Skip to main content

USDA 2023: Sorghum Grain Protein by NIR — Handheld vs. Benchtop

92 sorghum grain samples from the 2021 Kansas and Texas harvest, protein measured by combustion nitrogen and predicted by two near-infrared instruments — a benchtop laboratory unit and a handheld field unit. Published by the USDA-ARS grain quality program as U.S. public-domain data; mirrored here because two of its facts matter to every gluten-free brewer: sorghum grain protein spans 6.4 to 16.8%, and screening it no longer requires a laboratory.
Get the data

CSV, full precision: usda-2023-sorghum-protein-nir.csv · Archive DOI: pending · Cite: Bean, S.R., Peiris, S., Wu, S. *Data from: Performance of a handheld Micro · Full terms and version history under Reuse.

This is an external dataset. The record of authority is the USDA deposit on Ag Data Commons — doi:10.15482/USDA.ADC/1529552, license U.S. Public Domain. This page mirrors it unchanged and adds the brewing-side reading; the originating work is Bean, Peiris, and Wu at the USDA-ARS Grain Quality and Structure Research Unit (Manhattan, Kansas), with the companion open-access paper by Peiris, Bean, Wu, Sexton-Bowser, and Tesso in Foods (2023), doi:10.3390/foods12163101 (CC BY 4.0).

Methods, briefly​

  • Samples: 2021-harvest grain — test-market hybrids from private seed companies and public breeding programs grown at dryland sites in Assaria, Kansas and Sunray, Texas, an irrigated site in Colby, Kansas, plus samples from a Manhattan, Kansas breeding program.
  • Reference measurement: combustion nitrogen (LECO FP-828p, AACC method 46-30.01), N × 6.25, reported as-is moisture basis.
  • Instruments compared: a Perten DA-7250 benchtop NIR and a VIAVI MicroNIR OnSite-W handheld, spectra collected on intact grain.
  • Design: 59 samples built the calibrations; 33 held-out samples tested them. The deposit carries both sets; this mirror merges them into one CSV with a sample_set column and changes nothing else.

Explore this data​

Sort, filter, and plot measured against predicted values. The full tables below remain the printable, citable record; this is the same data, interactive.

Interactive explorer92 samples

Loading the explorer…

The dataset​

Protein content, % (g/100 g grain), as-is moisture basis. Values exactly as deposited.

Calibration set59 samples
Measured, %Benchtop NIR, %MicroNIR, %
13.0913.0813.44
11.1811.2111.28
12.4212.3912.30
13.4413.3813.35
12.1312.2212.87
10.9711.1110.89
7.147.637.52
11.5811.8712.17
11.8611.8412.54
10.2910.6510.24
10.7710.3310.55
11.3710.8810.50
9.709.5610.03
11.0911.0610.32
9.619.379.33
11.6811.7111.61
12.5512.6312.84
12.0312.5411.63
9.779.559.51
16.7615.6814.75
14.9415.3314.88
6.967.316.96
7.037.126.59
8.188.238.13
7.247.497.66
6.416.326.51
7.767.367.62
8.287.997.87
8.618.418.72
7.957.808.13
7.437.447.37
15.7214.8814.85
6.826.486.71
8.208.178.08
12.2912.1912.77
9.008.628.35
7.487.777.83
12.0712.0311.63
11.4211.8611.67
10.379.9310.25
11.0110.2910.23
11.2111.0411.54
12.9413.1013.26
11.6111.5212.17
11.7712.3412.61
10.7810.6510.58
8.749.1910.41
15.6615.9916.31
6.737.297.84
6.976.997.33
8.258.527.59
7.867.847.78
7.266.926.49
8.037.817.71
15.9415.6414.77
8.408.598.50
14.8715.8415.14
7.168.538.87
8.187.507.15
Validation set33 samples
Measured, %Benchtop NIR, %MicroNIR, %
13.7913.9713.55
14.2214.7014.61
12.9513.3311.81
10.7610.289.39
12.9812.5611.37
13.2212.9712.20
12.6512.1712.72
11.1811.3011.45
9.5510.119.82
11.5511.8811.21
8.948.718.72
9.7610.429.70
9.118.808.81
9.6210.029.73
10.4711.7710.98
11.5011.8212.12
14.5815.2115.33
14.5515.0715.67
12.0512.8813.35
14.6114.7014.93
13.8314.6915.36
16.2215.8715.01
14.7615.3015.45
12.0210.7810.66
10.2510.0110.21
12.7612.9713.08
13.3913.5913.67
8.167.807.49
11.7912.3412.32
13.3313.2613.01
10.7511.1210.68
8.588.488.75
9.379.199.17

