SOIL REFLECTANCE SENSING FOR DETERMINING

SOIL PROPERTIES IN PRECISION AGRICULTURE

J. A. Thomasson, R. Sui, M. S. Cox, A. AlRajehy

ABSTRACT. Soil samples were collected on a 0.4ha (1.0ac) grid from two agricultural fields in northeastern Mississippi.
The samples were measured for soil nutrient composition, soil texture, and diffuse reflectance between 250 and 2500 nm. The
data were examined for two purposes: (1) to understand the relationships between soil properties and reflectance spectra,
and (2) to understand the sources of variability in the reflectance spectra. From the raw reflectance spectra, 50nmband
averages were calculated. There were significant correlations between groups of the averaged spectra and soil properties,
but no single 50nm band was highly correlated to any soil property. Soil nutrients were better correlated with spectra in one
field, but texture was better correlated with spectra in the other. Only Ca and Mg in one field and clay and pH in the other
had multipleregressor model correlations with R2 values greater than 0.50. The relationships between soil properties and
reflectance spectra were not consistent between fields. Based on the study of variability in reflectance spectra, the results of
which were significantly dependent on the instrument used in this study, it was found that certain sections of the spectrum
are more useful for discriminating among soil samples with differing characteristics. Spectral regions of high discriminatory
power were 400 to 800 nm and 950 to 1500 nm.
Keywords. Soil properties, Spectrophotometer, Precision agriculture, Remote sensing, Reflectance.

recisionagriculture technologies are being developed to optimize farm profit and minimize environmental impact by adjusting production inputs to the
needs of individual areas within fields. For precision
agriculture to be successful, three things are required: (1) accurate sitespecific data about field conditions, (2) an understanding of relationships between the data and economic/
environmental benefits, and (3) the ability to vary inputs by
location.
One of the first and most important areas in which
precision agriculture has been commercially applied is in
managing the variability in soil fertility. The data for this
purpose come from laborious soilsample collection efforts
and expensive laboratory analyses. It is clear that new
methods for determining variability in soil characteristics
would be quite welcome. Since Mississippis Deltaregion
farmland consists of highly variable alluvial soils, and since
over 80% of the states row crops are produced there, it is an
ideal location for the eventual application of new soilrelated
precisionagriculture technologies.
Mississippi Delta soils have been mapped by the USDA
Natural Resources Conservation Service (formerly Soil
Conservation Service), but the surveys are typically 30 or

Article was submitted for review in September 2000; approved for

publication by the Power & Machinery Division of ASAE in October 2001.
Presented at the 2000 ASAE Annual Meeting as Paper No. 001044.
The authors are J. Alex Thomasson, ASAE Member Engineer,
Associate Professor, Ruixiu Sui, ASAE Member, Assistant Agricultural
Engineer, and Abdulrahman AlRajehy, Graduate Student, Department
of Agricultural and Biological Engineering; and Michael S. Cox, Associate
Professor, Department of Plant and Soil Science, Mississippi State
University, Mississippi State, Mississippi. Corresponding author: J. Alex
Thomasson, Box 9632, Mississippi State, Mississippi 39762; phone:
6623253282; fax: 6623253853; email: alex@abe.msstate.edu.

more years old and do not contain sufficient detail for

withinfield farm management. Since accurate data are very
important for precision agriculture, it is worth considering
whether optical sensors, which are adept at largescale and
rapid data collection, would be appropriate for collecting
data on soil conditions. Figure 1 includes spectral reflectance
curves that indicate that significant optical differences exist
among various soil types (data collected in association with
the research reported herein). Optical data from a field could
be obtained with either groundbased sensors or remote
sensing such as aerial or satellitebased imaging. It would be
ideal to be able to make useful soil maps within fields with
a rapid and relatively inexpensive method like remote
sensing. A number of soil properties important to crop
production are potentially detectable with remote sensing:
texture, mineralogy, organic matter content (O.M.), moisture
content, etc. A significant body of work has been developed
over the years concerning the optical properties of soils.
Some recent examples from the literature are given below.

LITERATURE REVIEW
LABORATORY INSTRUMENTBASED EXPERIMENTS
Starting in the early 1980s, there have been reports
(Krishnan et al., 1981; Pitts et al., 1986; Sudduth and
Hummel, 1991) of the use of nearinfrared (NIR) reflectance
for measuring soil properties, primarily O.M. Sudduth and
Hummel (1993a, 1993b, 1996) continued to develop this
technology and tested a portable NIR spectrophotometer for
measuring O.M., cationexchange capacity (CEC), and
moisture content. They found that, while O.M. could be
detected reasonably well in the field with such a system,
acceptable accuracy for a single calibration equation was
limited to a confined geographic area including Illinois,

Transactions of the ASAE

Vol. 44(6): 14451453

E 2001 American Society of Agricultural Engineers ISSN 00012351

1445

100

fine white sand

Reflectance (%)

80
silt
60
40

clay

20
top soil
0
250

500

750 1000 1250 1500 1750 2000 2250 2500

Wavelength (nm)

Figure 1. Spectral reflectance of reference soils.

