Fusion of Spectral Data From Multiple Handheld Analyzers (LIBS, XRF, and Raman) For Chemical Analysis and Classification of Soils
Fusion of Spectral Data From Multiple Handheld Analyzers (LIBS, XRF, and Raman) For Chemical Analysis and Classification of Soils
Fusion of Spectral Data From Multiple Handheld Analyzers (LIBS, XRF, and Raman) For Chemical Analysis and Classification of Soils
To search for other technical reports published by ERDC, visit the ERDC online library
at https://erdclibrary.on.worldcat.org/discovery.
ERDC/CRREL CR-20-1
September 2020
Chandra S. Throckmorton
Signal Analysis Solutions, LLC
22 Piney Grove Road
Bahama, NC 27503
Final report
Approved for public release; distribution is unlimited.
Abstract
An 18-month multidisciplinary project was undertaken by JRPlumer &
Associates, LLC and four subcontractors that had three technical objec-
tives: (i) to upgrade current handheld technology for chemical analysis by
X-ray fluorescence spectroscopy (XRFS), Raman spectroscopy (RS), and
laser-induced breakdown spectroscopy (LIBS); (ii) to design a multisensor
system based on these technologies for the rapid, in-situ chemical analysis
of soils and other materials of military interest; and (iii) to investigate the
classification/discrimination performance benefit that might be achieved
through advanced signal pre-processing and data fusion with XRFS, RS,
and LIBS analyses acquired for four suites of natural soils. Accomplish-
ments of the program in the latter area are described in this report.
DISCLAIMER: The contents of this report are not to be used for advertising, publication, or promotional purposes. Ci-
tation of trade names does not constitute an official endorsement or approval of the use of such commercial products.
All product names and trademarks cited are the property of their respective owners. The findings of this report are not to
be construed as an official Department of the Army position unless so designated by other authorized documents.
DESTROY THIS REPORT WHEN NO LONGER NEEDED. DO NOT RETURN IT TO THE ORIGINATOR.
ERDC/CRREL CR-20-1 iii
Contents
Abstract.................................................................................................................................... ii
Preface ..................................................................................................................................... vi
1 Introduction ...................................................................................................................... 1
1.1 Objective(s) ........................................................................................................ 1
1.2 Approach ............................................................................................................ 1
2 Background ...................................................................................................................... 3
2.1 Instrumentation ................................................................................................. 3
2.2 Soil ...................................................................................................................... 6
3 Methods .......................................................................................................................... 10
3.1 Background ...................................................................................................... 10
3.2 Soil sample preparation and analysis ............................................................11
6 Summary ......................................................................................................................... 42
References............................................................................................................................. 43
Acronyms ............................................................................................................................... 56
Figure 17. Classification accuracy for the CSU classification schema based
on consideration of spectra for Raman spectroscopy - RS (1), X-ray
fluorescence spectroscopy - XRFS (2), laser-induced breakdown
spectroscopy - LIBS (3), fused spectra for RS + XRFS (4), RS + LIBS (5), and
XRFS + LIBS (6), RS + XRFS + LIBS spectra (7), fused spectra + XRF
quantitative analyses (8), and fused spectra + XRF quantitative analyses +
laboratory chemical analyses (9). ...................................................................................... 36
Figure 18. Independent feature accuracy per CSU soil classification schema.
Features are sorted on a per schema basis, greatest accuracy to least accuracy.
The legends indicate classification performance that could be achieved by always
selecting the class with the most samples. .............................................................................. 37
Figure 19. Principal component analysis score plots for XRFS (upper left),
LIBS (lower left), RS (upper right), and chemistry (lower right) for the soil
series of CSU soil suite represented by multiple samples ............................................ 38
Figure 20. Principal component analysis scores plot from laboratory analysis of
the Lebanon/CRREL and Belmont soil sets. ......................................................................... 39
Figure 21. Principal components scores plot from laboratory analysis of the
Belmont soil pits. ...................................................................................................................... 40
Figure 22. Classification accuracy for the New Hampshire soil sets from
Belmont/Bean Hill (left) and Lebanon/CRELL (right) ........................................................... 41
Figure 23. Principal component analysis scores plot for fused XRFS and LIBS
spectral data sets for in-situ analysis of the three Belmont/Bean Hill soil pits. ................ 41
Tables
Table 1. Comparative capabilities of SciAps handheld XRFS, RS, and LIBS
analyzers. ................................................................................................................................... 5
Table 2. Perceived advantages of an integrated multi-analyzer system. ..................... 5
Table 3. Number of samples per class for each collection date. ......................................... 29
Table 4. Classification schema for CSU soil suite characterization. ................................. 35
ERDC/CRREL CR-20-1 vi
Preface
This study was conducted for ERDC T-15 Program under 474428, “Sup-
port of US Army Global Military Objectives: LIBS for Military Surveys.”
Parts of this program were funded under ERDC contract W913E518C0011
issued to JRPlumer & Associates LLC. The technical monitor was Dr. Jay
Clausen, CEERD-RN.
The Commander of ERDC was COL Teresa A. Schlosser and the Director
was Dr. David W. Pittman.
ERDC/CRREL CR-20-1 1
1 Introduction
An 18-month multidisciplinary effort was undertaken by JRPlumer & As-
sociates, LLC with three technical objectives: (i) to upgrade current com-
mercial handheld technology for chemical analysis by X-ray fluorescence
spectroscopy (XRFS), Raman spectroscopy (RS), and laser-induced break-
down spectroscopy (LIBS); (ii) to design a multianalyzer system based on
these technologies for the rapid, in-situ chemical analysis of soils and
other materials of military interest; and (iii) to investigate the classifica-
tion/discrimination performance benefit that might be achieved through
signal pre-processing and data fusion using XRFS, RS, and LIBS analyses
of natural soils. This initiative was supported by subcontracts from
JRPlumer & Associates, LLC to SciAps, Inc., Signal Analysis Solutions,
LLC, SoilHydrology Associates, LLC, and Applied Spectra, Inc. Accom-
plishments of this multidisciplinary R&D program pertain to three tech-
nical areas: Technology Development and Design, Soil and Military
Material Analysis; and Signal Processing, Chemometric Analysis, and Data
Fusion. Details of work completed under the first two program areas are
commercial proprietary, so this report describes the results only from the
third program area.
1.1 Objective(s)
A central hypothesis of the multisensor concept is that integration of three
chemical analyzers with largely complementary capabilities would result in
an operational performance surpassing that achievable by any of the indi-
vidual handheld analyzers. Thus, the overarching program objective was
to evaluate this hypothesis.
