Influence intact lupin hull (Lupinus angustifolius) has on near infrared predictions for protein and moisture by comparing wholegrain and grist calibrations

Abstract

Lupins (Lupinus angustifolius) are a type of pulse known for their high protein content, nutritional benefits and versatile uses in food. Lupins have a characteristic thick outer hull which may impact the development and application of NIR (near infrared) calibrations for assessing compositional traits. This study seeks to compare the predictive abilities of models based on wholegrain and ground grist for key quality traits, specifically protein and moisture content. To determine if the thick outer hull poses an obstacle for NIR light to penetrate and reflect from the seed, NIR spectroscopy of lupin samples was conducted for both wholegrain and ground grist forms. The same set of 120 samples was used to construct two prediction models (with a calibration set of 96 samples and an independent validation set of 24 samples). The results show that the predictions for moisture content were comparable between wholegrain and grist for the validation set, achieving r² values of 0.99 for both models. The models for protein content exhibited a greater distinction between wholegrain and grist, yielding r² values of 0.91 and 0.96, respectively. Overall, both sets of models exhibited high correlations with protein and moisture constituents and are indicative that calibrations developed on intact lupin hull using wholegrain NIR spectra can be used to predict protein content. Despite a decline in accuracy when compared with the grist results, the benefit of applying the wholegrain calibration outweighs the time taken to prepare the grist samples for NIR analysis.

Keywords

australian sweet lupin narrowleaf lupin hull NIR protein wholegrain grind

Introduction

Lupinus angustifolius, commonly known as the Australian sweet lupin or narrowleaf lupin, has become increasingly recognised in recent years due to its notable nutritional profile and versatility in food applications. Lupins and other pulses high in protein content are becoming increasingly sought after as viable protein-rich alternatives to meet dietary needs.¹

Australian sweet lupin seeds have high protein content, dietary fibre, essential unsaturated fatty acids, and are low in starch content.² The sweet lupin is generally low in alkaloid content resulting in reduced bitterness compared with other lupin species.³ Australian lupin production reached ∼1.3 million tonnes in 2023; an increase of ∼33% from 2022.⁴ The increase in lupin cultivation has emphasised the importance of establishing efficient and reliable methods to assess lupin composition, particularly in large-scale breeding programs and production areas. Near-infrared (NIR) spectroscopy is ideal for routine analysis of grain quality traits, with several advantages such as being non-destructive, rapid, and capable of handling high throughput. This capability supports breeders in identifying optimal breeding lines that exhibit higher protein content.⁵ Previous studies have explored NIR scanning of whole seed lupins and lupin kernel meal for various compositional traits, yielding varied results.^6–9

Lupins are characterized by a thick hull or seed coat, and this could pose a challenge for NIR analysis. The lupin seed coat represents on average 24% of the seed weight, and seed coat thickness can vary between ∼257 and 335 µm, dependent on variety and environmental conditions.^10,11 The thick hull, and high proportion of hull to cotyledon, may affect the penetration depth of NIR light which is dependent on multiple factors such as the physical proportions of the seed.¹² This could potentially lead to variations in the accuracy of lupin wholegrain predictions, especially as the seed coat only contains 3–4% protein content.¹³ This study aims to explore whether the intact lupin hull has a negative impact on NIR predictive accuracy for quality traits, protein and moisture content, by constructing two calibration models, intact wholegrain and ground grist, and comparing the statistical performance of predicting an independent validation set.

