Calibration
10
Prediction
PC-2 (32 %)
5
0 predictions, getting the attributes analysed in the same
-5
-10
-10 -5 0 5 10
PC-1 (37 %)
Fig. 3 e Principal component analysis scores plot using
Frontier spectra (4000e450 cm¡1) following baseline
correction.
effect across wavenumbers, which was corrected by pre-
processing.
The mean spectra for SM-3500 represents a typical vis-NIR
soil spectra as described elsewhere (e.g. Ben-Dor & Banin,
1995; Stenberg et al., 2010), featuring three prominent peaks
at about 1450, 1950 nm 2230 nm, corresponding to OH bonds of
adsorbed water (1450 and 1950 nm) and cationic water in the
crystal lattice of clay minerals (2230 nm). Main variations were
found near these peaks. In the NIRScan spectral range, which
was mainly the result of second and third overtones of
fundamental vibrations, only the prominent water peak at
around 1450 nm could be visually identified. In terms of noise,
the NIRScan spectrum appeared to be clear, and substantially
similar to that provided by the SM-3500. By examining the
standard deviation spectrum of the NIRScan compared to that
calculated from SM-3500 spectra, the former appeared to be
less variable than the latter.
When MIR and vis-NIR ranges are compared, it becomes
evident that MIR spectra had more detail; peaks are more
intense and easier to interpret. In the low MIR range (i.e. below
2000 cm1), abundant spectral information becomes compli-
cated to resolve, frequently attributed to the strong response
of quartz, which is almost absent in the NIR.
Sample variability was further explored through the
development of a PCA (Fig. 3), which can be also useful for the
detection of potential compositional outliers. According to the
PCA scores plot, no clear compositional spectral outliers were
found. Samples with high values of score 1 were characterised
by having a strong spectral response from carbonates, with
some of the samples potentially having dolomite.
3.3. Development of partial least squares regression
models
3.3.1. Assessment of the models obtained
Partial least squares regression calibration models were
developed for properties using all instruments (Table 2). As
indicated previously, concentration outliers were not
29
considered in model development. It is important to remark
that the main objective of this study was to compare the
performance of instruments and, for that purpose, the per-
formance of models developed using a set of soils prepared
in the same way (i.e. soils dried and sieved) and routine al-
gorithms such as PLSR were compared. However, we
acknowledge that in order to get optimal results in infrared
laboratory and at the same time using a trusted reference
method is critical. Further improvement can be obtained by
fine grinding the samples.
According to the models obtained, and other consider-
15 ations such as the fundamental spectral response of the
technique to the tested variables and agreement from previ-
ous publications, we ranked the soil variables in three cate-
gories (high, medium and low) of the potential to be predicted
by infrared technology:
 High predictive performance (properties ranked by the
prediction RPD of the best performing models): TC, pH,
sand, clay, CEC, TN, silt, Mg2þ, Ca2þ, DUL and OC
 Medium predictive performance: Naþ and ESP
 Low predictive performance: EC, Kþ, BD, SAT
Prediction performance for Naþ and ESP was classified as
“medium” since models were found only after fitting a rela-
tively large number of PLSR factors, which is indicative of
different sources of variability responsible for the prediction
of this property. The models “struggled” to predict low con-
centration samples. Low predictive performance was found
for EC and Kþ which do not have clear and direct response in
the infrared spectrum. Where successful models have been
reported, these were due to fortuitous and specific correla-
tions with infrared responsive attributes. For SAT and BD, it is
hard to establish a relationship between a property that was
analysed in situ and the spectra of a dried and disturbed soil.
This is especially difficult for BD, which is a property not only
determined by the soil composition/constituents which would
be determined in disturbed samples, but also affected by in
situ conditions, such as soil structure conditioned by man-
agement effects such as compaction from machinery.
Saturated moisture being derived from BD is similarly
affected. In addition, most of the BD and SAT values tended to
cluster together around a very limited range (in the high and
low end, respectively). The potential of the technique for the
prediction of BD can be also compromised by the unreliability
or variability in the analytical method. As discussed by Bellon-
Maurel and McBratney (2011) and Soriano-Disla et al. (2014),
predictions of BD using infrared spectroscopy have been car-
ried out by few researchers, with studies showing mixed re-
sults, which highlight the difficulty of predicting this variable
by infrared spectroscopy.
However, a recent approach has shown that gamma-ray
attenuation and viseNIR spectroscopy combine well to mea-
sure accurately the bulk density of soil cores scanned wet
under field conditions (Lobsey & Viscarra Rossel, 2016).
In general, predictive accuracies obtained in this study for
the MIR region were similar to the median R2 calculated by
Soriano-Disla et al. (2014) from a number of studies (Table 3).
Total nitrogen and OC were an exception, as the median R2 in
 30
Table 2 e Summary statisticsa of the partial least squares models developed for the prediction of soil propertiesb across soil instruments.
Instrument Analyte BD SAT DUL OC TC TN EC pH CEC Ca2þ Mg2þ Naþ Kþ ESP Sand Silt Clay
3 þ
g/cm mm/mm % 1:5 dS/m 1:5 w cmol /kg %
N_C 343 343 343 224 336 341 338 327 340 340 340 340 340 340 336 336 336
Range_C 0.8e1.8 0.31e0.68 0.07e0.46 0.02e3.33 0.04e6.19 0e0.27 0.03e1.84 4.4e10.3 1.0e47.5 0.6e34.2 0.2e19.9 0e15.5 0.1e3.2 0e43.6 9.0e97.0 0.8e44.0 2.0e80.2
SD_C 0.2 0.06 0.08 0.51 1.54 0.05 0.31 1.2 10.5 6.7 4.3 3.3 0.5 10.5 17.5 7.7 15.7
N_P 115 115 115 83 112 115 113 112 115 115 115 115 114 115 115 115 115
Range_P 1.0e1.7 0.32e0.60 0.08e0.46 0.02e3.10 0.05e5.43 0e0.20 0.04e2.13 4.5e10.0 2.3e45.4 1.5e34.6 0.3e16.3 0e11.2 0.1e3.5 0e41.1 13.3e93.0 1e38.0 5.0e69.3
SD_P 0.1 0.05 0.08 0.55 1.25 0.05 0.38 1.3 10.3 6.7 4.2 3.4 0.6 11.7 17.0 7.2 15.6