1.2 Approach
The program vision was that a multianalyzer capability would permit dis-
crimination of materials of interest presently beyond the capability of con-
temporary field-portable analyzers and at a price point significantly lower
than laboratory instrumentation. Thus, spectral data preprocessing and
fusion of the processed data streams from the individual analyzers were
studied to assess the extent to which classification/discrimination perfor-
mance could be improved. Anticipated performance enhancement would
ERDC/CRREL CR-20-1 2
A soil analysis component of the program was needed to baseline the per-
formance of the three individual commercial analyzers and to provide the
data needed for the signal processing and data analytics research. Four
suites of natural soils were acquired and then analyzed using the commer-
cial XRFS, RS, and LIBS analyzers individually. Various approaches to the
pre-processing of the LIBS data, which is characterized by high shot-to-
shot spectral intensity variation, were examined. Then, the processed data
for each sample analyzed was fused with the XRFS and RS data streams for
that samples, with the results for the fused results for the integrated ana-
lyzer data streams compared with that from the individual analyzers.
ERDC/CRREL CR-20-1 3
2 Background
2.1 Instrumentation
Until quite recently, chemical analysis for material characterization has
been possible only in the laboratory. Consequently, laboratory analysis
of materials collected in the field, typically tends to be limited to a small
number of samples that are analyzed using specialized instrumentation
such as inductively coupled plasma mass spectrometry, high performance
liquid chromatography, or gas chromatography. These processes are labor
intensive, time-consuming, and typically very costly because individual
samples must be collected and packaged on-site in the field, transported to
the laboratory, and then processed for analysis on what can be very expen-
sive instrumentation. However, with the recent advent of portable single
sensor instrumentation for chemical analysis in the field, a technological
opportunity existed to create a novel and unique field characterization and
forensic capability for the rapid, in-situ analysis of both soil and a wide
range of materials of military interest undisturbed in the field under ambi-
ent environmental conditions.
Each of these three technologies for chemical analysis – XRFS, RS, and
LIBS, are presently commercially available as handheld analyzers for use
outside the laboratory (Figure 1) and have common performance attrib-
utes like minimally-destructive analysis and rapid data collection and pro-
cessing. It is well recognized that these different analytical technologies
have different elemental sensitivities and analytical accuracies. However,
each modality has its own unique functional competency (Table 1), such as
differential sensitivity to a wide range of elements, point versus area analy-
sis, surface analysis versus depth profiling and analysis of elemental versus
ERDC/CRREL CR-20-1 5
There is strong potential for technology transition spinoff for other appli-
cations. For example, one such possibility would be its use for field foren-
sics by both the civil and military law enforcement communities.
2.2 Soil
To fully understand and exploit the analytical benefit a field-portable mul-
tianalyzer capability, it is important to understand the chemistry of target
of interest for this program – natural soil.
Eight chemical elements (O, Si, Al Fe, Mg, Ca, Na, & K) comprise the
bulk of the inorganic component of soils. Ions of these elements combine
in various ratios to form different minerals. Another 80+ elements natu-
rally occur in soil in minor and trace quantities and still others can be pre-
sent as anthropogenic contaminants. Soils are chemically different from
the rocks and minerals from which they form in that soils contain less of
the water-soluble weathering products, Ca, Mg, Na, and K and more of the
relatively insoluble elements such as Fe and Al. Old, highly weathered soils
normally have high concentrations of aluminum and iron oxides. The or-
ganic fraction of the near-surface soil, which is composed predominantly
of, carbon, hydrogen, oxygen, nitrogen and smaller quantities of sulfur
and other elements, usually comprises <6% of the soil mass by weight and
much less at depth. Yet, despite its small quantity it has a great influence
on soil chemical properties. The organic fraction enhances soil aggregation
ERDC/CRREL CR-20-1 7
and structure, enhances soil water retention, and increases cation ex-
change capacities, while also serving as a reservoir for the plant essential
nutrient elements (e.g. N, P, and S).
Soils are classified based on physical and chemical properties in their dif-
ferent horizons (e.g. composition, color, texture, structure, etc.) from the
surface to two meters depth. Currently, the NRCS soil database for the
United States
(https://www.nrcs.usda.gov/wps/portal/nrcs/detail/soils/survey/class/da
ta/?cid=nrcs142p2_053586) contains more than 20,000 detailed soil se-
ries descriptions that reflect the wide range of the five soil-forming factors
across the country. The NRCS soil taxonomy establishes hierarchies of
classes to permit the relationship among soils and between soils, and the
factors responsible for their character, to be understood. The soil series is
the lowest category within soil taxonomy. All soils within a single series
have uniform differentiating characteristics and arrangement of horizons.
This does not mean that all soils within a series are identical; it does mean
that they have a similar sequence of horizons, but the horizons may be of
different thickness, color, structure, within prescribed limits. All the soils
within a series will have developed in the same kind of parent material
with comparable drainage characteristics and will be of similar age. The ef-
fects of climate and biological activity will have been very similar. Conse-
quently, the soils within a series exhibit like properties and respond in like
fashion to usage or manipulation.
The single most important factor determining the accuracy of any soil map
is the map scale. Most intensive agricultural areas have soil maps at a scale
of 1:20,000 or less. All other areas, also known as rangeland areas, have
soil maps at a scale of 1:50,000 or greater. The smallest unit that can be
defined on a map at this scale is approximately 20 acres.
ERDC/CRREL CR-20-1 8
It is important to note that the method for mapping soils in rangelands ar-
eas of the western US differs from that in more intensively farmed regions.
In agricultural areas, soil map unit boundaries are identified primarily by
the change in soil properties determined from soil auguring and, more re-
cently, remote sensing tools such as electromagnetic induction meters. By
contrast, rangeland soil maps are based largely on the interpretation of
aerial photographs to identify proxy data that has been determined reflect
changes in soil properties, such as changes in landform, parent materials,
vegetation, etc. The initial identification of soil map units is validated with
a very small number of soil observations, commonly 4-5/km2. At the
rangeland scale, many soil map units are described as either complex or
compound Soil Mapping Units (SMU), compound SMU’s predominantly
consist of two Soil Series whereas complex SMU’s can contain more than
two Soil Series. Furthermore, in the description of the map units there is
no indication how the two or more Soil Series are distributed within the
unit.
Unit boundaries on all soil maps are shown as fixed (or definite) bounda-
ries that indicate the immediate change from one Soil Series to the adjoin-
ing Soil Series. However, two very different types of soil map unit
boundaries can be observed in the field - well-defined boundaries and gra-
dational boundaries. Well-defined boundaries are those where soil proper-
ties abruptly change across a landform or due to a change in lithology.
Such boundaries are easily identified by landform or lithological relation-
ships and the boundary does not change over time. By contrast, grada-
tional boundaries can change over time and may change according to the
soil properties being considered. Most of the soil map unit boundaries for
maps at a scale of 1:50,000 and above are boundaries based on landform
and/or lithological changes.
