Purchasedby USDAfor OffjcialUse

PORTABLE, NEAR-INFRARED SPECTROPHOTOMETER

FOR RAPID SOIL ANALYSIS

K. A. Sudduth, J. W. HummelMEMBER MEMBERASAE ASAE

ABSTRACT.A portable, near-infrared spectrophotometer was designed, fabricated, and performance tested. Significantfeatures included a circular variable filter monochromator, a fiber optic coupling for sensing of remote samples, and asoftware algorithm which corrected the instrument readings to a zero baseline. The instrument had a sensing range of1650 nm to 2650 nm, an optical bandwidth of under 55 nm, and could acquire a spectrum every 200 ms. The intended useof this device was for rapid estimation of soil organic matter as a control input for variable rate herbicide application.Keywords. Soil organic matter, Herbicides, Spectrophotometry, Variable rate application, Optics, Sensors.

Site-specific crop production, or "prescriptionfanning", is a management technique in which theapplication rates of such inputs as herbicides andfertilizers are varied within a field in response to

spatial differences in soil and agronomic properties. Dataquantifying these spatial differences could be collected byintensive sampling and subsequent laboratory analysis, butfor highest efficiency it would be desirable to obtain thesedata by means of automated instruments capable ofconducting analyses in the field.

Spectral reflectance measurements provide one possiblemeans for estimating soil properties in the field.Researchers have correlated soil properties with bothvisible and near infrared (NIR) reflectance data (Dalal andHenry, 1986; Gaultney et at, 1989; Gunsaulis et at, 1991;Henderson et aI., 1989; Krishnan et at, 1980;and Schreier,1977). Sudduth and Hummel (1991) evaluated visible andNIR reflectance data for estimation of the organic mattercontent of TIlinoissoils. NIR data analyzed by partial leastsquares regression provided the best correlations (r2= 0.92,standard error of prediction 0.34% organic matter) for30 soils at wilting point and field capacity moisture levels.The reflectance data used were from 1720-2380 nm on a60-nm spacing and bandwidth, for a total of 12 reflectancepoints.

This article describes the design, development, andevaluation of a rugged, portable NIR spectrophotometer

Article was submitted for publication in November 1991; reviewedand approved for publication by the Information and ElectricalTechnologies Systems Div. of ASAE in August 1992.

This research was carried out under USDA Cooperative Research andDevelopment Memorandum of Understanding 58-32U4-M-592 withAGMED, Inc., Springfield, IL. Trade names are used solely for thepurpose of providing specific information. Mention of a trade name,proprietary product, or specific equipment does not constitute a guaranteeor warranty by the USDA-Agricultural Research Service and does notimply the approval of the named product to the exclusion of otherproducts that may be suitable.

The authors are Kenneth A. Sudduth, Agricultural Engineer,Cropping Systems and Water Quality Research Unit, USDA-AgriculturalResearch Service, Columbia, MO; and John W. Hummel, SupervisoryAgricultural Engineer, Crop Protection Research Unit, USDA-Agricultural Research Service, Urbana, u...

VOL. 36(1): JANUARy-FEBRUARY 1993

that implemented the soil organic matter estimation methodchosen by Sudduth and Hummel (1991). The intended useof this instrument was to provide control input data forvarying the rates of soil applied herbicides within a field inresponse to organic matter variations. The NIRspectrophotometer might also be used to quantify spatialvariations in other properties, such as soil moisture, that areimportant in site-specific crop production.

OBJECTIVES

The overall design objective for the prototypespectrophotometer was to implement the optimum soilorganic matter prediction method chosen by Sudduth andHummel (1991) on a near, real-time basis. Specific designsub-objectives were to:. Utilize a bandpass of 60 nm or less over a minimum

sensing range from 1700 nm to 2420 nm.. Provide an essentially continuous (in wavelength)sensing method, to allow flexibility for additionaloptimization of the wavelengths selected.. Have potential, with additional refinements ifneeded, to generate an organic matter signal every4.5 s to support on-the-go herbicide application rateadjustment at realistic field speeds.. Predict soil organic matter content with a standarderror of prediction of less than 0.5% (or less than0.29% organic carbon) for the 30 TIlinoissoils in thecalibration dataset (Sudduth and Hummel, 1991).

. Tolerate harsh environmental conditions (e.g., dust,temperature fluctuations, shock loads, and vibration)that would be encountered in field operation.

SELECTION OF DESIGN ALTERNATIVESCommercial NIR instruments were surveyed in an

attempt to identify a unit meeting the design objectives.The available grating monochromator and tilting filterinstruments measured reflectance at more wavelengthsthan required and at a narrower bandwidth, and this excesscapability resulted in high costs. Additionally, theseinstruments were not designed for the environmentalstresses imposed by field operation, and would have been

185

limited to laboratory use. Available fixed-filter instrumentswere more rugged and less expensive, but obtainedreflectance measurements at a slower rate and at fewerwavelengths than required. Since no commerciallyavailable NIR instrument met all the design objectives,development of a prototype spectrophotometer specificallytargeted at NIR reflectance sensing of soil organic mattercontent was necessary.