Methods

Sample processing

A range of Australian lupin germplasm were acquired from lupin historical trials, commercially released varieties, spanning across harvest years 2015, 2017, 2020, and 2023 were used for calibration construction. The samples were sourced from the primary sweet lupin production regions in Victoria and South Australia. These samples, which were harvested in the dry season, were stored in screw lid plastic containers at ambient room temperature (∼16–21°C) in laboratory seed store. NIR spectra were acquired using a XDS Rapid Content Analyzer (FOSS, Hillerød, Denmark), which employs diffuse reflectance to capture wavelengths from 400 to 2498 nm. Prior to NIR spectroscopy, samples were moved to the same room as the XDS and let to equilibrate at room temperature (21 ± 1°C) for 24 h before analysis. Wholegrain scans were collected in triplicate, from ∼50 to 60 g samples, using a rectangular cell (repacked between scans).¹⁴ Subsequently, lupins were subsampled and ground into grist using a Laboratory Mill 3303 (Perten Instruments, Stockholm, Sweden). Particle size distribution was determined via mechanical sieving (AACC International Method 55–06.01), and found to average 238 µm, with the largest particles of 280 µm.¹⁵ The grist was then packed into ring cup cells (∼5 g sample) and scanned in triplicate (repacked between scans). Following this, lupin grist was analysed for moisture and protein (dry basis) content on the same day, with each sample run in duplicate. Moisture content was determined according to Moisture-Air Oven Methods (AACC International Method 44–15.02) via thermogravimetric analysis (TGM 8000, LECO Corporation, Saint Joseph, USA).¹⁵ For protein content determination, the Crude Protein-Combustion Method (AACC International Method 46–30.01) was modified to use a furnace operating at 1200°C with a nitrogen analyser (FP928, LECO Corporation, Saint Joseph, USA).¹⁵

Calibration development

WinISI software (v 4.6.11.14874, FOSS, Hillerød, Denmark) was used for calibration construction. The calibration set (n = 96) for model development consisted of samples grown prior to 2023, while the independent blind validation set (n = 24) consisted of samples grown in a 2023 trial at a single site. The validation set consists of six commercial varieties by four replicates grown at Gerang Gerung, Victoria in sandhill soil over 2023–24 season. While the environment and geographical information of the validation set is not represented in the calibration set, the six commercially released varieties are represented.

For the calibration set, spectra of the wholegrain and grist, the NIR calibration models were constructed using modified partial least squares (mPLS) regression with the following parameters: wavelength region: 1130–2450 nm; scatter: SNV and detrend; derivative: 1, gap: 4; smooth: 4; smooth 2: 1. The modified partial least squares calibration for moisture and protein used six spectral groups, with each group optimized individually for the number of PLS terms. For protein, the average number of mPLS terms selected during cross-validation across all groups was four, while for moisture it was six. A total of six PLS components were used in the final calibration models. The performance of each model was tested using the NIR prediction for the validation set. For calibration construction NIR spectra was collected in triplicate, averaged into a single spectra and assigned a constituent reference value.

Statistical analysis

Two-way ANOVA analysis was conducted using Genstat 24th edition (v24.1.0.1041, VSNi, Hemel Hempstead, UK) to compare the performance of both applications (wholegrain and grist) for each constituent (protein and moisture), using the regression residuals (reference – predicted) for each sample (n = 24).

Results & discussion

Moisture

The NIR regression models for moisture content showed consistent performance for both wholegrain and grist. The wholegrain model achieved low calibration errors for moisture with an SEC of 0.05 and SECV of 0.07, alongside R² and 1-VR values of 0.99 indicating a tight fit to the calibration data (Table 1, Figure 1(b)). The wholegrain model also performed well on the independent validation set with SEP of 0.18, and demonstrated its suitability for quantitative predictions with RPD of 3.2. Similar to wholegrain, the grist model exhibited strong correlations in addition to a notably low SEP of 0.05 and RPD >5. Both models exhibited highly replicable and precise moisture predictions for the validation set with low standard deviations, averaging below 0.04% (Table A1). The high performance for predicting moisture may be partly due to the narrow range of the validation set, 9.7–12.0%, which is around the expected harvest moisture content (Table A1). Although there were four prominent outliers in the wholegrain validation set with higher residuals, these did not substantially affect the overall performance of the model (Figure 2(b)). Both models proved to be effective tools for moisture content prediction, and the consistency of the predicted data across triplicate scans further supports the robustness of the NIR models (Figures 1(b) and 2(b)). While grist demonstrated slightly better precision, the overall difference between the two methods is not substantial enough to impact routine analysis. This suggests that either wholegrain or grist could be used for effectively predicting moisture content, with no impact on calibration performance from the lupin hull in the wholegrain application.