The NRCS has identified and mapped the over 20,000 different Soil Series
across the United States at different levels of classification from Family, to
Subgroup, Great Group, Suborder and Order. The NRCS soil taxonomy di-
vides natural soils into 12 Orders that exhibit only small differences in the
kinds and relative strengths of processes that tend to develop their internal
soil horizons: Alfisols (10% - in semiarid to moist areas), Andisols (1% - in
cool areas with moderate to high rainfall, Aridisols (12% too dry for meso-
phytic plants), Entisols (16% - little to no pedogenic development), Geli-
sols (9% - permafrost near the soil surface), Histosols (1% - high organic
content without permafrost), Inceptisols (17% - moderate development in
ERDC/CRREL CR-20-1 9
Thus, a major challenge for this project was how to select a representative
suite of the most common soils that might be encountered by the military
during operations in the field. One option was to select soils that represent
the 12 soil orders of the USDA soil taxonomy. A Soil Order is the category
with the highest level of generalization and abstraction, and each Soil Or-
der is very heterogeneous with respect to its physical and chemical proper-
ties that are not considered in the differentiation (NRCS 1999). As a
consequence, knowing only the Order of a soil is not sufficient to allow for
scientific repeatability of compositional measurements. Soil Series are the
most homogeneous classes in the system of taxonomy of the NRCS. They
are the lowest and most homogeneous category of the national soil classifi-
cation system and provide a detailed record of soil properties needed to
prepare soil evaluations for different applications in agriculture, construc-
tion, roads, etc. An official NRCS Soil Series description includes infor-
mation on location, taxonomic classification, a detailed soil profile
description, location of the typical soil profile, the range of descriptive
characteristics, competing series, geographic setting, geographically asso-
ciated soils, drainage and permeability, use and vegetation, distribution
and extent, and additional data. Therefore, a soil sample from an officially
recognized Soil Series taken at the “location of the typical soil profile”, rep-
resents a unique soil that is readily accessible, should additional sampling
be needed in the future for further testing and confirmation. However, it
was not possible to select a small set of representative soils from the
25,000+ soil series in the United States to cover the full range of soil char-
acteristics because any stratification scheme will necessarily fall short.
ERDC/CRREL CR-20-1 10
3 Methods
3.1 Background
JRPA undertook a capabilities performance investigation of the current
COTS handheld analyzers for XRFS, RS, and LIBS to understand perfor-
mance through a detailed program of comparative detection and quantifi-
cation of chemical elements and molecular species in soils. Two types of
soils were used for this purpose: (1) Standard Reference Materials (SRMs)
and (2) natural soils from across North America. A listing of these soils
and standards is provided in the on-line archive for this project in the
JRPA Multisensor Project folder on the Open Science Framework tool
(https://osf.io/z3wgk/). Although SRMs can serve as consistent calibra-
tion points for sensor performance, previous work by Hendrickx and his
colleagues on the detection of improvised explosive devices (e.g., Das et al.
2001; Miller et al. 2004; Van Dam et al. 2004; Hendrickx et al. 2006) un-
ambiguously corroborated that more will be learned from sensor develop-
ment research using representative field soils than commercially available
standard materials. The latter are unrepresentative of the wide range of
natural soils that are encountered in the field. Thus, test results in these
media will not necessarily give a realistic evaluation on the potential and
limitations of a novel sensor for field soil analysis.
A total of 272 natural soils were acquired from four sources for the project
analytical work. These are follows: (i) 50 rangeland soil samples from cen-
tral New Mexico, (ii) 21 soil samples from two sites in west-central New
Hampshire, (iii) a suite of 144 agricultural soils from across the United
States and Canada provided by Colorado State University, and (iv) 58 soils
from small arms ranges on military installations in Massachusetts, Vir-
ginia, Georgia, Idaho, Washington, and Alaska provided by USACE-ERDC-
CRREL. Each soil suite is described in the Appendix to this report. To-
gether, these soil suites met the following criteria:
(ii) The soils chosen should represent different climatic settings across the
country (from semi-arid to humid, because climate is an important soil
forming factor);
ERDC/CRREL CR-20-1 11
(iii) The soils should represent different geologic parent materials, because
parent material also is an important soil forming factor;
(iv) The soils should represent different ages because time is another im-
portant soil forming factor;
(v) The soil should be relatively thick and have distinguishable A and B ho-
rizons, so that well-defined, homogeneous samples can be collected; and
(vi) The soils should represent more than half of the twelve soil orders.
XRFS and LIBS spectra were preprocessed using z-scoring (mean subtrac-
tion and standard deviation division) normalization before further pro-
cessing and fusion. XRFS spectra were truncated to be the same length
and LIBS spectra had Ar spectral lines removed and baseline corrected. No
normalization was applied to the XRFS quantitative data and RS spectra
were utilized as collected.
Examples of typical XRFS, RS, and LIBS spectra obtained during this pro-
ject are presented in Figures 2-4, in this case for the playa lake soil at the
White Sands, NM sampling locality. These three figures illustrate clearly
that each of the handheld analyzers not only records similar compositional
information, as expected, but also can contribute unique information
about the composition of a soil sample.
ERDC/CRREL CR-20-1 13
Figure 2. Example XRFS spectrum for the playa soil from White Sands, NM, showing
the presence of Ca, Sr, S and Fe.
Figure 3. Raman spectrum for the playa soil from White San.
Figure 4. LIBS broadband spectrum for the playa soil from White Sands, NM,
showing presence of Ca and Sr together with the minor presence of Si and Na.
ERDC/CRREL CR-20-1 14
Due to the calibration process for XRFS, the spectra are often of slightly
different lengths. Machine learning algorithms cannot process feature vec-
tors of different length. Therefore, all XRFS spectra were truncated to en-
sure equal lengths. In order to ensure the classifier is not erroneously
trained to make decisions based on relative intensity of XRFS, RS, and
LIBS spectra, each intensity spectrum was normalized by subtracting the
mean and dividing by the standard deviation.
The LIBS data collection process using an inert gas to enhance plasma inten-
sity results in non-trivial emission line magnitudes at wavelengths of the Ar
purge gas, regardless of the possible presence of Ar within the sample be-
ing interrogated. If these emission lines are constant or random across the
identified spectral classes, then the partial least squares discriminant anal-
ysis (PLSDA) classifier should apply a zero or near-zero weight to the
ERDC/CRREL CR-20-1 15
As recorded, LIBS spectra contained emission lines for all wavelengths be-
tween 180 and 961 nm. With Ar removal, only emission lines in the ranges
of 190-675, 765-771, and 776-779 nm were retained. This process reduced
LIBS spectra from 23,431 emission lines to 14,823 emission lines. The im-
pact of argon removal was tested on several non-soil-based classification
tasks, and no negative impact was observed.
