The Zeltex ZX-800 is an NIR-emitting diode based near-infrared transmittance analyzer costing about 50% of other such analyzers used for grain analysis. Three ZX-800 master units were used to create universal calibrations for corn moisture, protein, oil, and starch, and for soybean moisture, protein, and oil. Samples originated from all major US growing regions, from 1996 and 1997 crops. Zeltex, using other units not included in the calibration process, tested transferability of these calibrations. Transferability ha been a historic concern for this particular technology; an optical standardization strategy was used to remove many unit differences as possible, and to include the rest in the calibration process. PLS regression with cross validation was used for calibration development.

The standard errors of prediction and standard deviation across slave units for the calibrations were:

Table 1. Calibration Statistics for Zeltex ZX-800

Constituent

Statistic

Value

Reported by USDA for Infratec a

Corn (Basis 15% moisture)

Moisture (%)	n c Range SEP d	977 b/141 b 8-25/23-25 0.60/0.57	N/A N/A N/A
Protein (%)	n Range SEP	338 5-12 0.63	7-10 0.44
Oil (%)	n Range SEP	247 4-9 0.44	4-7 0.40
Starch	n Range SEP	349 52-64 1.19	54-61 2.20

Soybeans (Basis 13% moisture)

Moisture (%)	n Range SEP	505 7-20 0.46	N/A N/A N/A
Protein (%)	n Range SEP	469 27-46 0.70	32-38 0.60
Oil (%)	n Range SEP	460 13-23 0.39	15-21 0.50

a	Infratec 1225 analyzers used in Official USDA grain inspections, as reported in Federal Register 6/30/98 (corn) and 6/13/98 (soybeans).
b	Low and high moisture calibrations.
c	Number of samples included.
d	Standard Error of Prediction, relative to chemistry.

In most factors, these compared favorably with reported results for other units. A procedure was developed to utilize multiple NIR brands for on-site analysis even when contractual requirements may specify one brand and/or calibration for contract settlement.

Near-infrared (NIRS) analyzers are rapidly becoming accepted as the instrument of choice for quality control of value-added grains such as high oil corn, low saturated fat soybeans, and other specialized products. Hurburgh et al. (1990) reported that NIRS units, if testing unground seed as received by a grain elevator, can perform analyses accurately in one minute per sample, a rate compatible with handling operations at most elevators. The cost of testing and segregation for value-added grains is estimated to be 2-4 cents per bushel (Hurburgh et al., 1993), which is reasonable compared to the 10-15 cents per bushel value increases represented by the current value-added grain products.

There are many network groupings of NIRS units worldwide. Hardy et al. (1995) and Rippke et al. (1995) reported the performance of one such network in the USA, unusual in that it contained units of more than one brand. Table 1 summarizes calibration and transfer statistics for this network, for corn and soybeans. In this table, data from all brands are pooled. These data exceed the performance proposed by the USDA Official inspection system as given in the Federal Register, July 1, 1998, for Infratec analyzers - the single brand used in the Official USDA inspection system.

Cost is always a concern for grain handlers operating on thin margins. Lower cost instruments are typically associated with reduced performance. There has been no study that attempted to quantify the accuracy of lower cost units and determine if such units could be used effectively in commerce.

The objectives of this study were:

1. To evaluate the Zeltex ZX800, an NIR-emitting diode-based unit, for measuring corn moisture, protein, oil, and starch, and for measuring soybean moisture, protein, and oil.
2. To develop a statistically sound operating protocol for using this unit in a marketing situation where another brand was being used as the standard.

Three ZX800 units were used in the Iowa State University Grain Quality Laboratory during the 1996 and 1997 crop seasons. Fresh harvest corn and soybean samples, with reference chemistry data, were scanned along with stored samples from previous crop seasons. The chemistry data was done by Woodson-Tenet, Des Moines, Iowa. This was the same sample set used to calibrate other brands as reported by Hardy et al. (1996) and Rippke et al. (1995).

In this work, corn samples with moisture over 25% were omitted from the Constituent (but not the moisture) calibrations because very high moisture corn is unlikely to be accepted as a value-added product, regardless of its potential advantages if allowed to mature more fully. Likewise, corn oil contents under 4% (basis 15% moisture) were eliminated because the market is not discriminating corn of normal and low oil. This strategy reduced the spectral variability and cross correlation that had to be modeled in the calibration.

The Zeltex unit uses 12 NIR-emitting diodes in transmission, in the 800-1100nm range. Samples (approximately 250g) are contained in the cuvette, which is moved progressively in front of the light source to provide up to 22 subsamples per scan. Instrument and grain temperatures are measured for each sample, with these terms used as regressors along with scan data.

Calibrations were generated by Partial Least Squares, using Unscrambler 7.0 (Camo AS, Norway) and cross validation for prediction statistics. Because availability of samples with chemistry was not limiting, there was no sample selection done before calibration. This is counter to many reports of PLS calibrations, in cases where chemical data was not as readily available. It is our belief that with many samples, PLS will fill in more areas of the Principle Component space, and will then equal the performance neural networks, which also use very large databases (BoBuchmann, 1997). Each year, our samples are selected for entry into the calibration process by the Combination of Constituents Matrix developed by Rippke et al. (1995), so diversity without overloading is included at the outset.

Scans of 40 samples per gain at -15°, 5°, and 45° C were added for temperature stabilization. One unit was designated as the primary master; optical regression equations by wavelength were applied to adjust the co-master data to more closely resemble the master, but not perfectly so. Spectral standardization removes approximately 50% of the variation between units and inclusion of other units in the calibrations has been documented to be an effective process for desensitizing calibrations for transferability.

From the statistics generated, an operating procedure was designed for users. The objective was to lower cost unit as well as am initial screening tool only to increase testing frequency to compensate for inherent variabilities.

Table 3 shows the statistics obtained from the ZX800 calibration and subsequent test at the Zeltex factory with other units not included in the calibration. The comparison with the current calibrations for other whole grain analyzers, on the same data set, shows some loss of performance in selected factors, but little loss in others. Figure 2 shows an example prediction graph from Unscrambler for one characteristic, corn protein.

These results were encouraging in that only loss in accuracy was of a lesser percentage than could be gained by testing one additional sample per delivery (e.g., two samples versus one reduces error by 30%). If this extra time is not limiting, as it might not be in times other than harvest, then data accuracy could be similar among brands. Likewise, the ZX800 was capable of differentiating classes of grain easily, as in high oil versus normal corn. This means that a buyer could use a lower cost test at receival, collect a composite sample, and use the composite for actual numerical settlement. This is an approach used in dairy and livestock marketing where delayed payment of quality premiums is common.

Based on corn and soybeans samples from the Iowa State University database:

1 Research supported by Zeltex, Inc. and the Iowa Agriculture and Home Economics Experiment Station, Iowa State University, Ames, IA.
2 Post Doctoral Research Associate, Laboratory Manager, and Professor, Department of Agricultural and Biosystems Engineering, Iowa State University, Ames, IA; Technical Manager, Zeltex, Inc., Hagerstown, MD.