Missouri, Indiana, and Ohio. Using the system in a broader

geographic area would require calibrations for other specific
regions.
ViscarraRossel and McBratney (1998) collected soils
from a site in New South Wales, Australia, and prepared them
by drying and sieving with a 2mm sieve. The soil samples
had varying amounts of clay, soil water, and O.M. Reflectance spectra were measured at 2nm intervals from 1300 to
2500 nm. Four wavelengths in that range (1600, 1800, 2000,
and 2100 nm) were considered in detail for their relationships
with clay, soil water content, and O.M. The most accurate
predictions of clay and water content were obtained at
2100 nm, while the least accurate were obtained at 1600 nm.
Clay content predictions were in good agreement with the
actual values. Water content predictions also were correlated
with actual water content values, but the prediction accuracy
was lower than for clay content. Results indicated significant (at the 5% level) responses to clay and soil water content
at all wavelengths considered, but found no significant
relationships with O.M.
BenDor and Banin (1995) stated that NIR spectroscopy
of soils was a good method for rapid and nondestructive
determination of clay content, specific surface area (SSA),
CEC, hygroscopic moisture (HIGF), calcium carbonate
content (CaCO3), and O.M. They studied 91 soil samples
collected from the A horizon of Israeli soils. All chemical and
spectroscopic determinations were carried out with the fine
portion of the soil samples (particle diameters smaller than
0.355 mm). Clay content, SSA, CEC, HIGF, O.M., and
CaCO3 were obtained with chemical analyses. The spectral
reflectances were recorded with an FTIR spectrometer
optimized to a specific NIR (1.0 to 2.5 nm) region. A constant
wavenumber resolution of 8 cm1 produced 3113 spectral
points in that region. Soil properties were found to fit into
three groups: Group I (clay content, SSA, CEC, and HIGF),
Group II (O.M.), and Group III (CaCO3 content). Group I
properties were correlated among themselves with R values
greater than 0.643, and they were all highly correlated to the
clay content and/or SSA. Group II values correlated moderately with Group I, having R values ranging from 0.412 to
0.665. The authors related this association to the adsorption
properties of soil (including SSA, CEC, and HIGF), either
directly or indirectly through the correlation with clay
content. Finally, Group III values did not correlate with any
of the other properties.

1446

Ingleby and Crowe (2000) used a UVVisNIR spectrophotometer to measure diffuse reflectance spectra of 420 soil
samples collected from five fields in Saskatchewan, Canada.
They developed multiple linear regression models, a separate
model for each field, to predict soil organic carbon content
(O.C.) with spectral reflectance data. They found strong
correlations between predicted and actual O.C., with R2
values ranging from 0.73 to 0.89 on the five fields.
Shibusawa et al. (1999) developed a portable spectrophotometer for collecting underground soil spectral reflectance, which was obtained at a depth of 35 cm in the
wavelength range of 400 to 1700 nm. Correlations between
spectral reflectance and soil properties such as moisture
content, pH, CEC, and O.M., were evaluated. The results
showed that these soil properties had a close correlation (R2 =
0.61 to 0.87) with spectral reflectance at certain wavelengths.
Dematt and Garcia (1999) used a spectroradiometer to
measure the spectral reflectance of soil samples collected
from basaltic soils in Brazil. The samples were in two depth
ranges, 0 to 20 cm and 40 to 60 cm. The spectral data were
in the range of 400 to 2500 nm. Their results indicated that
soils could be separated at the soil type level, that amorphous
and crystalline iron (Fe) in the soils influenced reflectance
differently, and that weathering of the basaltic soils was
correlated with differences in the spectral reflectance curves.
Further, they used multivariate analysis of the spectral data
to estimate, reasonably well, the amounts of clay, silt,
kaolinite, crystalline Fe, amorphous Fe, and Mg.
REMOTE SENSINGBASED EXPERIMENTS
Moran et al. (1997) reviewed several research efforts prior
to 1997 that used remote sensing to map soil properties
including loess thickness, O.M., CaCO3, soil nutrients
related to soil texture and drainage, iron oxide content, and
soil texture. However, they noted that remote sensing was not
being employed for this purpose regularly because soil
reflectance properties are often confused with variations in
soil moisture content, surface roughness, climatic conditions, solar zenith angle, and view angle. Nevertheless,
other researchers continue to look at remote sensing as an aid
in soil mapping.
Muttiah et al. (1998) developed a linear programming
procedure to classify cotton areas from Landsat and AVHRR
(Advanced Very High Resolution Radiometer) data. They
found that Landsat data were useful in distinguishing among
areas with a growing crop, bare soil, disked bare soil,
nondisked low straw, nondisked high straw, and disked soil
with high straw that had been redisked.
PalaciosOrueta and Ustin (1998) studied the relationship
between remotesensing reflectance data collected with the
AVIRIS (Advanced Visible/Infrared Imaging Spectrometer)
and Fe and O.M. contents, and texture, in soils from the Santa
Monica Mountains in California. They found that total Fe,
O.M., and sand were the main factors affecting the spectral
curve. Soils with low sand content had low reflectance. They
concluded that levels of Fe and O.M. could be discriminated
with sensors having the spectral resolution of the AVIRIS
sensor.
Barnes and Baker (1999) found that spectral classification
procedures based on aerial and satellite images could be used
to map soil textural classes in individual fields with a
reasonable degree of accuracy. Further, Barnes et al. (1996)

TRANSACTIONS OF THE ASAE

suggested ways in which spectral information from bare soil

could be applied to gridded soil sample data to improve
interpolation.
Coleman et al. (1993) studied spectral differentiation of
surface soils and soil properties with Landsat TM data and the
Barnes Modular Multiband Radiometer (MMR). They found
significant correlations among remotesensing reflectance
data in all seven Landsat bands and the soil variables studied:
soil texture, O.M., and Fe content. Prediction equations
generated from the same area with the MMR data produced
correlation coefficients of 0.665, 0.496, 0.633, 0.686, and
0.213 for sand, silt, clay, O.M., and Fe content, respectively.