Frontier F 5 5 7 13 4 14 15 12 11 13 14 15 11 13 12 12 12
RMSECV 0.1 0.04 0.05 0.24 0.29 0.03 0.22 0.6 4.8 3.3 2.4 2.2 0.4 7.3 9.0 4.3 8.9

b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6
R2_C 0.59 0.57 0.68 0.77 0.97 0.70 0.53 0.77 0.79 0.76 0.70 0.57 0.45 0.54 0.74 0.68 0.68
RPD_C 1.56 1.54 1.78 2.10 5.39 1.83 1.45 2.10 2.19 2.01 1.81 1.51 1.35 1.44 1.95 1.77 1.77
RMSEP 0.1 0.04 0.05 0.33 0.28 0.03 0.29 0.6 4.9 3.8 2.2 2.3 0.5 7.1 7.2 3.5 7.3
R2_P 0.42 0.39 0.71 0.64 0.95 0.68 0.41 0.82 0.79 0.69 0.72 0.56 0.36 0.63 0.82 0.76 0.78
RPD_P 1.27 1.24 1.84 1.66 4.39 1.75 1.30 2.36 2.13 1.78 1.88 1.49 1.24 1.63 2.36 2.04 2.13

ExoScan F 4 6 8 12 4 14 15 12 12 14 14 15 10 15 12 12 12
RMSECV 0.1 0.04 0.05 0.25 0.30 0.03 0.24 0.6 4.6 3.3 2.4 2.3 0.4 7.2 9.4 4.4 8.8
R2_C 0.56 0.57 0.68 0.76 0.96 0.73 0.45 0.77 0.81 0.76 0.71 0.54 0.43 0.55 0.71 0.67 0.69
RPD_C 1.52 1.52 1.77 2.06 5.20 1.93 1.33 2.08 2.29 2.02 1.84 1.44 1.32 1.46 1.86 1.74 1.78
RMSEP 0.1 0.04 0.05 0.31 0.30 0.02 0.27 0.5 4.6 3.8 2.2 2.0 0.5 6.3 8.1 3.8 7.7
R2_P 0.39 0.39 0.66 0.67 0.94 0.71 0.48 0.84 0.82 0.70 0.74 0.66 0.42 0.70 0.78 0.73 0.76
RPD_P 1.22 1.23 1.73 1.75 4.13 1.82 1.38 2.40 2.26 1.78 1.90 1.68 1.31 1.84 2.09 1.90 2.02

FlexScan F 5 5 8 12 4 15 13 12 12 14 12 15 9 15 13 11 12
RMSECV 0.1 0.04 0.04 0.24 0.31 0.03 0.23 0.6 4.6 3.2 2.4 2.1 0.4 6.8 8.8 4.3 8.5
R2_C 0.58 0.57 0.70 0.78 0.96 0.75 0.46 0.80 0.81 0.78 0.69 0.60 0.48 0.60 0.75 0.69 0.71
RPD_C 1.55 1.53 1.83 2.11 4.96 1.98 1.34 2.23 2.30 2.09 1.81 1.56 1.38 1.54 2.00 1.79 1.86
RMSEP 0.1 0.04 0.05 0.31 0.29 0.02 0.29 0.5 4.7 3.6 2.1 2.2 0.5 7.2 7.0 3.7 6.8
R2_P 0.40 0.39 0.68 0.68 0.95 0.78 0.42 0.83 0.81 0.71 0.74 0.62 0.39 0.63 0.83 0.75 0.81
RPD_P 1.26 1.24 1.77 1.77 4.24 2.12 1.30 2.45 2.21 1.85 1.96 1.58 1.27 1.63 2.43 1.97 2.29

SM-3500 F 8 8 9 12 10 12 7 13 12 13 9 14 11 14 9 11 9
RMSECV 0.1 0.04 0.06 0.29 0.42 0.02 0.24 0.6 5.5 3.4 2.6 1.9 0.4 5.8 11.4 5.1 9.4
R2_C 0.45 0.43 0.53 0.68 0.93 0.77 0.41 0.79 0.73 0.74 0.63 0.68 0.51 0.71 0.58 0.57 0.65
RPD_C 1.33 1.31 1.45 1.74 3.67 2.09 1.30 2.18 1.91 1.96 1.64 1.74 1.42 1.83 1.54 1.51 1.68
RMSEP 0.1 0.05 0.06 0.33 0.39 0.02 0.33 0.6 4.8 3.9 2.8 2.5 0.5 7.8 10.5 5.0 8.0
R2_P 0.34 0.31 0.55 0.65 0.91 0.73 0.25 0.80 0.79 0.67 0.58 0.51 0.45 0.57 0.63 0.54 0.74
RPD_P 1.19 1.15 1.48 1.69 3.22 1.89 1.16 2.22 2.14 1.72 1.49 1.35 1.32 1.49 1.63 1.43 1.94
 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6 31

Soriano-Disla et al. (2014) was larger than the R2 observed in

Summary statistics: Number of samples in calibration (N_C), standard deviation of calibration (SD_C), number of samples in prediction (N_P), standard deviation of prediction (SD_P), number of
model factors (F), root mean square error of the cross validation (RMSECV), coefficient of determination of calibration (R2_C), ratio of the SD to the RMSECV (RPD_C), root mean square error of the