The wavelength selection mechanism and photodetectorwere chosen first, since these components would dictatethe configuration of the remainder of the system. Thealternatives for the photodetector were a single elementdetector or an array detector. A single element detectorwould be used with a wavelength selection mechanism (inthis case, a monochromator) which scanned thewavelengths of interest sequentially onto the detector. Anarray detector would be used if the wavelength selectionmechanism (in this case, properly termed a spectrographbut usually called a monochromator) focused thewavelengths of interest into a line image on a flat focalplane, thus providing simultaneous sensing at allwavelengths.

Three design alternatives were considered for thewavelength selection mechanism: a grating mono-chromator, a prism, and a circular variable filter (CVF).The grating monochromator was the usual device used inNIR laboratory spectrophotometers (McClure, 1987), butenvironmental considerations made its use more difficult ina field instrument. Worner (1989) constructed a visiblespectrophotometer using a prism and a linear array detectorin an attempt to overcome the environmental problemsseen with gratings. Circular variable interference filters,which have spectral characteristics that can be varied bymeans of physical rotation, have been used in rugged fieldinstruments for portable color measurement (Jauch, 1979)and for airborne infrared spectral measurements (Hoviset at, 1967).

Using the above detector and monochromatoralternatives, five possible options were identified:1) Grating and single detector; 2) grating and arraydetector; 3) prism and single detector; 4) prism and arraydetector; and 5) CVF and single detector.

Several of these choices were eliminated early in thedesign process. Option 1, a grating monochromator with asingle detector, would require oscillation of the grating toscan all wavelengths of interest onto the detector, makingthe device prone to vibration-induced inaccuracies.Options 3 and 4 used a prism monochromator, which wasnot available as a stock item in the wavelength rangerequired. The costs and lead time associated with customprism design and fabrication were not desirable within thescope of this project. Use of the prism with a singledetector would require a scanning mechanism and wouldentail the same type of vibration problems seen in thegrating system. The nonlinear dispersion characteristics ofa prism would make it difficult to provide data at equalwavelength spacings with an array detector.

Final consideration then was between Option 2, gratingmonochromator and array detector, and Option 5, circularvariable filter monochromator and single element detector.The circular variable filter monochromator and singleelement detector combination was selected after a detailedvendor survey, primarily due to the greater flexibility

186

offered by this approach. A CVF was available with awavelength range of 1600 nm to 2900 nm and a bandwidthof approximately 55 nm (Optical Coating Laboratory, Inc.,Santa Rosa, CA). This provided extra capability on bothends of the required sensing range (1700 om to 2420 om)while meeting the 60 nm bandwidth requirement.Additional flexibility was realized with the CVF sincereflectance readings could be taken at any desired points inthe wavelength range, subject only to the limitations of thedata acquisition system. In contrast, the best combinationof grating monochromator and linear array detector did notcompletely cover the required range and allowed sensingonly from 1720 nm to 23-80nm. This combination had atheoretical bandwidth of 36 nm, but could output only16 reflectance readings, corresponding to the 16 elementsof the linear array detector. Other factors favoring thechoice of the CVF system were its greater tolerance of dustand vibration, and a reduced degree of complexity in theinterface electronics due to the use of one detector and datachannel versus 16 detector elements and channels. Onepossible drawback of the CVF system was its sequentialrather than simultaneous acquisition of data. It wastheorized that maximum accuracies with this sequentialwavelength scanning approach might require holding thesample stationary while data were being acquired so thatall wavelengths were scanned on an identical area of thesample.

DETAILED DESIGN DESCRIPTIONThe basic components of the NIR spectrophotometer

(fig. 1) were a broadband NIR source, CVF mono-chromator, fiber optic bundle before the sample, and a leadsulfide (PbS) photodetector to measure the energy reflectedfrom the sample. Output of the detector was conditioned bya pre-amplifier and input to a personal computer throughan analog-to-digital (AID) converter.

An aluminum housing formed the main structure of theinstrument and enclosed an aluminum filter disk containingthe quarter-segment CVF (fig. 2). Modulation of the lampoutput radiation for low frequency noise and drift rejectionwas accomplished with the filter disk itself, rather than bythe standard practice of a separate chopper disk. Byspinning the filter disk at a sufficient rate and using thethree-quarters of the disk which blocked the light path toperform the modulation function, the need for a separatemotor, chopping disk, and sensing electronics was avoided.A servo-controlled de motor-generator (Motomatic E-350,Robbins and Meyers, Hopkins, MN) allowed adjustment ofthe filter disk rotation speed. Operating speed in thisapplication was limited to 10Hz due to balancingconsiderations in the filter disk assembly, photodetectorresponse, and AID sampling rate.