OBJECTIVES
The goal of the research reported herein was to develop
optical methods that can provide data for precisionagriculture management decisions concerning the soil. Toward that
end, the objectives of this study were:
S To examine the relationships between soil reflectance
spectra and the physical and chemical properties of soil in
the study site.
S To analyze soil reflectance spectra to understand sources
of variation in the data.

MATERIALS AND METHODS

STUDY SITE
The study site consisted of two farm fields (referred to as
field 1 and field 3) in the northern Mississippi Delta region
near Vance, Mississippi. Fields 1 and 3 are approximately 111
and 162 ha in size, respectively. The fields are about 1.6 km
apart but in different counties, field 1 being in southern
Quitman County and field 3 being in northern Tallahatchie
County. The USDA/NRCS soil survey lists these two fields
as including several soil types (table 1). It can be seen in
remotely sensed images that a large amount of variation
exists in visible reflectance from the bare soil within the
fields, so it is to be expected that a large amount of soiltype
variability exists also.
SAMPLE COLLECTION AND PREPARATION
A total of 724 soil samples were collected from the study
site in July 1999: 276 from field 1 and 448 from field 3. The
location of each sample was selected with a differential GPS
receiver such that each sample represented about 0.4 ha
(1 ac). From each location, 400 to 700 g of soil was collected

Soil

to a depth of 150 mm to 200 mm with a 25mm diameter

manual soil probe. The soil core from one probe was
considered to be one subsample, and each sample was a
mixture of five subsamples. One of the five subsamples was
taken at the sampling position indicated by GPS. The other
four subsamples were taken from random points within 3 m
of the GPSindicated position.
Preparation consisted of drying soil samples in open air at
ambient temperature in a nonenvironmentally controlled
laboratory. The airdried samples were ground until the
entire sample passed through a 2mm screen.
SAMPLE MEASUREMENT TECHNIQUES
Chemical Measurements
Chemical contents of the soil samples were measured with
the Mississippi Soil Test method. This method predominantly involves creating a liquid extract from each sample that is
then subjected to atomicemission spectroscopy (AES).
Readings from the AES instrument indicate concentrations
of calcium (Ca), potassium (K), magnesium (Mg), sodium
(Na), phosphorus (P), and zinc (Zn) in the soil. Soil pH was
determined by measuring a mixture of soil and deionized
water with a pH probe.
Organic matter was not considered in this study because
soils in the Mississippi Delta typically contain less than 1.0%
organic matter, as opposed to 2% to 5% in the Midwest, and
organic matter values change rapidly due to the warm and
humid climate. Since the thrust of this work is towards the use
of remote sensing, which generally does not provide
information in real time, it was deemed reasonable not to
consider organic matter in this effort. Furthermore, since
optical remote sensing usually takes into account only the
surface layer of the soil, it was deemed unreasonable to
consider nitrogen variations in the soil, which are generally
measured with a deep core (approximately 1.0 m) sample.
Means and standard deviations of soil chemical properties
are given in table 2.
Physical Measurements
Soil texture was determined with the hydrometer method.
This method determines the approximate proportion of clay
(particle diameters < 2.0 m), silt (2.0 to 50 m), and sand
(50 to 2000 m) particles in a soil as follows (Research
Analytical Laboratory, 2000): 40 g of prepared soil is shaken
for 16 h with 100 ml of 5% sodium hexametaphosphate; the
suspension is transferred to a sedimentation cylinder and
brought to a total volume of 1 L with deionized water; after

Table 1. Soil types given by USDANRCS soil surveys for fields at study site.
Family
Subgroup

Field 1
Souva silt loam, nearly level phase
Dowling clay and silty clay
Dundee fine sandy loam, nearly level phase
Dubbs fine sandy loam, nearly level phase

Souva: Not given

Dowling: Not given
Dundee: Fine silty, mixed, thermic
Dubbs: Fine silty, mixed, thermic

Not given
Not given
Aeric Ochraqualfs
Typic Hapludalfs

Field 3
Alligator clay, depressional
Dundee silt loam, 0% to 2% slopes
Dundee silt loam, 0% to 3% slopes
Tensas silt loam, 0% to 3% slopes
Dubbs very fine sandy loam, 0% to 2% slopes
Forestdale silty clay loam, 0% to 3% slopes

Alligator: Fine montmorillonitic, acid, thermic

Dundee: Fine silty, mixed, thermic
Dundee: Fine silty, mixed, thermic
Tensas: Fine montmorillonitic, thermic
Dubbs: Fine silty, mixed, thermic
Forestdale: Fine mixed, thermic

Vertic Haplaquepts
Aeric Ochraqualfs
Aeric Ochraqualfs
Aeric Ochraqualfs
Typic Hapludalfs
Typic Ochraqualfs

Vol. 44(6): 14451453

1447

Table 2. Means and standard deviations for soil properties.