Bulk density (B), saturated moisture (SAT), drained upper limit (0.1 BAR) moisture (DUL), organic carbon (OC), total carbon (TC), total nitrogen (TN), electrical conductivity (EC), cation exchange
our present study. In the case of OC, reasons for a less suc-
12.5
0.37
1.26
11.2
0.49
1.39
cessful performance seem related to a) inaccuracies of the
7

analytical determination method, and to b) the very limited

concentration range, which made the model more influenced
by that analytical error. The relative low R2 for TN was related
0.37
1.26

0.45
1.35
6.1

5.4
9

to the narrow concentration range and low concentration

values. Thus, the error from the reference method had a
higher relative influence. Providing the low RMSEP of the
15.6
0.22
1.13
13.6
0.37
1.25

infrared method, we believe that most of the error was coming

7

from the reference method.

0.42
1.31

0.43
1.31

Our results for soil properties predicted by MIR spectros-

8.1

8.9
7

copy were compared with those (Hicks, Viscarra Rossel, &

Tuomi, 2015; Janik, Forrester, & Rawson, 2009; Pirie, Singh, &
0.27
1.16

0.27
1.14
0.5

0.5

Islam, 2005) obtained using conditions similar to our study

8

(i.e. diverse Australian soils and similar analytical methods).

0.42
1.31

0.45
1.33

In summary, and although comparisons were not straight-

capacity (CEC), exchangeable calcium, magnesium, sodium and potassium (Ca2þ, Mg2þ, Kþ, Naþ) and exchangeable sodium percentage (ESP).
2.5

2.6
7

forward, results obtained by previous studies were similar,

with small differences derived from the fact of using different
sample preparation (i.e. generally, ground samples provide
0.46
1.37

0.45
1.31
3.2

3.2
4

better results; e.g. Hicks et al., 2015), different set of soils,

different concentration ranges (wider concentration ranges
typically resulting in better models), and the influence of few
0.49
1.40

0.48
1.35
4.7

5.0
7

outlier samples.
In the case of vis-NIR, our study showed lower R2 values
than the median values observed in the review by Soriano-
0.55
1.50

0.62
1.58
7.0

6.5
7

Disla et al. (2014) for most of the properties. Our results were
prediction (RMSEP), coefficient of determination of prediction (R2_P), ratio of the SD to the RMSEP (RPD_P).

compared with those reported by other authors using soils

from similar geographical areas, similar analytical methods, a
0.66
1.71

0.63
1.62
0.7

0.8
10

large diversity (Dunn, Beecher, Batten, & Ciavarella, 2002; Pirie

et al., 2005; Viscarra Rossel & Webster, 2012). The results ob-
tained by these authors using vis-NIR were, in general, similar
0.25
0.35
1.24
0.35
0.20
1.09

to those obtained in our study, with differences that could be

9

attributed to the use of different concentration ranges,

different multivariate methods, etc.
0.03
0.72
1.87
0.03
0.60
1.59
13

3.3.2. Comparison of the performance of MIR/vis-NIR

instruments for the prediction of soil properties
0.82
0.72
1.87
0.88
0.60
1.42
11

According to the information provided in Tables 2 and 3, the

portable MIR instruments performed best, with the two in-
struments providing a similar accuracy and being similar to
0.39
0.58
1.30
0.43
0.44
1.29

the laboratory spectrometer (Frontier). The SM-3500 instru-

10

ment was the next most successful, being consistently less

accurate than MIR instruments, with the larger differences
observed for sand and silt, Mg2þ, TC and DUL. For the rest of
0.07
0.26
1.17
0.07
0.27
1.16

the properties, differences in accuracy between the vis-NIR

6

and MIR instruments were marginal. This information can

be summarised by the calculation of median R2 and RPD per
0.05
0.37
1.26
0.05
0.28
1.14

instrument across all properties: ExoScan (R2 ¼ 0.70,

9

RPD ¼ 1.82), FlexScan (R2 ¼ 0.71, RPD ¼ 1.85), SM-3500

(R2 ¼ 0.58, RPD ¼ 1.49), NIRScan (R2 ¼ 0.45, RPD ¼ 1.31),
0.39
1.28

0.30
1.16
0.1

0.1

which compared to the reference laboratory spectrometer,

9

Frontier (R2 ¼ 0.69, RPD ¼ 1.78).