An opto-interrupter was mounted within the instrumenthousing such that its optical path was broken once per filterdisk revolution by a small tab attached to thecircumference of the disk (fig. 2). The timing pulsegenerated by the opto-interrupter was conditioned to TTLlevels by a Schmitt trigger circuit. This TTL signal wasthen used to compensate for any variations in the filter diskspeed and to provide a positive angular position referencefor wavelength determination. The timing tab position usedinitially was opposite the CVF segment. Later tests were

TRANSACTIONS OF THE ASAE

AllGUW POSITIONSENSOII

\ :, ,\ :, ,

~ / CHOPPEIt

IIIAOE OFIlECTAIIGUWFILAIIENT

IlO1OIIaCONnIOWJt

ClIlCUW ENDOF FlIER BUNDLE

-1U.UIiIIIIA11ONCONE

QUARTZ WINDOW

[L J I-SAIIPLE CUP

Figure l-Optical system, including instrument housing with f"llterdisk and lamp (top), and sensor head assembly (bottom).

done with the tab adjacent to the CVF segment on the filterdisk (fig. 2) so that the timing pulse would coincide moreclosely with the analog reflectance signal.

A 50 W, 12 V, quartz halogen automotive-type lamp(fig. I) driven by a laboratory power supply was used asthe illumination source for the spectrophotometer. Thelamp mounting allowed three-axis adjustment for focusingand positioning the lamp image. A spherical bi-convex lens(fig. 1) was mounted in the upper surface of the instrumenthousing to focus the lamp image through the input slit andonto the surface of the CVF.

The wavelength of the light which passed through theCVF at any point in its rotation was a linear function of theangular position of that point relative to the leading edge ofthe filter. Therefore, to obtain monochromatic (or nearlyso) light from the system, a narrow radial slit (2-mm widex 10-mm radial length) was mounted about 5 mm abovethe surface of the filter. The projected image of the lampfilament was of a similar size and shape, so that only asmall portion of the lamp energy was blocked at the slit.

The monochromatic light from the CVF was directed tothe sample through a 61O-mmlong silica fiber optic bundlewith a useful transmission range from 350 nm to over2400 om (Volpi Fiber Optics, Auburn, NY). This sectionconverter bundle changed shape from a I mm x 10 mmrectangular section at one end to a 3.6 mm circularcross-section at the other. The rectangular section end ofthe fiber bundle was mounted approximately 5 mm belowthe surface of the CVF and in line with the input slit, thus

VOL. 36(1): JANUARY-FEBRUARY 1993

INmAL TIlliNGTAB POSITION

LOCATION or INPUTSLIT I!lAGE

FlNAL TIlliNGTAB POSITION

Figure 2-Circular variable filter disk, including initial and revisedtiming tab locations.

collecting the majority of the filter throughput. The circularcross-section end of the fiber optic bundle was attached tothe sensor head assembly with an adjustable mounting, toallow optimization of the location of the fiber exit conewith respect to the detector and the sample surface.

The sensor head assembly (fig. I) consisted of analuminum housing, an optically-flat quartz aperturewindow for dust exclusion, and an attached photodetector.The PbS photodetector was selected as preferable to theother available types (notably lead selenide and indiumarsenide) due to its lower cost, higher responsivity, andability to operate without cooling. Monochromatic lightfrom the fiber optic bundle passed through the quartzaperture window and illuminated a circular area on thesample surface. A portion of the energy diffusely reflectedfrom the sample passed back through the quartz apertureand was collected by the photodetector (OTC-22-53,OptoElectronics, Petaluma, CA). This detector had a usefulsensing range from approximately 1000 nm to 3500 nmand a 3 mm x 3 mm square sensing area. The detector wasoperated uncooled at ambient temperature to ensure anacceptable speed of response. The excitation andpreamplifier circuitry (fig. 3) for the PbS detector consistedof a high-gain single stage amplifier which was ac coupledto the output of the detector for insensitivity to lowfrequency drift in the detector output.

+liIY 58pfIlIea

33.2k5111"Iw

8

R.lloclance= (Vou! )

GAIN

10k r'>20 turn +

33.2k-lilY

~~ufT Tanl.

100

Figure 3-Detector excitation and preamplifier circuit.

187

DATA ACQUISITION SYSTEMData was acquired with a MetraByte DAS-16 analog

and digital input-output expansion board (Keithley,Taunton, MA) installed in an AT-compatible computerrunning at 12 MHz. One 12-bit differential analog channelon the DAS-16, configured with a:t 5 V range, was used tocollect data from the photodetector preamplifier. The TTLoutput from the filter disk timing tab was input as a digitalsignal to allow measurement of the filter disk period, gatethe AID converter, and synchronize data collection.