Although baseline drift was not observed to any great extent in the LIBS
spectra for the soils analyzed during this project, substantial baseline drift
was observed in some older LIBS data sets acquired with the same SciAps
handheld analyzer. Thus, there is the potential that such variations in base-
line level could confound the quality assessment of spectra (e.g. the calcu-
lation of signal-to-noise ratios) and negatively impact classification
performance. Therefore, a baseline correction algorithm was developed
that employs a 2-stage process that first removes the sparse peaks in the
LIBS spectrum in order to estimate the baseline response, and then sub-
tracts the estimated baseline from the original LIBS spectrum. In order to
remove the peaks in the LIBS spectrum, a Hampel filter (Davies and
Gather 1993) with a window size of 2% of the spectrum length was ap-
plied. For each sample, if the emission line magnitude was more than
three standard deviations from the filter window’s median, it was replaced
by the median. Once the peaks were removed from the spectrum, the base-
line of the spectrum was estimated by smoothing the residual with a me-
dian filter that was 0.5% of the spectrum length. This smoothed
baseline estimate was then subtracted from the original spectrum. The
baseline correction algorithm was added to the standard LIBS prepro-
cessing workflow.
ERDC/CRREL CR-20-1 16
Two methods SNR estimation with low computation loads were evaluated:
(i) the log of the peak spectrum value (Log-Max) and (ii) the ratio of the
peak spectrum value to the median spectrum value (Peak-To-Median).
Although the log of the peak spectrum value technically only estimates the
signal strength, LIBS spectra with low signal strength tend to be spectra
with low SNR. In addition to the two methods with low computation load,
two methods with higher computation loads were considered. In the first
method (Histogram-Split), an intensity histogram of the spectrum was
generated, and between class variance was used to separate spectrum
magnitudes into ‘noise’ and ‘signal’ values (Otsu 1979). Using these groups
of magnitudes, the signal energy and noise variance were estimated, and
SNR was calculated. In the second method (Gaussian-Outlier), a Hampel
filter as described above for baseline correction was used to remove the
peaks of the spectrum. In this method, the residual was used to estimate
noise mean and variance. All spectrum magnitudes that were three stand-
ard deviations from the noise mean were considered part of the signal and
used to estimate the signal energy. With the estimate of the signal energy
and noise variance, the SNR was calculated.
Four methods were compared for several different LIBS data sets. Classifi-
cation accuracy was estimated as increasing proportions of LIBS spectra
were culled based on thresholding of the four metrics. For each metric, a
limit was placed on culled spectra such that no sample’s spectra were en-
tirely discarded. The two low-computation-load methods were consist-
ently among the best performers across data sets. The Histogram-Split
approach tended to be a consistently poor performer and the Gaussian-
Outlier method performed poorly for one of the data sets. Of the four SNR
estimation metrics compared, the ratio of the normalized maximum spec-
trum value to the median spectrum value (Peak-to-Median method) most
often achieved the greatest improvement in classification performance
through spectra culling (removing spectra with lower SNR estimates).
Also, the Peak-to-Median method most often achieved increases in accu-
racy more rapidly than the other methods. This indicates a better estimate
of SNR than the other methods since the spectra that most impact classifi-
cation performance are identified earlier (i.e. the worst spectra are given
ERDC/CRREL CR-20-1 18
the lowest SNR estimates). The Peak-to-Median method was second only
to one other method in terms of calculation speed, indicating viability for
real- time use.
The potential benefit of spectra culling was tested on two geological data
sets. Spectra were culled based on thresholding of either the SNR estimate
or the similarity measure. The remaining spectra were classified, the confi-
dences were averaged per sample, and a final label assigned based on the
averaged confidences. This process was repeated for a range of thresholds
ranging from no culling to the maximum number of spectra that could be
culled without discarding any samples. Little benefit appeared to be gained
from culling outlier spectra if no samples are discarded in their entirety. Ra-
ther, any potential benefit of evaluating spectra quality will likely be in
terms of determining whether an entire sample should be remeasured.
Given these results, this process was not added to the standard prepro-
cessing workflow. However, there is potential for this method to be used
as a tool to evaluate a data collection in real-time.
4.1.6 Similarity
with itself. Perfect similarity is one for correlation and zero for the similar-
ity index and spectral contrast angle. Some of the 14 soil pellets show
strong heterogeneity in this spectral variability analysis indicating textural
heterogeneity within the pellet.
on Hempstalk et al. (2008) that uses available spectra from each analyzer
and class to estimate the mean and standard deviation at each wavelength.
With this statistical model, a number of simulated spectra matching the
number of measured spectra are generated. The outlier rejection, i.e.
NOTA classifier, is trained to set a boundary between the measured spec-
tra (not an outlier) and the simulated spectra (outliers). This process will
set a boundary as tightly as possible to the measured spectra (Figure 6).
The more accurate the statistical model, the tighter the outlier rejection
boundary. Thus, this process requires a significant amount of measured
data to ensure an accurate boundary.
Figure 6. Difference between typical classification (left) and outlier rejection (right).
Two cases were considered. In the first, the non-NOTA data set was the
New Mexico soil suite, and the NOTA data sets were the CSU soils, NIST
standards, OREAS standards, and military installation soils. In the second,
the non-NOTA data set was the military installation soil suite, and the
NOTA data set consisted of the New Mexico soils, CSU soils, NIST stand-
ards, and the OREAS standards. Figure 8 shows an example of percent re-
jections for the different data sets as a function of the NOTA classifier
operating points. The goal is to select an operating point that minimizes
the rejection non-NOTA correctly classified samples while maximizing the
rejection of NOTA samples. For example, selecting an operating point of
0.65 in Figure 8 would incorrectly reject approximately 9% of correctly
classified military installation soil samples while correctly rejecting 60-
97% of the NOTA samples, depending on the NOTA data set. In addition,
75% of the incorrectly classified non- NOTA samples would be rejected by
the outlier rejection classifiers.
ERDC/CRREL CR-20-1 22
Figure 8. XRFS outlier rejection rates for different data sets when the non-NOTA
data set is the military installation soil suite.
In Figure 9, the optimal operating point (assuming the cost of the two
types of errors is equal) is selected for each test case and sensor and the
proportion of rejections at that operating point is plotted. For the LIBS
and XRFS data, between 70% and 100% of NOTA samples are rejected, de-
pending on sensor and test case. Between 2% and 15% of non-NOTA sam-
ples were rejected. These results indicate the potential for outlier rejection
with these sensors. While the approach was able to reject 60% of NOTA
samples while rejecting 5% of non-NOTA samples for the New Mexico
soils test case using the RS data, all non-NOTA samples were rejected be-
fore any NOTA samples for the military installation soils test case. This is
likely due to the poor classification performance for the RS analyzer with
the military installation soils. Since it is unable to determine an accurate
model of the non-NOTA data, it is also unable to determine data that do
not match those models. For the military installation soil RS data, it was
not possible to reject any NOTA samples without rejecting all non-NOTA
samples (hence the selected operating point to minimize error cost was a
point that rejected no samples).