Table 1.

PLS regression statistics for standard equation and validation set.

Model	Constituent	N_cal	SEC_cal	R²_cal	SECV_cal	1-VR_cal	N_val	r²_val	SEP_val	RPD_val
Wholegrain	Moisture% (w/w)	96	0.05	0.99	0.07	0.99	24	0.99	0.18	3.9
Wholegrain	Protein% (w/w)	96	0.73	0.92	0.88	0.88	24	0.91	0.92	2.4
Grist	Moisture% (w/w)	96	0.06	0.99	0.08	0.98	24	0.99	0.05	14.3
Grist	Protein% (w/w)	96	0.62	0.94	0.67	0.93	24	0.96	0.44	5.1

N – sample size.

SEC – standard error of calibration.

R², r² – coefficient of determination.

SECV – standard error of cross validation.

1-VR – variance subtracted from one.

SEP – standard error of prediction.

RPD – residual prediction deviation (ratio of prediction to deviation).

Protein

The NIR models for protein content in lupin samples performed comparably well for wholegrain and grist applications, with both models exhibiting strong correlations of R² > 0.9 (Table 1). Protein content predictions displayed greater variation than the moisture content, which may be due to sample heterogeneity, as protein accumulates in the seed during grain development and maturity. Seed composition in any crop can vary greatly, and is a function of interactions between genetics, environmental factors and management practices.¹⁶ Although the spread between protein reference and predicted values were higher for some samples, the residual error remained within acceptable limits for both applications (Figures 1(a) and 2(a)). While some samples exhibited higher residuals in one method, the pattern of residuals and the specific samples with elevated residuals were observed to be a random response between the two methods (Table A1). The wholegrain protein model displayed elevated standard error in calibration and prediction sets, compared to the grist model (Table 1). Two-way ANOVA on prediction residuals showed a significant effect of the model used for both moisture and protein traits (p < 0.01), indicating consistent differences in model performance (Table 2). There were no significant effects of Variety or Method*Variety interactions, suggesting that prediction accuracy and the relative performance of the models were consistent across all varieties (Table 2). These results indicate that the grist model outperformed wholegrain in producing more accurate predictions, as shown by consistently lower residuals across traits and varieties (Table 2, Table A1). Despite a SEP difference of 0.5% between models (Table 1), the economic benefits of using NIR scanning on whole seeds to exercise selection pressure on breeding lines would facilitate higher throughput and non-destructive testing, which outweighs the impracticality of grinding large quantities of samples for processing.⁵ Despite this, both protein models exhibited strong correlations with acceptable residuals, and using either a wholegrain or grist model for practical applications such as screening lupin breeding lines for elevated protein content would be similarly performant with minimal differences.

Table 2.

Two-way ANOVA analysis between NIR prediction method (wholegrain and grist) residuals across six lupin varieties (n = 24).

Constituent	Source of variation	d.f.	s.s.	m.s.	v.r.	F pr.
Protein% (w/w)	Method	1	4.582	4.582	11.97	0.001
	Variety	5	2.221	0.444	1.16	0.347
	Method * variety	5	1.839	0.368	0.96	0.454
Moisture% (w/w)	Method	1	0.217	0.217	17.68	<.001
	Variety	5	0.017	0.003	0.28	0.922
	Method * variety	5	0.111	0.022	1.80	0.138

d.f – degrees of freedom.

s.s. – sum of squares.

m.s. – mean square.

v.r. – variance ratio.

F pr. – F probability/p-value.