The required AID sampling rate was approximately10 kHz, based on the geometry of the CVF, a 10 Hzmaximum speed of the filter disk, and a desire to obtainreflectance data on a 5 nm maximum spacing. A programwas written in compiled BASIC to transfer the AID datadirectly to an array while simultaneously allowing thecounter operation needed to time the rotation of the filterdisk. The program allowed data from up to 100 revolutionsof the filter disk to be acquired, displayed for rapid visualverification, and stored on disk for later analysis.

The data acquisition program was tested to determinethe maximum reliable sampling rate above which sampletiming errors would become unacceptable. Two functiongenerators were used to simulate the outputs of thespectrophotometer. The TTL output of one generatorsimulated the gating signal from the timing tab on the filterdisk, and also triggered the second function generatorwhich simulated the analog output of the photodetectorpreamplifier by means of a sine wave. Over the range from2 kHz to 12.5 kHz the reliability of the data acquisitionprocess was very good, with an error rate of less than10 out of 5,000 scans (0.2%) in all cases, and nodiscernable trend of error rate versus sampling rate (table1). A 15.2 kHz sampling rate was too fast for the system,causing an error rate of over 20%. Based on these tests, thedata acquisition program was configured with a 10kHzsampling rate, allowing a margin of safety before the pointwhere the frequency of errors became unacceptable.

OPTICAL PERFORMANCE TESTS AND

CALffiRATION

Once the prototype NIR spectrophotometer wasoperational, a sequence of performance tests andcalibrations was completed. The steps in this procedurewere: (1) optimize values for such basic operatingparameters as filter disk speed and sample distance;

Table 1. Number of bad data scansout of 5000 total scans at various

AID sampling rates

Sampling Rate Number of(kHz) Bad Scan

~O 85.0 65.5 86.7 108.0 3

10.0 512.5 7

15.2 > 1000

188

(2) document and analyze the general response of theinstrument with these settings; (3) calibrate instrumentoutput in terms of percent reflectance and wavelength; and(4) test the performance of the calibrated system in termsof such characteristics as bandwidth and stability. Althoughthe details of these steps were specific to the instrumentbeing developed, the general structure of this sequencecould be applied to any reflectance instrument.

DETECTOR RESPONSE

The manufacturer's specifications for the PbSphotodetector reported a typical time constant (the timerequired for the response to a step input to reach 63% of itsfinal value) of 250 JiSand a maximum time constant of400 J1Sfor an uncooled detector. At the design filter diskspeed of 10Hz, this time constant would correspond to13 nm to 21 nm and would cause some filtering of thereflectance signal. To minimize the degree of filtering andmaximize the signal, the detector output should reach itsfinal value (or very nearly so) during the time that thetheoretical pulse of light from the 2 mm entrance slit wasincident on anyone point on the CVF.

To test the detector response, a window was fabricatedin the filter disk directly opposite the CVF segment (fig. 2).During normal operation, this window was completelyshuttered, but for detector response testing a special shutterwith two openings was installed. The first opening to passby the light source during operation was a narrow slit1 mm wide x 10 mm long. The second opening was aradial segment of a ring, approximately 10 mm x 10 mm.The first opening was designed to allow measurement ofthe pulse response of the detector by approximating thepulse width caused by the interaction of the entrance slitand a point of interest on the CVF, while the second, wideropening was designed to allow the settled response of thedetector to be quantified. Detector output from a standardceramic reflecting surface was recorded at filter diskperiods from lOOms to 300 ms. The 1 mm pulse responsewas normalized by the settled response from the 10 mmopening (fig. 4). As a compromise between the increasingsignal level with longer disk period and the desire to collectdata as quickly as possible for field operation, a filter diskperiod of 200 ms (speed of 5 Hz) was selected forsubsequent tests.

---------------- .""'"8. 0.9"'"cr:-'""'"Il.U.~~ 0.80z

Nominal OperatingPoint (200 ms)

0.7100 150 200

FilterDiskPeriod, ms250 300

Figure 4-Relationship between photodetector response and filter diskperiod.

TRANSACTIONS OF THE ASAE

SAMPLE DISTANCE EFFECTS

To determine the optimum operating distance from thesample surface, the location of the sensing head was variedfrom 6.4 mm to 25.4 mm above a ceramic standard and theoutput signal recorded (fig. 5). Although operation at theminimum distance would provide the strongest signalpossible, it would have two negative effects. In a field unitsome allowance would be needed to compensate forvariations in instrument-to-sample distance. Decreasingthis distance would also decrease the area of the samplebeing sensed. For a nonuniform material such as soil,sensing of a relatively larger area would be desirable toaverage out any signal differences caused by sampleheterogeneity. As a compromise between signal strengthand these two effects, a nominal operating distance of15 mm was selected.