ERDC/CRREL CR-20-1 23
Figure 9. Rejection rates at the optimal operating point for which the costs of the two
types of errors (failing to reject NOTA samples and rejecting correctly classified non-
NOTA samples) were assumed equal.
The software for real-time data processing developed for the project has
the following features:
This capability was field tested during in August 2019, when the data
streams from the three analyzers were in real time wirelessly transmitted
during analysis to a laptop computer.
During the project, three reference cases were added to the software for
real-time estimation of class confidences for any unknown sample: New
Mexico soils, military installation soils, and New Hampshire soils. For the
latter, only LIBS processing is possible due to the limited data collection.
For the first two, individual data processing was possible for XRFS, RS,
and LIBS data, as well as data fusion processing data from all three analyz-
ers. At the conclusion of the project, the software polls an input directory
into which the user places sensor spectra CSV files. The software automati-
cally determines the analyzer type of the new data, processes the data with
the appropriate spectra classifier, and displays a confidence that the sam-
ple belongs to each of the three classes in the library. Once data from all
three analyzers has been added, the spectra classifier outputs are fused
and confidence that the sample belongs to each class is displayed. Once
the user has completed analysis of the sample, the user selects an archive
button to send the data and a report of the results to a time-stamped
folder. Processing is then cleared and ready for data from a new sample.
ERDC/CRREL CR-20-1 27
parent material, climate, flora and fauna, and topography) could be dis-
criminated. These samples were analyzed by JRPA in January, March, Au-
gust, and December 2019 to evaluate the benefits of sensor fusion
performance through multiple analysis of the same sample suite.
The 5-class schema for New Mexico soils was used - Rio Salada, Sedillo
Hill, Socorro Canyon Upper/Lower, White Sands, and Willard. The six soil
pellets from each locality were measured on four collection dates (Decem-
ber 2018, January 2019, March/April 2019, and August 2019). Spectral fu-
sion was first undertaken for the January 2019 data collection using the
spectral data from the three analyzers (XRFS, RS, and LIBS), first individ-
ually, then LIBS with XRFS, and finally the spectral data for all three ana-
lyzers.
The result of this fusion exercise is shown in Figure 12, where the classifi-
cation accuracy for RS is seen to approach 80%, that for LIBS better than
85%, and that for XRFS almost 90% for the January 2019 data collec-
tion. So, it is not surprising that the accuracy for 2- or 3-analyzer fusion
was better than 90%. For the March/April 2019 data collection, this di-
minished to just under 75% for RS, but increased to better than 95% for
both LIBS and XRFS. By comparison, accuracy for dual analyzer fusion
involving RS was lower than for January 2019, that for LIBS plus XRFS
was improved, and that for the fusion of spectral data for all three sensors
was marginally only diminished. Fusion of the results from both dates ex-
hibited full classification accuracy for the 3-analyzer fusion.
ERDC/CRREL CR-20-1 29
Figure 12. Classification accuracy for the New Mexico soil suite based on
consideration of spectra for XRFS, RS, and LIBS; fusion of spectra for RS + XRFS,
LIBS + XRFS; and for RS + LIBS + XRFS for spectral data collected in Jan 2019,
Mar/Apr 2019, and for data fusion from both dates.
In addition, two LIBS systems were used in each of these data collections.
The number of samples per class varied per data collection (Table 3) with
an incomplete set of samples analyzed in December 2018 and August
2019, so that cross-validation analysis for these collection dates is not pos-
sible.
Aug 2019
Dec 2018 Jan 2019 Mar 2019
(*LIBS SciAps)
Rio Salada 2 14 6 12
Sedillo Hill 2 8 6 6*
Socorro Canyon 11 11 12*
White Sands 6 6 6
Willard 6 6 6*
Data analysis early in the project suggested that training and testing on
separate collection dates resulted in poor classification performance for
LIBS data. This observation raised the following questions:
Figure 13. Single analyzer classification results for three different methods of
training.
2019 collection dates ranged from 82% to 93% correct. Pooling the data
across all collection dates and using cross-validation resulted in 95% cor-
rect accuracy. There are no cross-validation results for Dec 2018 since the
data set is incomplete. Testing the Jan 2019 data against classifiers trained
with the Mar/Apr2019 or Aug 2019 data sets results in an average accu-
racy of 49% correct, with little
If, however, the Jan 2019 data are classified using a classifier trained on
both the Mar/Apr 2019 and Aug 2019 data sets, accuracy increases to the
cross-validation accuracy of 82% correct. The classification based the
quantitative XRFS analyses is thoroughly robust to collection date,
whereas results from all three analyzers are all negatively impacted by
training and testing across collection dates. Pooling across collection dates
mitigates this issue to varying degrees. The same methods of classification
accuracy estimation were used for the multiple sensor analysis as for the
single-sensor test (cross-validation, train/test across individual dates of
collection, and train/test with pooled collection dates). Two fusion cases
were considered: (i) RS + XRFS + LIBS and (ii) XRFS + LIBS. The spectral
data acquired by the two LIBS analyzers (JRPA and SciAps) were kept sep-
arate such that each collection date resulted in two sets of accuracy results.
Figure 14 shows results for fusing XRFS and LIBS sensor data. Cross-vali-
dation results are consistently near 100% correct. As with the individual
sensors, training with pooled data tends to result in higher accuracy than
training with an individual collection date. Figure 15 shows results for fus-
ing XRFS, RS, and LIBS data. Results are similar to those for the single-
analyzer results for just XRFS or LIBS. Ignoring the Dec 2018 collection
date, accuracy for the 3-analyzer fusion with pooled collection date train-
ing appears to be marginally more robust across collection dates than for
the XRFS and LIBS sensor fusion. It is important to note that each collec-
tion date has a small number of samples, so it is possible that the impact of
pooling is not just due to capturing variability across dates of collection. It
might also be due to the increase in the overall amount of information.
ERDC/CRREL CR-20-1 32
Figure 14. Classification results for XRFS and LIBS fusion for three different methods
of training. For each data collection, XRFS data is paired with that from either the
JRPA LIBS analyzer or the SciAps LIBS analyzer.
Figure 15. Classification results for fusing the Raman, XRFS, and LIBS sensors using
three different methods of training. For each data collection, Raman and XRF are
paired with either the JRPA LIBS sensor or the SciAps LIBS sensor.
Figure 16. Classification accuracy for the military installation soils based on RS,
XRFS, and LIBS data; fused 2-analyzer data, and fused 3-analyzer data.