Insights & implications

Overall, both models performed well in predicting moisture and protein content with acceptable residual errors. There was little difference between grist (protein, R² = 0.94) and wholegrain (protein, R² = 0.92) calibrations in lupin (Table 1). This finding is consistent with previous NIR research findings that demonstrated for major constituents there were little difference in R² and error (SEP) between ground and wholegrain malt calibrations.¹⁷ Due to the small difference between these models, it can be inferred that the lupin hull did not adversely impact the prediction of protein or moisture in the wholegrain lupins.

In a practical industry-focussed setting external to laboratory research purposes, such as on-farm sensing or grain receival, it is difficult to justify additional labour to grind and process lupins for a minimal improvement in model predictions. Another benefit of wholegrain scanning is that even low volumes of seed can be predicted non-destructively, which is useful for screening breeding lines. The differences between grist and wholegrain predictions could be viewed as effectively negligible when compared with the resources spent to prepare large quantities of grist samples for scanning.

To determine the impact of the lupin hull on NIR predictions, this study focused on evaluating the prediction accuracy and precision for major constituents, protein and moisture, which comprise the majority of the composition. Further research could primarily investigate how well the grist and wholegrain NIR models perform in predicting other, minor constituents that make up a smaller portion of the cotyledon composition. To extend upon these research findings, investigation could be conducted into how fibre, lipoxygenase, protein composition, alkaloids and phenolic compounds predict through the lupin hull for wholegrain NIR. For future work in this area, it would be ideal to expand the number of varieties of sweet lupin and include a range of growing environments from Western Australia, South Australia, Victoria and New South Wales, as different varieties will not only have different compositions, but may also have different hull thickness levels depending on the growing conditions and soil types.

Conclusion

NIR models were constructed and found to be feasible for predicting protein and moisture content in both lupin wholegrain and grist applications. Both constituents exhibited strong correlations in both models and are indicative that the intact lupin hull did not negatively impact predictions for moisture or protein content. While the grist model displayed a better predictive performance, the accessibility of wholegrain scanning would likely be more suitable in agricultural applications such as screening breeding lines.

Footnotes

Acknowledgements

This research was funded as part of the “DEE2312-002RTX Capitalising on pulse protein food and feed opportunities domestically and internationally” project, a collaboration between Agriculture Victoria and the Australian Grains Research and Development Corporation (GRDC). The authors wish to acknowledge the Crop Quality laboratory at the Horsham SmartFarm, Dr Jason Brand and Dr Audrey Delahunty for their support in analysing and growing the trials.

ORCID iDs

Nicholas Rigas

Pankaj Maharjan

Cassandra Walker

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by Agriculture Victoria and the Australian Grains Research and Development Corporation (GRDC) through project DEE2312-002RTX Capitalising on pulse protein food and feed opportunities domestically and internationally.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Appendix

Table A1.

NIR average predictive values for moisture and protein across 24 blind validation lupin samples.