Field 1
Field 3

Soil
property
Clay (%)
Sand (%)
P (ppm)
K (ppm)
Ca (ppm)
Mg (ppm)
Zn (ppm)
Na (ppm)
pH

Mean

S.D.

Mean

S.D.

37.32
32.45
57.51
237.4
2108
116.7
4.684
92.07
5.976

15.30
8.957
22.78
118.9
453.9
38.51
5.225
44.21
0.3928

35.50
36.10
43.23
176.8
1611
147.0
2.971
65.37
5.736

13.12
8.520
30.05
152.0
578.2
119.0
4.117
78.13
0.4465

a 2hr temperature equilibration, the suspension is stirred

vigorously for 60 s to resuspend the particles; an ASTM No.
152H hydrometer is carefully placed in the suspension and
used to take two readings, one at 40 s and another at 6 to 8 h
(depending on suspension temperature); the percentages of
clay, silt, and sand in the soil are calculated from the
hydrometer readings. In practice, clay and sand contents are
calculated from the readings, and silt is found by subtracting
the percentages of clay and sand from 100%. Hydrometer
readings for sand were checked against the sieving method on
approximately one of every 10 samples. Means and standard
deviations of soil physical properties are given in table 2.
Spectrophotometric Measurements
The reflectance spectra for each soil sample were
collected with a Cary 500 UV/Vis/NIR spectrophotometer.
This spectrophotometer is equipped with a diffusereflectance accessory that incorporates an integrating sphere.
Because of the integrating spheres geometry, it is able to
collect almost all the reflected radiation, remove any
directional preferences, and present an integrated signal to
the detector. The collected spectra for each sample consisted
of 970 reflectance values, each with an averaging time of
0.1 s, at wavelengths from 250 nm to 2500 nm. The spectral
resolution selected in the 250 to 792 nm range was 1 nm, and
that used in the 792 to 2500 nm range was 4 nm.
Soil samples were prepared by placing a small amount of
soil into a small plastic sample holder with a window on one
side. The sample holder was designed and fabricated
specifically so that soil samples could be presented to the
spectrophotometer. The sample holder was 25 mm tall,
25 mm in outside diameter, and 20 mm in inside diameter. It
had a window at one end and a removable cap at the other end.
The window was a piece of sapphire glass with a thickness of
1 mm and a diameter of 22 mm. Sapphire was chosen because
its transmission from 200 to 6000 nm is roughly constant (at
about 85% in the case of these 1mm thick windows).
Before collecting soil reflectance spectra in a given data
collection session (i.e., at least once per day), a spectral
reflectance baseline was recorded, after instrument warm
up, with a reference disk covering the sample port of the
diffuse reflectance accessory. The reference disk in this case
was a manufacturerprovided, secondarywhitestandard,
polytetrafluoroethylene (PTFE) disk calibrated relative to a
perfectly diffuse reflector. In order to account for the light
attenuation caused by the optical window in the sample
holder, the same type of sapphire glass as used in the sample
holder was placed on top of the PTFE disk during baseline
collection.

1448

A sifted soil sample was added to the sample holder until

it was full (typically about 6.6 g of soil), and the soil was
retained in the holder by placing the removable cap over the
open end of the holder. Each soil sample was then mounted
over the sample port with the sample holders window
pressed against the sample port, and the sphere collected the
energy reflected from the soil sample surface. The standardized spectral reflectance of a sample was calculated as the
ratio of the flux reflected by the sample to that reflected by
the reference disk under identical geometrical and spectral
illumination conditions. After each reflectance measurement, the cap was opened and the sample in the holder was
removed. Then the optical window of the holder was cleaned
before adding the next sample.
EXPERIMENT 1: RELATIONSHIPS BETWEEN SPECTRAL
REFLECTANCE AND CHEMICAL/PHYSICAL DATA
Data Collection
One spectral measurement was collected for each of the
724 samples in this part of the study. Soil texture measurements were also collected for all samples without replication.
Soil chemical properties were collected in triplicate for all
samples, with the average used in the analysis. Due to the
large number of samples, the effect of error sources on
spectral data was investigated with a separate study (Experiment 2, described below) on a reduced number of soil
samples rather than by repeated measurements of the entire
sample set.
Data Analyses
Multiplelinear regression was used to develop models of
reflectance spectra that were most highly correlated with a
soil property of interest. Since there were 970 reflectance
values for each sample, it was necessary to significantly
decrease the amount of spectral data that would be used in
developing models. For this purpose, each sample spectrum
was divided into 45 wavebands, each one being 50 nm wide,
and an average reflectance was calculated for each 50nm
waveband. The center wavelengths for the 50nm wavebands were as follows: 275, 325, 375, , 2425, and 2475. It
is reasonable to divide the spectrum in this manner for two
reasons: (1) sensor development often employs optical filters
with bandwidths in the 10 to 100 nm range, and (2) files of
spectral data including a sizeable number of samples are of
such magnitude that they often do not lend themselves to
rapid computer analysis at the spectral resolution of the data
(i.e., some reduction of data volume is often required prior to
analysis).
These 50nm averages were analyzed with the SAS
procedure Proc REG (SAS, 1999) to quantify correlations
between reflectance and soil properties. The maximum
model size with this procedure was 45, so it was important to
consider what number of spectral variables would constitute
the appropriate model size. The REG procedure with the
select = Cp option optimizes the value of the Mallows Cp
statistic as it calculates regression parameters and statistics
with different numbers of regressors. The output of Proc REG
with the select = Cp option gives an estimate of the most
appropriate model among all those possible. The Cp statistic
strikes a balance between the improvement in fit by adding
a new regressor to the model and the corresponding increase