RMSECV

RMSEP
RPD_C

RPD_P

We have demonstrated that portable MIR instrumentation

R2_C

R2_P

can provide similar performance to benchtop instruments

F

leading to increased potential for the in-field prediction of soil

properties. These conclusions were similar to those found
NIRScan

Reeves et al. (2010) for the prediction of TC and TN in soils, by

Forrester et al. (2015) for the prediction of phosphorus buffer
b
a

index in Australian soils, and by Webster et al. (2016) for the

32 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6

Table 3 e Median root mean square error of the prediction (RMSEP), coefficient of determination (R2) and ratio of standard
deviation to RMSEP (RPD) across mid-infrared (Frontier, ExoScan and FlexScan) and vis-NIR (SM-3500) instruments from
Table 2 together with median R2 values from a review (the number of studies used for calculation in brackets) for the
prediction of soil properties.a
Property RMSEP R2 RPD R2 (Soriano-Disla et al., 2014)
MIR Vis-NIR MIR Vis-NIR MIR Vis-NIR MIR Vis-NIR
BD (g/cm3) 0.11 0.12 0.40 0.34 1.26 1.19 NA NA
SAT (mm/mm) 0.04 0.05 0.39 0.31 1.24 1.15 NA NA
DUL (mm/mm) 0.05 0.06 0.68 0.55 1.77 1.48 NA NA
OC (%) 0.31 0.33 0.67 0.65 1.75 1.69 0.93 (20) 0.83 (33)
TC (%) 0.29 0.39 0.95 0.91 4.24 3.22 0.93 (16) 0.89 (27)
TN (%) 0.02 0.02 0.71 0.73 1.82 1.89 0.90 (14) 0.86 (31)
EC (1:5 dS/m) 0.29 0.33 0.42 0.25 1.30 1.16 0.26 (7) 0.60 (1)
pH (1:5 w) 0.55 0.59 0.83 0.80 2.40 2.22 0.75 (8) 0.79 (15)
CEC (cmolþ/kg) 4.67 4.81 0.81 0.79 2.21 2.14 0.85 (13) 0.84 (8)
Ca2þ (cmolþ/kg) 3.76 3.89 0.70 0.67 1.78 1.72 0.82 (8) 0.80 (7)
Mg2þ (cmolþ/kg) 2.20 2.80 0.74 0.58 1.90 1.49 0.74 (7) 0.83 (14)
Naþ (cmolþ/kg) 2.16 2.52 0.62 0.51 1.58 1.35 0.32 (6) 0.61 (2)
Kþ (cmolþ/kg) 0.47 0.46 0.39 0.45 1.27 1.32 0.37 (12) 0.71 (14)
ESP (%) 7.14 7.84 0.63 0.57 1.63 1.49 NA NA
Sand (%) 7.22 10.49 0.82 0.63 2.36 1.63 0.83 (11) 0.78 (15)
Silt (%) 3.66 5.04 0.75 0.54 1.97 1.43 0.63 (8) 0.67 (9)
Clay (%) 7.31 8.04 0.78 0.74 2.13 1.94 0.80 (14) 0.78 (15)
a
Bulk density (B), saturated moisture (SAT), drained upper limit (0.1 BAR) moisture (DUL), organic carbon (OC), total carbon (TC), total nitrogen
(TN), electrical conductivity (EC), cation exchange capacity (CEC), exchangeable calcium, magnesium, sodium and potassium (Ca2þ, Mg2þ, Kþ,
Naþ) and exchangeable sodium percentage (ESP).

prediction of total petroleum hydrocarbons in contaminated
soils.
The higher accuracy of MIR compared with vis-NIR pre-
dictions is in agreement with previous work (reviewed by
Reeves, 2010; Soriano-Disla et al., 2014 and Viscarra Rossel
et al., 2006). However, these reviews compared results that
were obtained on different soils and properties and with
different concentration ranges, with analytical reference data
sourced from different laboratories. For comparison of vis-NIR
with MIR for soil predictions, it is important to use the same set
of reference soils and laboratory analytical data. In general,
results from those studies adopting that approach (e.g. Madari
et al., 2005; McCarty & Reeves, 2006; Pirie et al., 2005; Vohland,
Ludwig, Thiele-Bruhn, & Ludwig, 2014; Yang & Mouazen, 2012)
are also in agreement with the present study, with MIR
generally providing more accurate predictions than vis-NIR for
the majority of soil attributes. Our present study confirmed
this in a highly diverse set of samples, for a wide range of
properties, and using portable MIR and NIR instrumentation.
Mid-infrared spectra contain more, and better resolved, in-
formation than NIR spectra. Visible-NIR is highly sensitive to
sample variability, especially from the point of view of sample
particle size (Brunet, Barthe s, Chotte, & Feller, 2007; Madari
et al., 2005). Particle size variability is one of the main char-
acteristics of the data set that was used in this study.
This generic tendency of MIR models performing better
than vis-NIR was not observed in some articles. Peng et al.
(2014) compared the performance of benchtop vis-NIR and
MIR spectroscopy for the prediction of OC and clay in n ¼ 125
samples collected from a Danish field. Predictions were simi-
larly accurate for both analytes using both ranges. However,
the diversity of soils and range of particle size was limited
comparing with our present study. Rabenarivo et al. (2013)
compared the performance of benchtop NIR and MIR for the
prediction of TC, TN and two biological analytes using dried
and ground soil representative of the main soil types used for
agriculture in Madagascar (n ¼ 360). Contrary to previous
studies, they found that NIR outperformed MIR, suggesting its
possible superiority for tropical soils. They attributed this to
the different mineralogy of tropical soils, typically rich in
haematite, goethite and/or gibbsite. These minerals have a
strong response in the MIR which could overlap with organic
matter (OM) peaks, thus affecting the derived calibrations for
properties such as carbon and nitrogen.
In terms of potential field applications, the presence of
water in field soils could adversely affect MIR predictions for
some soil properties (e.g. Ji et al., 2016). In this study, using a
prototype portable MIR instrument (variable-filter-array
DRIFT; Wilks Enterprise, USA) operating in the 1811e898 cm1
(5522e11,136 nm) range, the authors obtained improved pre-
dictions for a number of soil properties using air-dried
compared with in situ spectra (i.e. OM, Fe, Cu, CEC, Ca, Mg,
BD), while similar accuracies were found for the remaining
properties (i.e. N, P, K, Na, pH). Using the same prototype,
Dhawale et al. (2015) showed that predictions for OC and
particle size did not greatly differ between PLSR models that
used spectra from either air dry or wet soils. However, pre-
dictions were classified as poor and not reliable. Thus, there is
a need to test the influence of variable soil water status on the
prediction of a range of soil properties across contrasting soils
using full range handheld/portable MIR and vis-NIR in-
struments. Independently of the results of further tests, the
MIR technology might be suitable for field applications in dry
regions (e.g. southern Australia and southern Europe),
 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6 33