SYSTEM OUTPUT CHARACTERISTICS

The general output characteristics of the system weredocumented at a filter disk speed of 5 Hz. A ceramic diskwas used as the reflectance object due to its nearly constantreflectance over the wavelength range of interest. Thedigitized signal obtained for a typical ceramic diskreflectance scan (fig. 6) exhibited several noteworthyfeatures. The rapid signal rise near AID sample point 500occurred as the leading edge of the CVF passed theentrance slit. The roll-off of the response with increasingAID point number (and increasing wavelength) was due tothe transmission characteristics of the fiber optic bundleand possibly the illumination characteristics of the lamp.Several nonlinearities were present in the signal, mostnotably a notch near point 700 likely due to the presence ofsome unidentified absorber in the optical system.

Another characteristic of the response curve was thenon-zero baseline found in the non-illuminated portions ofthe scan (outside the range from about point 500 topoint 875). This non-zero, dynamic baseline was caused bythe ac coupling of the detector output in the preamplifier(fig. 3). The ac coupling eliminated possible long-term driftproblems and allowed optimal use of the available AIDrange. However, this ac coupling required application of asoftware correction to adjust the digitized reflectance datato a zero baseline.

The output signal of the instrument with no illuminationprovided an indication of the amount of excitation-

1.2

Q)

i!! 0.80a.8!a:"C 0.6.~'"E(; 0.4z

Nominal Operating

Point (15 mm)

0.2

00 4 8 12 16 20

Sensor to Sample Distance, mm

24 28

Figure 5-Relationship between signal level and sample distance.

VOL. 36(1): JANUARY-FEBRUARY 1993

1600

1200

~ 800:J0

()0

:«400

a

400a 400 800

NO Point Number

1200

Figure 6-Digitized signal for a typical ceramic disk reflectance scan.

independent noise present in the data. The noise was quitesmall in magnitude, approximately 2 AID counts or 0.1%of the full-scale reflectance signal, and consisted of a high-frequency random noise superimposed on a 60 Hzsine-wave. With no radiation incident on the detector, apositive offset of approximately 5 AID counts wasobserved.

CERAMIC DISK REFERENCE CALIBRATION

To compensate for changes in illumination, detectorresponse, and other optical system variations, each sensorreading was referenced to the reading from a ceramic disk,a substance widely accepted for standardization of NIRinstruments. The ceramic reference also enabledconversion of the instrument response to a percentreflectance (or decimal reflectance) basis.

Two nominally identical ceramic disks were used in thecalibration procedure. One 50-mm diameter disk was usedas the working reflectance standard for the prototypespectrophotometer, while the other disk was sent to theUSDA-ARS Instrumentation and Sensing Laboratory,where its reflectance characteristics were obtained bycomparison with a standard sample of slightly compressedsulfur (fig. 7). A series of 10 paired readings of bothceramic disks was then completed with the prototypespectrophotometer and used to compute the decimalreflectance characteristics of the ceramic disk used forsensor calibration (fig. 7).

WAVELENGTH CALIBRATION

During early laboratory tests, a significant amount ofsub-periodic filter disk speed variability was observed.Relocation of the filter disk timing tab (fig. 2) reduced thisvariability by providing timing references as close aspossible to the beginning and to the end of the analogreflectance signal. The relocated timing tab provided thereference near the end of the reflectance signal, while thereference at the beginning of the signal was obtained fromthe half-peak point on the initial step response of the CVF(fig. 6). Each individual AID point was then located atsome fraction of the elapsed time between the initialreference and the final reference. This period, duringwhich filter disk speed variations would reduce theaccuracy of the wavelength calibration, was minimized at

189

0.85

- Ceramicdisk calibratedat ISL- - - - Ceramicdisk used in laboratorytests

0.84aa

g 0.83'"UQ)

16a:: 0.82EQ)"CD

IL

I/, I

-" , ../-'--' \ ''- \ ,,_/0.81

0.81600 1800 2000 2200

Wavelength, nm

Figure 7-Reflectance characteristics of the ceramic reference diskscanned at the USDA-ARS Instrumentation and Sensing Laboratory(ISL) and the working ceramic reference disk used to calibrate theprototype spectrophotometer.

2400 2600

approximately 40% of the total filter disk period. Themaximum uncertainty in wavelength estimation wasapproximately 2.5 om.