A single data for sets for the sets of 10 surface soils at small arms ranges on
six military installations were acquired in Mar/Apr 2019. A 6-class classifi-
cation schema was used with the following classes: MMR-Joint Cape Cod
(MA), Ft. Eustis (VA), Ft. Benning (GA), ING Camp Kimama (ID), Ft.
Lewis (WA), and Ft. Wainwright (AK). The result of this fusion exercise is
shown in Figure 16, where the classification accuracy for RS is around
60%, and that for XRFS and LIBS >90%. Thus, the fusion results for XRFS
+ LIBS and for all analyzers exhibit a high accuracy that is >95%.
ERDC/CRREL CR-20-1 34
There is no single classification schema for the CSU soil suite, which can be
classified in multiple different ways. Therefore, it was decided to examine
other descriptive features ascertainable from the Soil Series descriptions
for each soil (i.e. soil thermal regime, soil moisture regime, geologi-
cal/mineralogical character, and geological source) in the context of the
five soil-forming factors (climate, parent material, landscape setting, bi-
ota, and time) to see if any of these influences could be teased out from
the statistical analysis. For purposes of the statistical analysis, analytical
data for the CSU sample suite was grouped for processing according to a 5-
class schema – NRCS Soil Order, thermal regime, moisture regime, miner-
alogical character, and geological source (Table 4). The feature sets used
were spectra from the individual XRFS, RS, and LIBS analyzers, the quan-
titative XRF analyses, the laboratory chemical analyses for the samples,
and a combination of all features. Results are shown in Figure 17 and the
classification matrices for the random forest classifier with text features are
shown in Figure 18. For the sensor fusion, samples were discarded that
were not measured using all the handheld analyzers and/or did not have a
full suite of analytical metadata. This resulted in the loss of a few samples
from the processed dataset.
ERDC/CRREL CR-20-1 35
Toward this end, the spectral data from the three analyzers, first individu-
ally, next fused together 2x2, then fused all together, and finally all infor-
mation about the samples (i.e. spectral data, XRFS quantitative analyses,
and laboratory chemical analyses) fused together. As illustrated in Figure
17, a moderate level of success was realized when classifying soils based on
soil thermal regime, soil moisture regime, and geological source. This re-
sult is not surprising given that only one sample was available per soil,
when what was really needed for robust statistical analysis was 5-10 sepa-
rate samples of each soil.
ERDC/CRREL CR-20-1 36
Figure 17. Classification accuracy for the CSU classification schema based on
consideration of spectra for Raman spectroscopy - RS (1), X-ray fluorescence
spectroscopy - XRFS (2), laser-induced breakdown spectroscopy - LIBS (3), fused
spectra for RS + XRFS (4), RS + LIBS (5), and XRFS + LIBS (6), RS + XRFS + LIBS
spectra (7), fused spectra + XRF quantitative analyses (8), and fused spectra +
XRF quantitative analyses + laboratory chemical analyses (9).
Figure 18. Independent feature accuracy per CSU soil classification schema.
Features are sorted on a per schema basis, greatest accuracy to least accuracy. The
legends indicate classification performance that could be achieved by always selecting
the class with the most samples.
space for the best selection of features can contain many local perfor-
mance maxima that can confound the search for the true global maxi-
mum.
Figure 19. Principal component analysis score plots for XRFS (upper left), LIBS
(lower left), RS (upper right), and chemistry (lower right) for the soil series of CSU
soil suite represented by multiple samples
three samples from the Bean Hill locality in Belmont, NH. Additionally,
three soil pits were excavated by SoilHydrology Associates at each of the
Bean Hill sample sites, with the different soil horizons in each pit sampled
and described.
Laboratory analysis was first undertaken for soil pellets prepared for the
Lebanon/CRREL and Belmont soil sets and then for the soil pits at the
Belmont location. The fusion feature vectors for the XRFS, RS, LIBS spec-
tral data were reduced to 2-dimensions using PCA. This analysis demon-
strated that the multi-analyzer approach can distinguish between different
soil types and identify soils that are similar (Figure 20). Laboratory analy-
sis of pellets from the three soil pits at Bean Hill in Belmont (Figure 21)
documents the clear distinction of the ash and charcoal layers and other
layers in soil pit and the compositional differences from the upper to lower
soil layers.
Classification for each individual analyzer and their spectral data fusion is
shown on the right side of Figure 22. Discrimination between the Leba-
non/CRREL and Belmont/Bean Hill soil sets is 100% correct for the
XRFS, RS, and LIBS analyzers individually. Therefore, nothing can be
gained through sensor fusion. By contrast, there is sufficient composi-
tional variability for the different layers of the three Bean Hill soil pits,
shown on the left side of Figure 22, that pit discrimination classification is
extremely poor for the RS spectra (< 30% correct) and not strong for ei-
ther LIBS (< 70% correct) or XRFS (<80%) spectra. Similarly, in this sec-
ond instance, data fusion does not provide any classification benefit.
Because the ultimate objective of this project was to develop a robust capa-
bility that could be used in the field, a short field study was conducted at
the Belmont, New Hampshire site to test the performance of the handheld
analyzers under field conditions with real time data transmission to a base
station. XRFS and LIBS analyses were undertaken for this purpose; a lo-
gistical issue prevented use of RS analyzer. Field measurements in the
three Bean Hill soil pits (Figure 23) demonstrated that the XRFS and LIBS
perform with confidence under ambient environmental conditions, which
included a rain shower during the fieldwork.
ERDC/CRREL CR-20-1 41
Figure 22. Classification accuracy for the New Hampshire soil sets
from Belmont/Bean Hill (left) and Lebanon/CRELL (right)
Figure 23. Principal component analysis scores plot for fused XRFS and LIBS
spectral data sets for in-situ analysis of the three Belmont/Bean Hill soil pits.
ERDC/CRREL CR-20-1 42
6 Summary
Funded under US Army ERDC-CRREL contract W913E518C0011, an 18-
month multidisciplinary effort was undertaken by JRPlumer & Associates,
LLC between September 2018 to March 2020. This project had three tech-
nical objectives: (i) to upgrade current handheld technology for chemical
analysis by X-ray fluorescence spectroscopy (XRFS), Raman spectroscopy
(RS), and laser-induced breakdown spectroscopy (LIBS); (ii) to design a
multisensor system based on these technologies for the rapid, in-situ
chemical analysis of soils and other materials of military interest; and (iii)
to investigate the classification/discrimination performance benefit that
might be achieved through advanced signal pre-processing and data fusion
with XRFS, RS, and LIBS analyses acquired for four suites of natural soils.
This initiative was supported by sub-contracts to SciAps, Inc., Signal Anal-
ysis Solutions, LLC, SoilHydrology Associates, LLC, and Applied Spectra,
Inc. Accomplishments of the program in the latter area are described in
this report.