especially when most of the commercial soil sampling for
agricultural purposes occurs at the end of summer before the
autumn (rainfall) break.
The vis-NIR instrument outperformed the miniature NIR
instrument. The better performance of the full range vis-NIR
instrument for a range of soil properties is not surprising,
and is in agreement with previous studies (e.g. Lee et al., 2009;
Mouazen, De Baerdemaeker, & Ramon, 2006) and attributed to
the wider spectral range of the former. However, the NIRScan
instrument did show some indicative potential for the pre-
diction of CEC, pH, TC and TN. Although the R2 and RPD values
were not high (R2 from 0.60 to 0.63 and RPD from 1.42 to 1.62),
this may be sufficiently accurate for some applications where
high accuracy is not required. In such situations, it would an
advantage to use instruments which are easier to transport,
lighter and cheaper.
We compared our results with other similar studies
reporting the performance of instruments with a restricted
range in the NIR similar to the one used in our present study.
Comparisons are challenging since we used different soils and
concentration ranges, and only few studies for a similar
spectral range (950e1650 nm) are available. A lower error for
the prediction of TN was obtained in our study compared with
that obtained by An, Zheng, Liu, Sun (2014) (RMSEP ¼ 0.043%)
for a similar range using a portable spectrometer and selected
wavelengths in the 940e1550 nm range. Similar results were
found by Mouazen et al. (2006) for a range of properties,
although better predictions were obtained for CEC and Ca2þ,
which is probably related to the fact that a wider spectral
range was used in their study (400e1700 nm) and more
restricted sample diversity, especially in particle size. Similar
to our study, Lee et al. (2009) found significant accuracy
reduction for many soil properties when using the
975e1770 nm range compared to the full vis-NIR range. Thus,
and producing noisier and more variable spectra. The lack of
reduction in prediction accuracy, from instruments with
lower quality spectra, was thought to be due to the mitigating
effects of spectral pre-processing and redundancies in multi-
variate modelling.
In the case of the miniature NIR instrument, we deter-
mined if the less successful models (compared with the full
vis-NIR range) were attributed to the limited spectral range or
to the quality of the spectra produced. Models for the NIRScan
for selected properties were compared with those using the
SM-3500 in the same range as for the NIRScan. Models ob-
tained for the SM-3500 were almost the same as those ob-
tained for the NIRScan (Fig. 4). The poorer performance of the
miniature NIR instrument was therefore related to the
restricted spectral range.
Thus, and for both vis-NIR and MIR instruments, the
spectral range seems to play a more important role than
spectra quality. In terms of reproducibility, and independent
on the instrument used, the adherence to strict and rigorous
laboratory protocols is crucial to get optimum results. Other-
wise, inconsistent errors may arise which may not be
accounted by multivariate pre-processing and modelling,
having a negative effect on predictions.
3.3.4. Optimal spectral ranges for prediction of soil properties
Due to manufacturing, cost or marketing considerations,
many new small or portable infrared instruments have
4.5
3000-5000 nm
4 2500-5555 nm
FronƟer 2000-5000 nm
3.5
5522-11136 nm
3
RPD

and although further studies are needed, it appears that 2.5

mixed results have been found in previous studies using the
2
950e1650 nm with models using this range being less suc-
cessful than those using the full vis-NIR range. 1.5
1
3.3.3. Impact of the quality of the spectra on the performance OC TC TN pH CEC Ca Mg Sand
of predictions Analyte
When comparing the performance of different instruments
4.5
having different or similar spectral range, it is reasonable to
1650-2450 nm
hypothesize that if one instrument produces spectra of higher 4 SM-3500 400-1650 nm
quality and more reproducible, this may have an effect on the 3.5 400-700,1650-2450 nm
prediction accuracy. Thus, comparing predictions provided by 950-1650 nm
3
RPD

different instruments may have a confounding effect.