Wavelength calibration was accomplished with a sheetof white polystyrene, a material with well-definedabsorption bands in the NIR region (Williams, 1987). Themajor polystyrene absorption bands centered at 1682 nm,2165 nm, and 2470 nm, identifiablein prototypesensorscans of the polystyrene sheet (fig. 8), were used in thecalibration step. Each polystyrene dataset was ratioed witha paired ceramic reading to obtain reflectance data whichwas then processed to find the location of the absorbancepeaks (or reflectance troughs). A linear regression wasused to relate the location, in terms of fraction of the timingperiod, of the three polystyrene reflectance minima to theirknown wavelengths. An excellent correlation was found,with an r2 of 0.99999 and a standard error of thewavelength estimate of 1.18 nm. Application of thewavelength calibration equation showed that the sensor hada usable data range from under 1650 nm to over 2650 om,well in excess of the design requirement.

0.5

0.4

aa

Q)g 0.3

~Q)'"Q)a::E 0.2Q)"£

0.1

00 0.1 0.2 0.3

Fraction of Gate

0.4

Figure 8- Typical polystyrene reflectance curve obtained with theprototype spectrophotometer.

190

BANDWIDTH DETERMINATION

The bandwidth of the CVF was investigated by insertinga fixed filter with a 2230 nm center wavelength andnominal 100m bandwidth in the optical path between thelight source and the CVF. The halfpeak bandwidth of thereflectance curve of a ceramic standard was obtained atseveral filter disk periods with the fixed filter in place.

The apparent bandwidth of the CVF with the standard200 ms period was 52 nm (fig. 9), within the system designrequirements. The bandwidth with a 300 ms period was50 nm, while with a 100 ms period it was 61 nm. Thestandard 2 mm entrance slit was not a limiting factor, as theCVF bandwidth remained at 52 nm when the slit width wasdecreased to 1 mm. The apparent bandwidth determined bythis test was somewhat higher than the true bandwidth ofthe CVF, since the light illuminating the CVF was notmonochromatic, but rather exhibited the 10 nm bandwidthof the fixed filter. A computer simulation using triangularfilter functions predicted that the apparent bandwidthwould be approximately 10% greater than the actualbandwidth. As this level of accuracy was sufficient,deconvolution to obtain the true CVF bandwidth was notattempted.

OPTICAL RESPONSE STABILITY

Variation in the optical response of the sensor wasmonitored by means of 10 repeated readings of the sameceramic disk over a 2-h period. Little variation was seenwhen comparing the baseline corrected mean sensorreadings. To magnify any differences present, each readingwas normalized by dividing the response pointwise by theceramic response measured at the beginning of the 2-hperiod. Two types of variation (fig. 10) were seen in thenormalized response curves as compared to the unityresponse which would have been obtained with a perfectlystable system. An approximately constant offset wasobserved in the 1650 to 2100 nm portion of the curve,indicating a drift in the lamp output or other opticalcharacteristics of the system.

More variation was seen in the portion of the curve past2100 nm, most notably a peak or trough located near2200 nm, a location corresponding to the notch in the

100

80

60.$c::J

8 40()

~20

0

-202100

0.52200 2300

Wavelength, nm

Figure 9-Response curve with a 2230 nm center wavelength, 10 nmbandwidth filter placed in the optical path, showing the 52 nmapparent half-amplitude bandwidth ofthe system.

2400

TRANSACTIONS OF THE ASAE

1.05

1.03

0

~ 1.01c:as0C/)0

.~ 0.99"u

0.97

0.951600 1800 2000 2200

Wavelength,nm

2400 2600

Figure to-Teu readings of the same ceramic disk during a 2 hstabllity test, normalized by ratioing each successive curve to the first(time zero) reflectance curve.

system response curve (fig. 6). These variations betweenscans were due to sub-periodic variations in the filter diskspeed, which caused the calibration between wavelengthand AID sample point to shift. The influence of a givenwavelength shift on sensor output was magnified inportions of the response curves with steeper slopes andwith rapidly changing slopes such as the notch area notedabove.

To verify this relationship, a ceramic scan wasreferenced to itself, with wavelength shifts of up to 5 om ineach direction. The results of this simulation (fig. 11) werevery similar in shape to the deviations observed betweensuccessive ceramic readings (fig. 10), providing additionalevidence that the majority of the non-offset differencebetween ceramic readings was attributable to sub-periodicvariations in filter disk speed.

DATA PROCESSING ALGORITHMSA sequence of operations was necessary to convert the

raw digitized data obtained from the spectrophotometerand stored by the data collection program to the percent

1.05

1.03

iiiiiiiii!

0~II: 1.01c:as0C/)0

.~ 0.99Wu

""="""-'~/ ~'~'-~ k..;:I"\. 'J' -, ~ , V \~,\. \ /. \ ~~-

\, \ ,f ". \...\, , \ "-i ; '-,.A. '''Yi ; \ \i . ", \J i I

\ I\'offl" I

Ii Ii I

--Ll..J2600

0.97- +5.0nm_n_n +2.5nm--- -2.5nm

- 5.0nm

0.951600 2000 2200

Wavelength, nm

Figure ll-The effect of wavelength offset on the ratio of a ceramicreflectance curve to itself.