TRANSACTIONS OF THE ASAE

in variability. Equation 1 gives details of how Cp is

calculated:
2

Cp = p +

(s 2 ^ )(n p)

(1)

^ 2

where
p = number of model parameters (regressors)
s2 = error mean square for the candidate model
2

^
= estimate of 2 (population variance)
s
n = number of data points.
When the value of Cp is approximately equal to the
number of regressors in the model, a reasonable model is
indicated. It should be pointed out that Proc REG uses the
2

^
error mean square from the most complete model as s
. If this
is not a good estimate, then the bias portion of Cp can be
negative, in which case Cp can be less than p (Walpole and
Myers, 1993). According to Myers (1990), the lowest value
of Cp in a group of regression models generally indicates the
most appropriate model. The SAS procedure, Proc REG with
the select = Cp option selects the model with the lowest
value of Cp .
The models correlating reflectance spectra with soil
properties were developed separately for the two fields in the
study area. Because of the interest in having broadly
applicable models for determining soil properties, the models
developed for each field were applied to the data from the
other field. The best model for each soil property in each
field, as judged by having the lowest Cp value, was tested on
the other fields data, and comparisons were made between
models.

EXPERIMENT 2: DIFFERENTIATING VARIABILITY IN THE

SPECTRAL REFLECTANCE DATA
Data Collection
Reflectance spectra of 21 soil samples, selected at random
from the total of 724 samples, were measured repeatedly in
two different ways. First, five subsamples were collected
from each soil sample, and the spectral reflectance of these
five subsamples was measured once. Then, one of the five
subsamples was scanned repeatedly for a total of five times
without removing the sample holder from the sample port of
the diffuse reflectance accessory. Repeating measurements
in this way allowed the variability among samples, the
variability within samples, and the variability associated
with measurement error to be evaluated independently. This
experiment was important, because, of the total reflectance
variability within a group of soil samples, it is important to
know what portion is a result of measurement error, what
portion is a result of nonhomogeneity of individual samples
including sample preparation and presentation to the instrument, and what portion relates to true differences in soil
samples from different parts of a field.
Data Analyses
Spectral data for evaluating measurement variability were
analyzed with the SAS procedure Proc SUMMARY (SAS,
1999) to determine the coefficient of variation (CV) at all
wavelengths from 250 to 2500. First, the CV was calculated
for all five subsamples of the 21 samples, not including the
replications for measurement error. This CV was considered

Vol. 44(6): 14451453

as total CV. Second, the CV was calculated for all five

subsamples of each individual sample, not including the
replications for measurement error. This calculation gave
21 sample CV values, which were averaged together into a
CV considered as withinsample CV. Finally, the CV was
calculated for all five replications of the one replicated
subsample of each sample. This calculation also gave
21 sample CV values, which were averaged together into a
CV considered as measurement CV. It is duly noted that
this method of calculating and considering CVs is not a
rigorous statistical approach to partitioning variability.
However, with the number of variables (wavelengths) being
considered here, a traditional analysis of variance was
viewed as unnecessarily cumbersome. The hope was that
certain spectral bands would be found to be particularly good
with regard to considering total variability and the partitioning of it (i.e., certain bands would exhibit high total CV and
low withinsample and measurement CVs). It should be
noted that certain wavebands were expected to be poor areas
of the spectrum to make soil measurements. For instance,
wavelengths around 800 nm are affected by the detector
change that occurs in the spectrophotometer at 792 nm.
Furthermore, wavelengths from roughly 2175 nm to 2275 nm
are affected by an absorption peak caused by the vibration
mode of the hydroxyl ion (Baumgardner et al., 1985).

RESULTS
EXPERIMENT 1: RELATIONSHIPS BETWEEN SPECTRAL
REFLECTANCE AND CHEMICAL/PHYSICAL DATA
It was clear from the analyses that no individual 50nm
bands were highly correlated with any soil property under
consideration. The only soil properties that had singleregressor models with R2 values above 0.10 were Ca (R2 = 0.17)
and Mg (R2 = 0.23), and only in field 1. While these
singleband correlations were statistically significant at the
5% level, they were not strong enough to be useful in
designing spectral sensors for soilproperty measurement.
The best multipleregressor spectralreflectance model for
each soil property, selected with the REG procedure, is given
for fields 1 and 3 in tables 3 and 4, respectively. The models,
all of which were significantly correlated (at the 5% level) to
the property of interest, are those with the lowest Cp value.
Several things are evident upon review of these tables.
First, soil nutrients were more highly correlated with
reflectance spectra on field 1 than on field 3, while soil
texture properties were more highly correlated with reflectance spectra on field 3 than on field 1. Second, the only soil
properties that had multipleregressor models with R2 values
greater than or equal to 0.50 were Ca (R2 [ 0.7) and Mg
(R2 [ 0.7) in field 1 and clay (R2 [ 0.6) and pH (R2 [ 0.5)
in field 3 (the R2 values for clay, P, and pH in field 1 were
nearly 0.50). Third, the soil properties that were somewhat
well correlated with soil reflectance spectra on one field were
generally not as well correlated in the other field. Finally, the
wavelengths included in the best correlation models were not
similar between fields. Figures 2(a) through 2(i) include
betweenfield comparisons of the spectral wavelengths that
were most correlated with a given property. There is no
obvious evidence that particular wavelengths were broadly
applicable for a given soil property.