Compared to benchtop instruments, portable instruments 2.5
such as FlexScan and ExoScan have reduced spectral ranges, 2
lower signal to noise ratio, and produce noisier and more
variable spectra. In addition, when deployed in the field, 1.5
portable instruments have to face climatic conditions that 1
may affect their performance. This was, for example, the case OC TC TN pH CEC Ca Mg Sand
of ExoScan, an instrument that exhibited more variation than Analyte
FlexScan probably as a result of instrument use. However, the
quality of the spectra seemed not to play a major role in in- Fig. 4 e Ratio of the standard deviation of the reference
strument performance for the prediction of soil properties. values to the root mean square error of the prediction (RPD)
This was confirmed by the similar prediction accuracies found values for the prediction of soil properties by partial least
for the Frontier, FlexScan and ExoScan, despite the two latter square regressions using different spectral ranges scanned
having reduced spectral ranges, lower signal to noise ratio, with Frontier and SM-3500.
34 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6
restricted spectral ranges. We therefore assessed a number of
different spectral ranges in both the MIR (using the Frontier
instrument) and vis-NIR (using the SM-3500) to determine the
optimal range for prediction of soil properties.
Within the spectral ranges tested by using the Frontier
spectra, both 2000e5000 nm and 2500e5555 nm ranges
seemed to be the optimum for prediction of the soil properties
considered here. In general, the accuracy of these reduced
spectral range models was very similar to those provided by
the full MIR spectral range, apart from TC and sand. The
3000e5000 nm range was similarly successful, although pH
was predicted with less accuracy than with the other ranges.
The 5522e11136 nm range was the least successful for most
properties. Similar results were reported by Dhawale et al.
(2015) for the prediction of OM using the 5522e11136 nm
range. In the same study, higher R2 values were reported for
the prediction of sand, but also the RMSE, indicating that the
high R2 values were partially influenced by the wide concen-
tration range. Errors were partially attributed to those from
the reference analysis: 8% for the hydrometer method (Gee &
Bauder, 1986). Ji et al. (2016) also used the same spectral range
for the prediction of soil properties in two sites using ground
samples (after being air-dried). In general, they obtained bet-
ter prediction accuracies for CEC, Ca, Mg and OC, and worse
for TN and pH. However, comparisons are not straightforward
since prediction accuracy in Ji et al. (2016) was dependent on
the site tested, and their samples were ground, while in our
study samples were only sieved to <2 mm.
In agreement with Sudduth and Hummel (1991) and Lee
et al. (2009), the most useful information in the vis-NIR re-
gion seems to be contained in 1650e2450 nm range. Models
using this range were similar to those obtained by using the
full vis-NIR range (however, better accuracy was found for the
latter for the prediction of TC and CEC), and considerably
better than those using 950e1650 nm only. The inclusion of
the visible range (400e700, 1650e2450 nm models) only
improved the accuracy for the prediction of TC and sand
comparing with the 1650e2450 nm range. If using the low
range, it was useful to include the visible region, with models
using the 400e1650 nm range being generally better than
those using the 950e1650 nm range.
Thus portable/miniature instruments producing high
quality spectra in the NIR-MIR range of 1650e5000 nm would
be ideal for prediction of the soil properties assessed in this
study. The inclusion of the NIR range in such instruments will
be an advantage despite MIR being more accurate since NIR is
theoretically less affected by variable soil moisture as found in
field samples.
4. Conclusions
Under the conditions of this study, the portable MIR in-
struments e the Agilent ExoScan and FlexScan instruments e
provided the best performance for the prediction of soil
properties, having a performance similar to reference labo-
ratory instrumentation. The SM-3500 instrument (vis-NIR)
was the next most successful, showing minimal differences in
predictive accuracy, apart for predictions of sand, silt, Mg2þ,
TC and DUL. In general, predictive performance could be
regarded as highly satisfactory, and allow confidence in the
use of infrared technique for the prediction of soil properties.
The performance of the miniature instrument
(950e1650 nm) for prediction of soil properties was not up to
the same standard as the other portable instruments. This
was mainly due to the limited spectral range of this instru-
ment rather than to the quality of the spectra produced. The
absence of a relationship between spectra quality and accu-
racy of predictions was also observed when the predictions
from MIR instruments were compared.
Optimum restricted spectral ranges in the NIR and MIR
regions for prediction of soil properties were 1650e2450 nm in
the NIR and 2500e5000 nm in the MIR. Thus, the development
of new portable/miniature producing high quality spectra in
the NIR-MIR range of 1650e5000 nm would be ideal for such
predictions.
As the next step, ongoing work is in process to evaluate the
best performing instruments under field conditions where
large variation on climate and soil conditions is expected.
Acknowledgements
The authors wish to acknowledge the support of the Grains
Research and Development Corporation for funding this
project (GRDC project CSO00045 “Soil infrared capability”), the
provision of soils and analytical data by Linda Karssies and
Peter Wilson (CSIRO National Soil Archive) and valuable
comments by Kirsten Verburg (CSIRO Agriculture and Food).
We thank Dr Graham Miller, Global Channel Development
Manager Mobile FT-IR for Agilent Technologies (UK), for the
provision of the FlexScan instrument for this study.
Appendix A. Supplementary data