1800 2400

VOL. 36(1): JANUARy-FEBRUARY 1993

reflectance data needed to calibrate the instrument and toestimate soil properties.

BASELINE CORRECTION

A dynamic baseline correction algorithm was developedto convert the ac coupled raw reflectance signal (fig. 6) to adc signal with a baseline level of zero AID counts.Referring to the schematic of the detector and preamplifier(fig. 3), the parameter which varied directly with theradiation incident on the detector was the current throughthe photodetector, iin' The parameter digitized by the datacollection system was the voltage output of the preampli-fier, Vout.Analysis of the preamplifier circuit showed thatthe general relationship between these two parameters wasgiven by:

iin= kl + k2 Vout + k3 f Voutdt(1)

The output voltage, Vout' was converted to raw AIDcounts, Draw' by application of a gain and offset in thedigitization process:

Draw= ao + al Vout (2)

Vout= a3 + a4Draw (3)

Draw was the parameter stored as data by the computer.However, the desired reading was the corrected AIDcounts, DeOIT'which represented the digitization of theinput current. A sequence of steps was required to calculateDeOITfrom Draw' Equation 1 was digitized and rewrittenwith Drawreplacing Voutand DeOITreplacing iin:

n

DcOIT = CI + C2 Draw + C31: Draw + C41

(4)

Equation 4 was simplified by setting c2 to one, since theunits of Drawwere the same as the units of DeOIT'Also, c4was removed from under the summation:

n

DcolT = Draw + CI + C4 n + C31: Draw1

(5)

Calculation of DeOITfrom Drawthen required values forc}, c3' and c4 to be determined. These values varied witheach particular data curve, due to the differences inreflected energy between samples. To determine c}, c3' andc4 it was necessary to make use of the fact that DeOITshould, on the average, be zero in the baseline portions ofthe curve (from points 1 to 500 and 875 to 1400 in fig. 6).In these two portions of the data curve, equation 5 could berewritten as:

}9}

n

Draw=-c]-c4n-c3L Draw]

Equation 6 was fit to the data in the baseline areas of thecurve using a least-squares multiple linear regression,where the independent variables at each data point were theindex number of the data point and the summation of theraw AID data up to that point. The dependent variable wasthe raw AID data value. The regression was applied to al67-point section of the baseline on either side of the dataportion of the curve, corresponding to 16.67 ms at the10kHz sampling rate, to account for the mean of any60 Hz noise present in the data. An excellent fit to the rawbaseline data was obtained, with typical r2 valuesexceeding 0.999. Once c1' c3' and c4 were determined,equation 5 was then applied pointwise to the raw AID datato generate the baseline corrected data curve (fig. 12).

WAVELENGTH CALIBRATION

The polystyrene wavelength calibration equation wasapplied to the data to convert AID point number to thecorresponding wavelength.

DATA POINT INTERPOLATION

Due to wavelength calibration differences and variationsin filter disk period, a given AID point did not correspondto the same wavelength for all reflectance readings.Therefore, the baseline corrected sensor response data wereinterpolated to a standard wavelength spacing for pointwisedifferencing of soil and ceramic readings and additionalanalysis. Points were generated every 5 nm from 1600 nmto 2700 nm using piecewise cubic spline interpolationalgorithms (Spath, 1974).

REFLECTANCE CALCULATION

The interpolated, baseline corrected raw data obtainedfrom the sensor were converted to decimal reflectance(percent reflectance + 100) by comparison against dataobtained from the ceramic disk with known reflectancecharacteristics (fig. 7). Repeated readings of the ceramicdisk were done on a frequent basis, so that compensation

1600

1200

"'§ 800a00~

400

0

-400400 600 800

AID Point Number

Figure 12-Typical raw ceramic reflectance curve and the same curveafter application of the baseline correction algorithm.

192

(6)could be made for lamp output fluctuations or otherchanges in the optical path of the sensor.

EXAMPLE OUTPUTThe prototype NIR spectrophotometer worked well in

laboratory use. Ten-scan average soil reflectance dataobtained from this instrument were compared to dataobtained with a high-precision research grade spectro-photometer at the USDA-ARS Instrumentation andSensing Laboratory (fig. 13). Data from the two sourcesagreed quite well, both in overall shape and level, for awide range of soils. The differences observed wereattributable to slightly different soil moistures, especially atthe 1.5 MPa nominal moisture tension level. The soilorganic matter estimates obtained by application ofmultivariate calibration techniques to these data wereinsensitive to the reflectance differences induced bychanges in soil moisture content (Sudduth, 1989).