1449

Table 3. Field 1 models relating soil properties to reflectance spectra.

Soil
property

No. of
regressors

R2

Zn
Sand
pH
P
Na
Mg
K
Clay
Ca

15
17
13
19
17
16
19
19
15

0.3131
0.2925
0.4902
0.4879
0.3907
0.7332
0.4030
0.4917
0.7210

Soil
property

No. of
regressors

R2

Zn
Sand
pH
P
Na
Mg
K
Clay
Ca

10
10
20
13
20
15
7
20
13

0.1722
0.3324
0.5143
0.2071
0.2738
0.1530
0.0437
0.5981
0.2177

Wavelengths (nm) included in model

275 325 425 775 825 875 1025 1175 1275 1425 1475 1875 1925 2225 2375
275 325 425 475 575 675 725 1025 1075 1225 1275 1475 1575 1825 2375 2425 2475
275 375 425 475 525 575 675 725 775 1225 1275 1425 1475
275 325 475 525 575 625 725 1075 1225 1425 1475 1875 1925 1975 2025 2075 2325 2375 2475
375 425 475 525 575 675 725 825 1325 1425 1525 1875 1925 2225 2325 2375 2475
425 475 725 775 1025 1125 1225 1275 1375 1425 1675 1775 2225 2275 2325 2475
275 325 375 425 525 575 625 675 825 1075 1175 1225 1325 1475 1625 1975 2025 2225 2275
375 475 625 675 725 1025 1125 1225 1275 1475 1525 1675 1875 2075 2175 2275 2375 2425 2475
275 425 525 575 725 975 1125 1375 1425 1575 1925 2225 2275 2375 2475
Table 4. Field 3 models relating soil properties to reflectance spectra.
Wavelengths (nm) included in model
525 725 925 1025 1175 1525 1575 2025 2125 2325
275 425 475 575 625 1375 1425 1475 2425 2475
275 325 375 425 475 525 575 625 775 825 1025 1075 1175 1225 1325 1625 1975 2325 2425 2475
375 425 525 675 725 825 1025 1125 1225 1425 1525 1575 2175
275 375 475 575 625 675 1025 1075 1375 1425 1475 1525 1675 1825 1875 1925 1975 2175 2225 2425
425 475 575 775 1025 1075 1325 1425 1475 1525 1575 1875 1975 2125 2175
375 475 525 575 1175 1325 1425
275 525 575 625 725 975 1075 1175 1275 1325 1425 1475 1525 1575 1675 1825 1975 2175 2275 2325
275 325 425 675 725 775 825 1325 1425 1625 2125 2225 2375

Further, when the best models for field 1 were applied to

field 3 and vice versa (table 5), lower R2 values were
obtained, except in the case of K in field 1 based on the model
from field 3. This anomaly is not important because the
difference in R2 values is small, the R2 values are very low,
the field1based model contained 19 variables as opposed
to the seven in the field3based model, and the
field1based model is not significant. On the whole, a
model developed on one field did not apply as well to the
other field. However, it is evident that the models developed
in one field were predictive for certain soil properties in the
other field, e.g., R2 = 0.63 for Ca and 0.68 for Mg in field 1
based on the model developed for field 3, and R2 = 0.50 for
clay in field 3 based on the model developed for field 1. In
addition, the correlations between all of these models and the
property of interest (except K in field 3) were statistically
significant at the 5% level.
EXPERIMENT 2: DIFFERENTIATING VARIABILITY IN THE
SPECTRAL REFLECTANCE DATA
Figures 3a and 3b are graphs of the total, withinsample,
and measurement CV for the randomly selected samples in
experiment 2. It is clear that certain sections of the spectrum
have more variation than others. It can also be seen that, as
was expected, measurement CV is less than withinsample
CV, which is less than total CV. However, at certain portions
of the spectrum, measurement CV approaches withinsample CV, and at certain portions withinsample CV approaches
total CV.
Figure 4 includes the ratios of nonwithinsample CV to
total CV and nonmeasurement CV to total CV. The height
underneath each curve indicates the portion of the variability
not attributable to the associated factor. For instance, at any
wavelength, the percentage under the curve of nonwithin
sample CV to total CV indicates the amount of total
variability remaining for discrimination of soilproperty
differences in the population. Figure 5 also includes the ratio