A total of 272 soils from four sources were analyzed for the project - 50
rangeland soil samples from central New Mexico, 21 soil samples from two
sites in west-central New Hampshire, a suite of 144 agricultural soils from
across the United States and Canada, and 58 soils from small arms ranges
on military installations in Massachusetts, Virginia, Georgia, Idaho, Wash-
ington, and Alaska. An investigation of spectral data preprocessing was
undertaken that examined the benefit that could be realized from Ar emis-
sion line removal from LIBS spectra, spectra baseline correction, spectral
truncation and normalization, spectral culling through signal-to-noise ra-
tio analysis, similarity analysis, uncertainty mitigation and adaptive quan-
tization, and outlier rejection. A conceptual approach for real-time signal
processing was developed and demonstrated through a field test. Finally, a
hierarchical classification process was developed to combine spectral data
from multiple analyzers and this was applied to the soil data acquired dur-
ing the project. Using data from multiple collection dates it was observed
that fusion of XRFS-RS-LIBS spectral data provided better classifica-
tion/discrimination performance than for analysis by any individual ana-
lyzer.
ERDC/CRREL CR-20-1 43
References
Das, B.S., Hendrickx, J.M.H., and Borchers, B. (2001). “Modeling transient water
distributions around landmines in bare soils.” Soil Science 166: 163-173.
Davies, L. and Gather, U. (1993). “The identification of multiple outliers.” Journal of the
American Statistical Association 88: 782-792.
Egli, M., Hunt, A.G., Dahms, D., Raab, G., Derungs, C., Raimondi, S. and Yu, F. (2018).
“Prediction of soil formation as a function of age using the percolation theory
approach.” Frontiers in Environmental Science 6: 108.
Hempstalk, K., Frank, E., and Witten, I.H. (2008). “One-class classification by combining
density and class probability estimation.” In Proceedings of the Joint European
Conference on Machine Learning and Knowledge Discovery in Databases.
Antwerp, Belgium: Springer, 505-519.
Hendrickx, J.M.H., Alkov, N., Hong, S.-h., Van Dam, R.L., Kleissl, J., Shannon, H.,
Meason, J., Borchers, B., and Harmon, R.S. (2006). “New Mexico Tech landmine,
UXO, IED detection sensor test facility: measurements in real field soils” In
Proceedings of the International Society for Optical Engineering, SPIE, 6217:
278-289.
Kruse, F.A., Lefkoff, A.B., Boardman, J.W., Heidebrecht, K.B., Shapiro, A.T., Barloon,
P.J., Goetz, A.F.H. (1993). “The spectral image processing system (SIPS)—
interactive visualization and analysis of imaging spectrometer data.” Remote
Sensing of the Environment 44: 145–163.
Merritts, D.J., Chadwick, O.A., and Hendricks, D.M. (1991). “Rates and processes of soil
evolution on uplifted marine terraces, Northern California.” Geoderma 51: 241–
275.
Miller, T.W., Hendrickx, J.M.H., and Borchers, B. (2004). “Radar detection of buried
landmines in field soils.” Vadose Zone Journal 3: 1116-1127.
NRCS (1999). Soil Taxonomy - A Basic System of Soil Classification for Making and
Interpreting Soil Surveys. Washington, DC: US Department of Agriculture,
Natural Resources Conservation Service.
Soil Survey Staff (2014). Keys to Soil Taxonomy, 12th Edition. United States Department
of Agriculture, Natural Resources Conservation Service, 372 pp.
Van Dam, R.L., Hendrickx, J.M.H., Borchers, B., Hong, S.-h., Miller, T.W., and Harmon,
R.S. (2004). “A controlled outdoor test site for evaluation of soil effects on land
mine detection sensors.” In Proceedings of the International Society for Optical
Engineering, SPIE, 5415: 1335-1344.
ERDC/CRREL CR-20-1 44
White, A. G., Blum, A. E., Schulz, M. C., Bullen, T. D., Harden, J. W., and Peterson, M. L.
(1996). “Chemical weathering rates of a soil chronosequence on granitic
alluvium: I. quantification of mineralogical and surface area changes and
calculation of primary silicate reaction rates.” Geochimica Cosmochimica Acta,
60: 2533–2550.
ERDC/CRREL CR-20-1 45
Within the budgetary limits available for soil collection in New Mexico, it
was not possible to select a truly representative small set of soils from such
a large state of heterogeneous geological character. Therefore, instead of
using a statistical approach for the sampling design, criteria were devel-
oped to test the capability of the multi-analyzer approach to distinguish
different soil types and to identify similar soil types. These criteria, posed
as questions, are: (i) Can the multi-analyzer approach detect which soils
are different from one another?; (ii) Can the multi-analyzer approach de-
tect which soils are the same?; and (iii) Can the multi-analyzer approach
discriminate between soils that are somewhat different, but have common
underlying characteristics?
ERDC/CRREL CR-20-1 46
Initially, eight survey samples were collected from four soil series in cen-
tral New Mexico by SoilHydrology Associates in August of 2018 for the de-
velopment of a sampling protocol as well as initial testing and analysis by
JRPA. The sampling protocol developed consisted of taking a sample of
about 500 g from the top 0–5 cm of the soil (Fig. 6.1). After air-drying in
the laboratory at room temperature, the samples were sieved to remove all
particles and root debris exceeding a size of 2 mm. The sieved samples
were sent to JRPA for further processing and analysis.
This protocol was then followed to sample soils from six locations across
central-north New Mexico in November and December 2018 Six samples
were collected from each location along a straight transect approximately
22 m apart. Described in Table A1, six soils sampled for the project were:
(i) the loamy fine sand of the Bluepoint Soil Series at the Rio Salado, (ii)
the very stony sandy loam of the Laborcita-Pilabo-Lemitar Soil Series
Complex in Upper Socorro Canyon, (iii) the very stony sandy loam of the
Laborcita-Pilabo-Lemitar Soil Series complex in Lower Socorro Canyon,
(iv) the gravelly sandy loam of the Millett-Sedillo Soil Series complex at
Sedillo Hill, (v) the gypsiferous sand of the Lark-Transformer Soil Series
association at White Sands, and (vi) the strongly saline silty clay loam of
the playa soil at Willard. At Rio Salado, two sets of samples were gathered:
one at depth 0–3 cm and one at 10–13 cm. Thus, a total of 42 soil surface
samples formed the primary suite of samples analyzed for the project.
Figure A1. Geographic locations of the six New Mexico soil sampling sites.
ERDC/CRREL CR-20-1 47
Soil series are classified not only based on the characteristics of the surface
soil layer but on all soil horizons and characteristics down to a depth of 2
m depth (Soil Survey Staff, 2014) Often, there is no direct relationship be-
tween the properties of the 0–3 cm soil surface layer and its soil series
name. Nevertheless, in order to provide more understanding of the context
of the soil samples used in this project a short description with visuals is
presented below for each of the New Mexico soil locations.