SUMMARY AND CONCLUSIONSA portable, near-infrared spectrophotometer intended

for in-field use was designed, fabricated, and performancetested in the laboratory. The intended use of this devicewas rapid estimation of soil organic matter as a controlinput for variable rate herbicide application, usingpreviously developed methods (Sudduth and Hummel,1991).

Significant features included a circular variable filtermonochromator for ruggedness and simplicity of operation,fiber optic coupling for sensing of remote samples, and asoftware algorithm which corrected the ac-coupledinstrument readings to a zero baseline. Performance testsdocumented a sensing range of 1650 to 2650 nm, an opticalbandwidth of under 55 nm, and an optimal data acquisitionrate of 5 Hz. Soil reflectance curves obtained with thisinstrument in the laboratory agreed well with data obtainedusing a research grade spectrophotometer.

0.6

0.5- 1.5MPaTension- - -. 0.033 MPaTension

00

::::. 0.4CD0c:'"~ 0.310a:1::~ 0.2(;;a.

0.1

0 -400 800 1200 1600 2000 2400

Wavelength, nm1000

Figure 13-Comparison between mean reflectance curves from theUSDA-ARS Instrumentation and Sensing Laboratory (ISL) and three

replicate prototype spectrophotometer curves for Ade Loamy Sand attwo moisture tension levels.

TRANSACTIONS OF THE ASAE

ACKNOWLEDGMENT.The authors acknowledge thecontributions of Robert C. Funk and others at AGMED,Inc., to the design and fabrication of the spectro-photometer. This invention is covered under U.S. Patent5,038,040, which is assigned to AGMED, Inc.

REFERENCESDalal, R C. and R J. Henry. 1986.Simultaneous determinationof

moisture, organic carbon, and total nitrogen by near infraredreflectance spectrophotometry.Soil Sci. Soc. Am. J. 50:120-123.

Gaultney, L. D., J. L. Shonk, D. G. Schultze and G. E. VanScoyoc. 1989.A soil organic matter sensor for prescriptionchemical application. In Agricultural Engineering -Proceedings of the 11th International Congress,Dublin,4-8 September 1989,eds. V. A. Dodd and P. M. Grace, 3065-3072. Rotterdam: Balkema.

Gunsaulis, F. R, M. F. Kocher and C. L. Griffis. 1991. Surfacestructure effects on close-range reflectance as a function of soilorganic matter content. Transactionsof the ASAE 34(2):641-649.

Henderson, T. L., A. Szilagyi, M. F. Baumgardner, C. T. Chen andD. A. Landgrebe. 1989.Spectral band selection forclassification of soil organic matter content. Soil Sci. Soc.Am.J.53:1778-1784.

Hovis, W. A., Jr., W. A. Kley and M. G. Strange. 1967.Filterwedge spectrometer for field use.Appl. Optics 6(6):1057-1058.

Jauch, K. M. 1979. A portable rapid spectrophotometer for colourstudies.J. Physics E: Sci. Instrumentation 12:1171-1175.

VOL. 36(1): JANUARy-FEBRUARY 1993

Krishnan, P., J. D. Alexander, B. J. Butler and J. W. Hummel.1980. Reflectance technique for predicting soil organic matter.Soil Sci. Soc. Am. J. 44: 1282-1285.

McClure, W. F. 1987. Near-infrared instrumentation. In Near-

infrared technology in the agricultural andfood industries,eds. P. C. Williams and K. H. Norris, 89-106. St. Paul, MN:American Association of Cereal Chemists.

Schreier, H. 1977. Quantitative predictions of chemical soilconditions from multispectral airborne, ground, and laboratorymeasurements. In Proc. 4th Canadian Symposium on RemoteSensing, 106-112. Ottawa: Canadian Aeronautical and SpaceInstitute.

Spath, H. 1974. Spline Algorithmsfor Curves and Surfaces.Winnipeg: Utilitias Mathematica Publishing.

Sudduth, K. A. 1989. Near infrared reflectance soil organic mattersensor. PhD. diss., University of Illinois, Urbana-Champaign.Public. No. 90-15736. Ann Arbor MI: University Microfilms,Inc. (Diss. Abstr. Int. B 51(1):308).

Sudduth, K. A. and J. W. Hummel. 1991. Evaluation of reflectancemethods for soil organic matter sensing. Transactions of theASAE 34(4): 1900-1909.

Williams, P. C. 1987. Commercial near-infrared reflectanceanalyzers. In Near-infrared technology in the agricultural andfood industries, eds. P. C. Williams and K. H. Norris, 107-142.St. Paul, MN: American Association of Cereal Chemists.

Womer, C. R 1989. Design and construction of a portablespectrophotometer for real time analysis of soil reflectanceproperties. Unpublished M.S. thesis, Library, University ofIllinois, Urbana-Champaign.

193

" ,<'