Supplementary data related to this article can be found at
http://dx.doi.org/10.1016/j.biosystemseng.2017.06.017.
references
 , M., Blanco, M., Moyano, D., Broad, N. W., O'Brien, N.,
Alcala
Friedrich, D., et al. (2013). Qualitative and quantitative
pharmaceutical analysis with a novel hand-held miniature
near infrared spectrometer. Journal of Near Infrared
Spectroscopy, 21(6), 445e457.
An, X., Li, M., Zheng, L., Liu, Y., & Sun, H. (2014). A portable soil
nitrogen detector based on NIRS. Precision Agriculture, 15(1),
3e16.
Bellon-Maurel, V., & McBratney, A. (2011). Near-infrared (NIR)
and mid-infrared (MIR) spectroscopic techniques for
assessing the amount of carbon stock in soils e Critical
review and research perspectives. Soil Biology and Biochemistry,
43(7), 1398e1410.
Ben-Dor, E., & Banin, A. (1995). Near-infrared analysis as a rapid
method to simultaneously evaluate several soil properties. Soil
Science Society of America Journal, 59(2), 364e372.
Ben-Dor, E., Ong, C. C. H., & Lau, I. C. (2013). Reflectance
measurement of soils in the laboratory: Standards and protocols.
Perth: CSIRO Publications.
 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6
Brunet, D., Barthe s, B. G., Chotte, J. L., & Feller, C. (2007).
Determination of carbon and nitrogen contents in Alfisols,
Oxisols and Ultisols from Africa and Brazil using NIRS
analysis: Effects of sample grinding and set heterogeneity.
Geoderma, 139(1e2), 106e117.
Burk, L., & Dalgliesh, N. (2013). Estimating plant available water
capacity. Grains Research and Development Corporation, Canberra
(Australia). Available at https://grdc.com.au/Resources/
Publications/2013/05/Estimating-Plant-Available-Water-
Capacity (Last accessed 20 March 2017).
Dhawale, N. M., Adamchuk, V. I., Prasher, S. O., Viscarra
Rossel, R. A., Ismail, A. A., & Kaur, J. (2015). Proximal soil
sensing of soil texture and organic matter with a prototype
portable mid-infrared spectrometer. European Journal of Soil
Science, 66(4), 661e669.
Dunn, B. W., Beecher, H. G., Batten, G. D., & Ciavarella, S. (2002).
The potential of near-infrared reflectance spectroscopy for
soil analysis e A case study from the Riverine Plain of south-
eastern Australia. Australian Journal of Experimental Agriculture,
42(5), 607e614.
Forrester, S. T., Janik, L. J., Soriano-Disla, J. M., Mason, S.,
Burkitt, L., Moody, P., et al. (2015). Use of handheld mid-
infrared spectroscopy and partial least-squares regression for
the prediction of the phosphorus buffering index in Australian
soils. Soil Research, 53(1), 67e80.
Gee, G. W., & Bauder, J. W. (1986). Particle-size analysis. Methods of
Soil Analysis, 383e411 (Part 1).
Gobrecht, A., Roger, J. M., & Bellon-Maurel, V. (2014). Major issues
of diffuse reflectance NIR spectroscopy in the specific Context
of soil carbon content estimation. A review. Advances in
Agronomy, 123, 145e175.
Hicks, W., Viscarra Rossel, R. A., & Tuomi, S. (2015). Developing
the Australian mid-infrared spectroscopic database using data
from the Australian soil Resource information System. Soil
Research, 53(8), 922e931.
Janik, L. J., Forrester, S. T., & Rawson, A. (2009). The prediction of
soil chemical and physical properties from mid-infrared
spectroscopy and combined partial least-squares regression
and neural networks (PLS-NN) analysis. Chemometrics and
Intelligent Laboratory Systems, 97(2), 179e188.
Janik, L. J., Merry, R. H., & Skjemstad, J. O. (1998). Can mid infrared
diffuse reflectance analysis replace soil extractions? Australian
Journal of Experimental Agriculture, 38(7), 681e696.
Ji, W., Adamchuk, V. I., Biswas, A., Dhawale, N. M., Sudarsan, B.,
Zhang, Y., et al. (2016). Assessment of soil properties in situ
using a prototype portable MIR spectrometer in two
agricultural fields. Biosystems Engineering, 152, 14e27.
Knadel, M., Stenberg, B., Deng, F., Thomsen, A., & Greve, M. H.
(2013). Comparing predictive abilities of three visible-near
infrared spectrophotometers for soil organic carbon and clay
determination. Journal of Near Infrared Spectroscopy, 21(1),
67e80.
Kuang, B., Mahmood, H. S., Quraishi, M. Z., Hoogmoed, W. B.,
Mouazen, A. M., & van Henten, E. J. (2012). Sensing soil
properties in the laboratory, in situ, and on-line. A review.
Advances in Agronomy, 114, 155e223.
Lee, K. S., Lee, D. H., Sudduth, K. A., Chung, S. O., Kitchen, N. R., &
Drummond, S. T. (2009). Wavelength identification and diffuse
reflectance estimation for surface and profile soil properties.
Transactions of the ASABE, 52(3), 683e695.
Lobsey, C., & Viscarra Rossel, R. (2016). Sensing of soil bulk
density for more accurate carbon accounting. European Journal
of Soil Science, 67(4), 504e513.
Madari, B. E., Reeves, J. B., III, Coelho, M. R., Machado, P. L. O. A.,
De-Polli, H., Coelho, R. M., et al. (2005). Mid- and near-infrared
spectroscopic determination of carbon in a diverse set of soils
from the Brazilian national soil collection. Spectroscopy Letters,
38(6), 721e740.
35
Martens, H., & Naes, T. (1989). Multivariate calibration. Chichester:
John Wiley & Sons.
Matejovic, I. (1997). Determination of carbon and nitrogen in
samples of various soils by the dry combustion.
Communications in Soil Science and Plant Analysis, 28(17e18),
1499e1511.
McCarty, G. W., & Reeves, J. B., III (2006). Comparison of near
infrared and mid infrared diffuse reflectance spectroscopy for
field-scale measurement of soil fertility parameters. Soil
Science, 171(2), 94e102.
Mouazen, A. M., De Baerdemaeker, J., & Ramon, H. (2006). Effect of
wavelength range on the measurement accuracy of some
selected soil constituents using visual-near infrared
spectroscopy. Journal of Near Infrared Spectroscopy, 14(3),
189e199.
Mouazen, A. M., Saeys, W., Xing, J., De Baerdemaeker, J., &
Ramon, H. (2005). Near infrared spectroscopy for agricultural
materials: An instrument comparison. Journal of Near Infrared
Spectroscopy, 13(2), 87e97.
Nguyen, T. T., Janik, L. J., & Raupach, M. (1991). Diffuse reflectance
infrared Fourier transform (DRIFT) spectroscopy in soil