NRCS description: The Bluepoint Soil Series consists of very deep, some-
what excessively drained soils that formed on dunes and sand sheets in eo-
lian materials derived from mixed rock sources. The typical NRCS profile
for the Bluepoint Series is:
ERDC/CRREL CR-20-1 48
NRCS description: The Nolam Soil Series (Figs. 6.6 to 6.8) consists of very
deep, well drained, moderately slow permeable soils that formed in allu-
vial sediments derived from rhyolite and andesite on terraces and pied-
monts. The typical NRCS profile is:
In addition to the differences between the NRCS typical profiles and the
local profiles observed in Socorro Canyon, there is one important addi-
tional local observation that pertains to the location of the boundary be-
tween the Arizo and Nolam series shown on the NRCS soil map (Fig. 6.6).
The orange boundary presented on the map is not correct, as the actual
boundary is located along the fault line (i.e. the black line on the map).
This is the nature of a 1:50,000 scale soil map and the issue of boundary
location on such maps. This problem of misplaced soil boundary locations
on such large-scale maps is not well understood by the portion of the soil
science community who lack geomorphological expertise.
ERDC/CRREL CR-20-1 51
NRCS description: The Ladron Soil Series (Figs. 6.10 and 6.11) consists of
deep, well-drained, moderately permeable soils that formed in alluvium.
These soils form on knolls and fan terraces. The typical soil profile is:
NRCS description: The Millett Soil Series consists of very deep, well-
drained soils formed in alluvium weathered from old gravelly alluvium
known as ‘rim gravels’. This material is composed of quartzite, sandstone
and some basic igneous rocks. Millett soils are on stream and fan terraces
on plateaus and footslopes and shoulders on stable landslides. The typical
soil profile is:
scale of the NRCS maps the actual soils sampled described at the Bean Hill
site may not necessarily fall within these NRCS soil series.
The soil at site BH-1 (Fig. 6.17) is mapped as the Pillsbury FSL (647B). This
soil is located on a small boulder terrace adjacent to a small stream. Soil
horizons at the site are as follows:
• Litter 10-0 cm,
• Ah horizon @ 0-10 cm,
• AB horizon @ 10-16 cm,
• Bw horizon @ 16-24 cm,
• C horizon @ 24-55 cm.
The soil at site BH-2 is mapped as the Canterbury FSL (167 B). This soil is
located in the upper backslope position on a heavily forested drumlin. Soil
horizons at the site are as follows:
• Litter @ 6-0 cm,
• Ah horizon @ 0-18 cm
• AB horizon @ 18-25 cm,
• Bw horizon @ 25-34 cm,
• C horizon @ 35-65 cm.
The Soil at site BH-3 (Figure 6.19) is mapped as Moosilauke FSL (415B).
This soil is located at the base of a small drumlin on a lower footslope posi-
tion. Soil horizons at the site are as follows:
• Litter @ 5-0 cm,
• Ah @ 0-10 cm,
• Bwj1 @ 10-25 cm,
• Bwj2 @ 25-36 cm,
• Ash and charcoal layer varying between 0-8 cm thick,
• 2Bwb @ 7 cm,
• 2Cb @ 12 cm.
The first two soils at the BH-1 and BH-2 sites appear similar in terms of
degree of development and, therefore, probably in age. They are forming
ERDC/CRREL CR-20-1 54
Acronyms
ASI Applied Spectra, Inc.
COTS Commercial-off-the-shelf
CRM Certified reference material
CRREL Cold Regions Research and Engineering Laboratory
CSU Colorado State University
ERDC Engineer Research and Development Center
eV Electron volt
FWHM Full width at half maximum
GUI Graphical user interface
Hz Hertz
ICP-MS Inductively coupled plasma mass spectrometry
JRPA JRPlumer & Associates, LLC
KeV Kilo electron volt
LA-ICP-MS Laser ablation inductively coupled plasma mass spec-
trometry
LIBS Laser-induced breakdown spectroscopy
LOSO Leave-one-sample-out
µ Micron
µm Micrometer
mA Milliamp
mJ millijoule
mm Millimeter
ns Nanosecond
PC Principal component
PCA Principal component analysis
QA/QC Quality assurance/Quality control
NIST National Institute of Science and Technology
nm Nanometer
NOTA None of the above
NRCS Natural Resources Conservation Service
R&D Research and development
RDTE Research, Development, Test and Evaluation
RS Raman spectroscopy
RSD Relative standard deviation
SAS Signal Analysis Solutions, LLC
SciAps SciAps Inc.
SHA SoilHydrology Associates, LLC
SMU Soil mapping unit
SNR signal-to-noise ratio
USACE U.S. Army Corps of Engineers
USDA U.S. Department of Agriculture
XRFS X-ray fluorescence spectroscopy
Form Approved
REPORT DOCUMENTATION PAGE OMB No. 0704-0188
Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the
data needed, and completing and reviewing this collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing
this burden to Department of Defense, Washington Headquarters Services, Directorate for Information Operations and Reports (0704-0188), 1215 Jefferson Davis Highway, Suite 1204, Arlington, VA 22202-
4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to any penalty for failing to comply with a collection of information if it does not display a currently
valid OMB control number. PLEASE DO NOT RETURN YOUR FORM TO THE ABOVE ADDRESS.
1. REPORT DATE 2. REPORT TYPE 3. DATES COVERED (From - To)
September 2020 Final
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Fusion of Spectral Data from Multiple Handheld Analyzers (LIBS, XRF and Raman) W913E518C0011
for Chemical Analysis and Classification of Soil 5b. GRANT NUMBER
14. ABSTRACT
An 18-month multidisciplinary project was undertaken by JRPlumer & Associates, LLC and four subcontractors that had three technical
objectives: (i) to upgrade current handheld technology for chemical analysis by X-ray fluorescence spectroscopy (XRFS), Raman spectros-
copy (RS), and laser-induced breakdown spectroscopy (LIBS); (ii) to design a multisensor system based on these technologies for the rapid,
in-situ chemical analysis of soils and other materials of military interest; and (iii) to investigate the classification/discrimination perfor-
mance benefit that might be achieved through advanced signal pre-processing and data fusion with XRFS, RS, and LIBS analyses acquired
for four suites of natural soils. Accomplishments of the program in the latter area are described in this report.
16. SECURITY CLASSIFICATION OF: 17. LIMITATION 18. NUMBER 19a. NAME OF RESPONSIBLE PERSON
OF ABSTRACT OF PAGES
a. REPORT b. ABSTRACT c. THIS PAGE 65 19b. TELEPHONE NUMBER
Unclassified Unclassified Unclassified UU