1450

of nonwithinsample CV to total CV, but after it has been

normalized by dividing the value at each wavelength by the
maximum value over all wavelengths (in this case, at
740 nm). It can be seen in figure 5 that the spectrum between
250 and 2500 nm can be divided into several sections with
regard to the ability to discriminate among the soil samples
studied. The region over which maximum discriminatory
power (90% to 100% of maximum for the ratio of total CV
minus withinsample CV to total CV) exists is between
roughly 500 and 800 nm. Regions of good discriminatory
power (80% to 90% of maximum) exist between 400 and 500
nm and between 950 and 1500 nm. Regions of moderate
discriminatory power (50% to 80% of maximum) exist
between 250 and 400, 875 and 950, 1500 and 2125, and 2275
and 2375 nm. All remaining regions (800 to 875, 2125 to
2275, and 2375 to 2500 nm) can be considered as having poor
discriminatory power (less than 50% of maximum). It bears
repeating that these regions are somewhat dependent on the
instrument used. The regions where measurement CV is high,
most notably the region around 800 nm, have poor discriminatory power largely because of instrument characteristics.

DISCUSSION
Ideally, one would find certain reflectance spectra that
could be measured and consistently related to soil properties
on a wide variety of agricultural fields. The work in this study
indicates that, while relationships between reflectance
spectra and the soil properties under consideration are
detectable, they are not so strong and consistent as to allow
realtime opticsbased mapping of soil chemical and
physical properties with the methods examined. This work
compares well to other studies (e.g., Ingleby and Crowe,
2000) that have reported inconsistencies between fields in the
relationships among soil properties and reflectance spectra.
It is possible that finer resolution than 50 nm or consideration

TRANSACTIONS OF THE ASAE

of other portions of the spectrum may be more promising for

certain soil properties. In addition, other optical measurement techniques besides diffuse reflectance could be investigated. One thing that appears clear is that more fundamental

research as to how soil constituents affect optical properties

is needed. Natural soils are very complex, and the solution for
realtime soilproperty measurement is likely to be complex
also.

Field
number

Clay models
3
1
250

500

750

1000

(a)

1250

1500

1750

2000

2250

2500

1750

2000

2250

2500

1750

2000

2250

2500

1750

2000

2250

2500

1750

2000

2250

2500

1750

2000

2250

2500

1750

2000

2250

2500

1750

2000

2250

2500

1750

2000

2250

2500

Wavelength (nm)

Field
number

Sand models
3
1
250

500

750

1000

(b)

1250

1500

Wavelength (nm)

Field
number

P models
3
1
250

500

750

1000

1250

1500

(c)
Wavelength (nm)

Field
number

K models
3
1
250

500

750

1000

(d)

1250

1500

Wavelength (nm)

Field
number

Ca models
3
1
250

500

750

1000

1250

1500

(e )
Wavelength (nm)

Field
number

Mg models
3
1
250

(f)

500

750

1000

1250

1500

Wavelength (nm)

Field
number

Zn models
3
1
250

500

750

1000

(g)

1250

1500

Wavelength (nm)

Field
number

Na models
3
1
250

500

750

1000

1250

1500

(h)
Wavelength (nm)

Field
number

pH models

(i)

3
1
250

500

750

1000

1250

1500

Wavelength (nm)

Figure 2. Comparison of model independent variables (spectral reflectance wavelengths) between fields.

Vol. 44(6): 14451453

1451

Coefficient of variation (%)

Table 5. Results for estimating soil properties in one field

with models developed with data from the other field.
R2
Soil property
Origin of model
Clay (field 3)
Sand (field 3)
P (field 3)
K (field 3)
Ca (field 3)
Mg (field 3)
Zn (field 3)
Na (field 3)

Field 1
Field 1
Field 1
Field 1
Field 1
Field 1
Field 1
Field 1

0.4271
0.1811
0.3795
0.1544
0.6263
0.6755
0.1897
0.2718
0.4440
0.4363

Clay (field 1)
Sand (field 1)
P (field 1)
K (field 1)
Ca (field 1)
Mg (field 1)
Zn (field 1)
Na (field 1)
pH (field 1)
pH (field 1)

Field 3
Field 3
Field 3
Field 3
Field 3
Field 3
Field 3
Field 3
Field 3
Field 1

90

Relative variability (%)

0.4980
0.3024
0.1545
0.0549
0.1567
0.0828
0.1426
0.1386

100

80
70
60
50
40
30
20
10
0
250

500

750

1000

1250

1500

1750

2000

2250

2500

Wavelength (nm)
(Total CV minus withinsample CV)/(Total CV)

(Total CV minus measurement CV)/(Total CV)

Figure 4. Ratios of amongsample CV to total CV.

55
50
45
40
35
30
25
20
15
10
5
0
250

500

750

1000

1250

1500

1750

2000

2250

2500

Wavelength (nm)
Total CV

Withinsample CV

Measurement CV

Figure 3a. Coefficient of variation of soilsample reflectance spectra.

Figure 5. Relative discriminatory power at various wavelength regions.

Coefficient of variation (%)

15

12

0
250

S
500

750

1000

1250

1500

1750

2000

2250

2500

Wavelength (nm)
Total CV

Withinsample CV

Measurement CV

Figure 3b. Detailed view of coefficient of variation of soilsample reflectance spectra.