studies. Australian Journal of Soil Research, 29(1), 49e67.
Nocita, M., Stevens, A., van Wesemael, B., Aitkenhead, M.,
Bachmann, M., Barthe s, B., et al. (2015). Soil spectroscopy: An
alternative to wet chemistry for soil monitoring. Advances in
Agronomy, 132, 139e159.
Palm, C., Sanchez, P., Ahamed, S., & Awiti, A. (2007). Soils: A
contemporary perspective. Annual Review of Environment and
Resources, 32, 99e129.
Peng, Y., Knadel, M., Gislum, R., Schelde, K., Thomsen, A., &
Greve, M. H. (2014). Quantification of SOC and clay content
using visible near-infrared reflectanceemid-infrared
reflectance spectroscopy with jack-knifing partial least
squares regression. Soil Science, 179(7), 325e332.
Pirie, A., Singh, B., & Islam, K. (2005). Ultra-violet, visible, near-
infrared, and mid-infrared diffuse reflectance spectroscopic
techniques to predict several soil properties. Australian Journal
of Soil Research, 43(6), 713e721.
Rabenarivo, M., Chapuis-Lardy, L., Brunet, D., Chotte, J. L.,
Rabeharisoa, L., & Barthe s, B. G. (2013). Comparing near and
mid-infrared reflectance spectroscopy for determining
properties of Malagasy soils, using global or LOCAL
calibration. Journal of Near Infrared Spectroscopy, 21(6), 495e509.
Rayment, G. E., & Higginson, F. R. (1992). Australian laboratory
Handbook of soil and water chemical methods. In Australian Soil
and Land Survey Handbooks Series (Vol. 3). Melbourne: Inkata press.
Rayment, G. E., & Lyons, D. J. (2011). Soil chemical methods-
Australasia. Collingwood: CSIRO Publishing.
Reeves, J. B., III (2010). Near- versus mid-infrared diffuse
reflectance spectroscopy for soil analysis emphasizing carbon
and laboratory versus on-site analysis: Where are we and
what needs to be done? Geoderma, 158(1e2), 3e14.
Reeves, J. B., McCarty, G. W., & Hively, W. D. (2010). Mid-versus
near-infrared spectroscopy for on-site analysis of soil. In
R. A. Viscarra Rossel, et al. (Eds.), Proximal soil sensing, Progress
in soil science 1 (pp. 133e142). The Netherlands: Springer
ScienceþBusiness Media.
Reimann, C., Birke, M., Demetriades, A., Filzmoser, P., &
O'Connor, P. (2014). Chemistry of Europe's agricultural soils,
Part B. General background information and further analysis
of the GEMAS data set. In Geologisches Jahrbuch Reihe B, band B
103. Stuttgart: Schweizerbart Science Publishers.
Savitzky, A., & Golay, M. J. E. (1964). Smoothing and
differentiation of data by simplified least squares procedures.
Analytical Chemistry, 36(8), 1627e1639.
Shepherd, K. D., & Walsh, M. G. (2002). Development of
reflectance spectral libraries for characterization of soil
properties. Soil Science Society of America Journal, 66(3), 988e998.
36 b i o s y s t e m s e n g i n e e r i n g 1 6 1 ( 2 0 1 7 ) 2 4 e3 6
Shepherd, K. D., & Walsh, M. G. (2007). Review: Infrared
spectroscopy e Enabling an evidence-based diagnostic
surveillance approach to agricultural and environmental
management in developing countries. Journal of Near Infrared
Spectroscopy, 15(1), 1e19.
Shi, Z., Wang, Q. L., Peng, J., Ji, W. J., Liu, H. J., Li, X., et al. (2014).
Development of a national VNIR soil-spectral library for soil
classification and prediction of organic matter concentrations.
Science China Earth Sciences, 57(7), 1671e1680.
Soriano-Disla, J. M., Janik, L., McLaughlin, M. J., Forrester, S.,
Kirby, J., Reimann, C., et al. (2013). The use of diffuse
reflectance mid-infrared spectroscopy for the prediction of the
concentration of chemical elements estimated by X-ray
fluorescence in agricultural and grazing European soils.
Applied Geochemistry, 29, 135e143.
Soriano-Disla, J. M., Janik, L. J., Viscarra Rossel, R. A.,
MacDonald, L. M., & McLaughlin, M. J. (2014). The performance
of visible, near-, and mid-infrared reflectance spectroscopy for
prediction of soil physical, chemical, and biological properties.
Applied Spectroscopy Reviews, 49(2), 139e186.
Stenberg, B., Viscarra Rossel, R. A., Mouazen, A. M., &
Wetterlind, J. (2010). Visible and near infrared spectroscopy in
soil science. Advances in Agronomy, 107(C), 163e215.
Sudduth, K. A., & Hummel, J. W. (1991). Evaluation of reflectance
methods for soil organic matter sensing. Transactions of the
ASAE, 34(4), 1900e1909.
Sudduth, K. A., & Hummel, J. W. (1996). Geographic operating
range evaluation of a NIR soil sensor. Transactions of the
American Society of Agricultural Engineers, 39(5), 1599e1604.
Viscarra Rossel, R. A., Behrens, T., Ben-Dor, E., Brown, D. J.,
Dematte ^, J. A. M., Shepherd, K. D., et al. (2016). A global
spectral library to characterize the world's soil. Earth-science
Reviews, 155, 198e230.
Viscarra Rossel, R. A., Walvoort, D. J. J., McBratney, A. B.,
Janik, L. J., & Skjemstad, J. O. (2006). Visible, near infrared, mid
infrared or combined diffuse reflectance spectroscopy for
simultaneous assessment of various soil properties. Geoderma,
131(1e2), 59e75.
Viscarra Rossel, R. A., & Webster, R. (2012). Predicting soil
properties from the Australian soil visible-near infrared
spectroscopic database. European Journal of Soil Science, 63(6),
848e860.
Vohland, M., Ludwig, M., Thiele-Bruhn, S., & Ludwig, B. (2014).
Determination of soil properties with visible to near- and mid-
infrared spectroscopy: Effects of spectral variable selection.
Geoderma, 223e225(1), 88e96.
Webster, G. T., Soriano-Disla, J. M., Kirk, J., Janik, L. J.,
Forrester, S. T., McLaughlin, M. J., et al. (2016). Rapid prediction
of total petroleum hydrocarbons in soil using a hand-held
mid-infrared field instrument. Talanta, 160, 410e416.
Williams, P. C. (1987). Implementation of near-infrared
technology. In P. C. Williams, & K. H. Norris (Eds.), Near-
infrared technology in the agricultural and food industries (pp.
145e169). St. Paul: American Association of Cereal Chemists.
Yang, H., & Mouazen, A. M. (2012). Vis/near and mid-infrared
spectroscopy for predicting soil N and C at a farm scale. In
T. Theophile (Ed.), Infrared spectroscopy e Life and biomedical
sciences (pp. 185e210). Rijeka: InTech.