Summary statistics​

Computed from the dataset as deposited (mean absolute error and r² against the combustion reference; the paper reports its own validation statistics of R² 0.94 / RMSEP 0.52% for the benchtop and R² 0.87 / RMSEP 0.76% for the handheld):

SetnProtein range, %Benchtop MAE, ptsBenchtop r²MicroNIR MAE, ptsMicroNIR r²
Calibration596.41 – 16.760.300.9770.460.945
Validation338.16 – 16.220.430.9490.590.892

What these numbers mean for brewing​

  • Sorghum grain protein is a lot-level variable, and the spread is wide. 6.4 to 16.8% across one harvest year — a 2.6× range. Grain protein is the upstream driver of soluble protein and FAN in the malt, and FAN is the sorghum brewer's designed input (yeast nutrition): the Hartwick production malts ran 46–87 mg/L FAN partly because of where their grain sat on exactly this kind of spread. A brewer or maltster who doesn't know their lot's protein is guessing at the nutrition problem before the mash starts.
  • Screening no longer requires a laboratory. The study's point is instrumental: the benchtop unit predicted combustion protein at R² ≈ 0.95, and the handheld — a field instrument — at R² ≈ 0.89, both with mean absolute error under 0.6 protein points on held-out samples. The authors' conclusion is that the handheld is acceptable for screening intact grain. For a craft maltster sorting incoming sorghum lots, that is the difference between sending samples away and sorting at the truck.
  • The honest limits. This dataset predicts protein and nothing else — not diastatic power, not FAN, not extract. It is grain, not malt. And screening precision (±0.6 points) is sorting precision, not specification precision — a declared laboratory value on a certificate is still the number a contract should reference, which is exactly the declarations discipline GFB-SME requires.

Record​

Mirrored and first published on this site 2026-08-11. Corrections to date: none. Because this is an external record, corrections split two ways: an error in our mirror or reading notes is ours — report it here and it gets logged in this section. An error in the data itself belongs upstream — report it to the originating authors via the Ag Data Commons record, and if the deposit is revised, this mirror follows it (logged here).

Reuse: download, cite, tell us​

→ Download as CSV — both sets merged, sample_set column added, values unchanged.

The canonical files live in the USDA deposit: Ag Data Commons, doi:10.15482/USDA.ADC/1529552. This dataset is not separately archived on Zenodo — it already has a permanent federal DOI, and a second archive would add noise, not persistence.

Cite the originating deposit, not this page:

Bean, S.R., Peiris, S., Wu, S. Data from: Performance of a handheld Micro NIR instrument in comparison with laboratory benchtop NIR instrument for determining protein levels in harvested sorghum grain samples. USDA Ag Data Commons, 2023. https://doi.org/10.15482/USDA.ADC/1529552

For the analysis behind it: Peiris, K.H.S., Bean, S.R., Wu, X., Sexton-Bowser, S.A., Tesso, T. Performance of a Handheld MicroNIR Instrument for Determining Protein Levels in Sorghum Grain Samples. Foods 12(16):3101, 2023. https://doi.org/10.3390/foods12163101

If the brewing-side reading here was useful, this page can be cited as: Gluten Free Brewer, "USDA 2023: Sorghum Grain Protein by NIR," 2026, https://glutenfreebrewer.com/research/external/usda-2023-sorghum-protein-nir

Source and Validation Notes

Mirrored programmatically from the two CSV files in the USDA deposit (calibration, 59 rows; validation, 33 rows) on 2026-08-11: files downloaded from the deposit's own download URLs, merged with a sample_set column, byte-order marks and empty annotation columns removed, no value changed. The deposit's license is U.S. Public Domain; the mirror and these notes carry this site's usual CC BY 4.0. The originating laboratory is named because the work is published federal research; no endorsement of this site by USDA-ARS or the authors is implied.

Research needed (2)