CONCLUSIONS

Based on this work, the following conclusions were

drawn:
S 50nmband averages of reflectance spectra from 250 to
2500 nm were significantly correlated (at the 5% level) to
soil properties on the two fields in this study, but no single

1452

50nm bands were highly correlated with any soil property under consideration.
Soil nutrients were more highly correlated with multiple
reflectance spectra on one of the fields, while soil texture
properties were more highly correlated with reflectance
spectra on the other.
The only soil properties that had multipleregressor
models with R2 values greater than 0.50 were Ca and Mg
on one field and clay and pH on the other field.
Soil properties that were well correlated with reflectance
spectra on one field were not as well correlated on the
other.
Wavelengths included in the best correlation models were
not similar between fields, and models developed on one
field did not apply as well to the other field.
Certain sections of the spectrum have more variation than
others in terms of measurement error, sample
inhomogeneity, and differences among samples.
For the instrument used in this study, the spectrum
between 250 and 2500 nm can be considered in several
sections with regard to the ability to discriminate among
the soil samples studied; the regions of highest
discriminatory power were between 400 and 800 nm and
between 950 and 1500 nm.

TRANSACTIONS OF THE ASAE

REFERENCES
Barnes, E. M., and M. G. Baker. 1999. Multispectral data for soil
mapping: possibilities and limitations. ASAE Paper No.
991138. St. Joseph, Mich.: ASAE.
Barnes, E. M., M. S. Moran, P. J. Pinter, Jr., and T. R. Clarke. 1996.
Multispectral remote sensing and sitespecific agriculture:
Examples of current technology and future possibilities. In Proc.
3rd Intl. Conf. Precision Agric., 845854. Madison, Wisc.:
ASACSASSSA.
Baumgardner, M. F., L. F. Silva, L. L. Biehl, and E. R. Stoner.
1985. Reflectance properties of soils. Advances in Agronomy 38:
144. Orlando, Fla.: Academic Press, Inc.
BenDor, E., and A. Banin. 1995. Nearinfrared analysis as a rapid
method to simultaneously evaluate several soil properties. Soil
Sci. Soc. Am. J. 59: 364372.
Coleman, T. L., P. A. Agbu, and O. L. Montgomery. 1993. Spectral
differentiation on surface soils and soil properties: Is it possible
from space platforms? Soil Sci. 155(4): 283293.
Dematt, J. A. M., and G. J. Garcia. 1999. Alteration of soil
properties through a weathering sequence as evaluated by
spectral reflectance. Soil Sci. Soc. Am. J. 63: 327342.
Krishnan, P., B. J. Butler, and J. W. Hummel. 1981. Closerange
sensing of soil organic matter. Trans. ASAE 24(2): 306311.
Moran, M. S., Y. Inouye, and E. M. Barnes. 1997. Opportunities
and limitations for imagebased remote sensing in precision
crop management. Remote Sens. Environ. 61: 319346.
Muttiah, R. S., G. Lawson, R. Srinivasan, and J. Atwood. 1998.
Impact of satellite resolution on cropland feature discrimination.
ASAE Paper No. 983077. St. Joseph, Mich.: ASAE.
Myers, R. H. 1990. Classical and Modern Regression with
Applications. 2nd ed. Boston, Mass.: PWSKent.

Vol. 44(6): 14451453

PalaciosOrueta, A., and S. L. Ustin. 1998. Remote sensing of soil

properties in the Santa Monica Mountains: I. Spectral analysis.
Remote Sens. Environ. 65: 170183.
Pitts, M. J., J. W. Hummel, and B. J. Butler. 1986. Sensors utilizing
light reflection to measure soil organic matter. Trans. ASAE
29(2): 422428.
Research Analytical Laboratory. 2000. Textural analysis:
Hydrometer. Available at: http://ral.coafes.umn.edu/soil.htm.
(Research Analytical Laboratory, Soil Methods Page). St. Paul,
Minn.: Univ. of Minnesota.
SAS. 1999. SAS Users Guide. Cary, N.C.: SAS Institute, Inc.
Shibusawa, S., M. Z. Li, K. Sakai, A. Sasao, and H. Sato. 1999.
Spectrophotometer for realtime underground soil sensing.
ASAE Paper No. 993030. St. Joseph, Mich.: ASAE.
Sudduth, K. A., and J. W. Hummel. 1991. Evaluation of reflectance
methods for soil organic matter sensing. Trans. ASAE 34(4):
19001909.
_____. 1993a. Portable, nearinfrared spectrophotometer for rapid
soil analysis. Trans. ASAE 36(1): 185193.
_____. 1993b. Soil organic matter, CEC, and moisture sensing with
a portable NIR spectrophotometer. Trans. ASAE 36(6):
15711582.
_____. 1996. Geographic operating range evaluation of a NIR soil
sensor. Trans. ASAE 39(5): 15991604.
ViscarraRossel, R. A., and A. B. McBratney. 1998. Laboratory
evaluation of a proximal sensing technique for simultaneous
measurement of soil clay and water content. Geoderma 85:
1939.
Walpole, R. E., and R. H. Myers. 1993. Probability and Statistics
for Engineers and Scientists. 5th ed. New York, N.Y.:
MacMillan.

1453

1454

TRANSACTIONS OF THE ASAE