Sample ID	Variety	Lab protein% (w/w) drybasis^a	NIR_WG protein% (w/w) drybasis	NIR_Gr protein% (w/w) drybasis	Lab moisture% (w/w)^a	NIR_WG moisture% (w/w)	NIR_Gr moisture% (w/w)
1	Uniharvest	34.3	34.5 ± 0.27	33.9 ± 0.34	11.7	11.5 ± 0.05	11.8 ± 0.03
2	Coyote	31.1	31.1 ± 0.38	31.0 ± 0.04	9.9	9.9 ± 0.02	9.9 ± 0.04
3	PBA Bateman	28.6	29.9 ± 0.35	29.2 ± 0.23	10.1	10.1 ± 0.05	10.1 ± 0.01
4	PBA Gunyidi	31.4	32.0 ± 0.75	31.9 ± 0.40	10.1	10.1 ± 0.02	10.1 ± 0.05
5	Lawler	29.3	29.8 ± 0.20	29.6 ± 0.37	10.2	10.2 ± 0.05	10.2 ± 0.2
6	Mandelup	29.2	29.7 ± 0.41	29.2 ± 0.11	10.8	10.6 ± 0.02	10.8 (SD < 0.01)
7	PBA Bateman	28.8	29.2 ± 0.66	29.1 ± 0.23	10.7	10.6 ± 0.03	10.6 ± 0.02
8	Uniharvest	34.8	36.2 ± 0.38	34.8 ± 0.27	12.0	11.6 ± 0.06	12.1 ± 0.03
9	Coyote	27.9	28.3 ± 0.12	28.4 ± 0.02	10.2	10.2 ± 0.02	10.2 ± 0.02
10	PBA Gunyidi	32.3	32.5 ± 0.68	32.3 ± 0.27	11.9	11.5 ± 0.01	11.9 ± 0.01
11	Mandelup	28.5	29.1 ± 0.26	28.3 ± 0.09	10.7	10.6 ± 0.02	10.7 ± 0.02
12	Lawler	31.1	31.7 ± 0.31	30.7 ± 0.43	10.5	10.4 ± 0.02	10.5 ± 0.04
13	Mandelup	30.2	31.2 ± 0.61	29.9 ± 0.13	10.7	10.6 ± 0.06	10.8 ± 0.03
14	Uniharvest	35.2	37.7 ± 0.80	34.6 ± 0.33	10.2	10.2 ± 0.02	10.2 ± 0.01
15	Lawler	31.4	30.4 ± 0.20	30.7 ± 0.20	10.3	10.2 ± 0.02	10.2 ± 0.03
16	Coyote	29.9	29.8 ± 0.38	29.5 ± 0.17	10.3	10.4 ± 0.02	10.3 ± 0.02
17	PBA Gunyidi	30.9	31.3 ± 0.45	31.0 ± 0.10	9.9	9.9 ± 0.01	9.9 ± 0.04
18	PBA Bateman	30.0	30.0 ± 0.63	29.3 ± 0.25	12.0	11.7 ± 0.03	12.0 ± 0.11
19	Coyote	30.8	30.0 ± 0.53	29.9 ± 0.20	11.0	10.9 ± 0.02	10.9 ± 0.05
20	Lawler	29.9	31.6 ± 0.62	30.5 ± 0.61	10.4	10.3 ± 0.02	10.4 ± 0.08
21	Uniharvest	36.9	37.1 ± 0.72	36.7 ± 0.14	9.7	9.8 ± 0.04	9.7 ± 0.04
22	PBA Bateman	30.3	31.2 ± 0.44	30.7 ± 0.13	10.5	10.4 ± 0.01	10.5 ± 0.07
23	PBA Gunyidi	31.1	32.4 ± 0.73	31.5 ± 0.42	11.5	11.3 ± 0.05	11.5 ± 0.04
24	Mandelup	29.6	29.8 ± 0.54	29.3 ± 0.17	11.9	11.5 ± 0.06	11.9 ± 0.03
	Mean	31.0	31.5	30.9	10.7	10.6	10.7
	Range	27.9–36.9	28.3–37.7	28.3–36.7	9.7–12	9.8–11.7	9.7–12.1
	Pooled repeatability (SD)		0.51	0.27		0.03	0.05

Reference testing was conducted in duplicate and reported as average.

NIR values are displayed as mean of triplicate scans ± standard deviation (SD) reported to two decimal places to better reflect measurement variation.

WG refers to wholegrain.

Gr refers to grist (ground lupin).

^aThe Lab Protein & Lab Moisture reference values have a standard error (SET) of ±0.3% and ±0.2% respectively.

References

Erbersdobler

Barth

Jahreis

. Legumes in human nutrition nutrient content and protein quality of pulses. Ernahrungs Umsch 2017; 64: 134–139.

Smulikowska

Wasilewko

Mieczkowska

. A note on the chemical composition of the cotyledons and seed coat of three species of sweet lupin. J Anim Feed Sci 1995; 4: 69–76.

Sedláková

Straková

Suchý

, et al. Lupin as a perspective protein plant for animal and human nutrition - a review. Acta Vet 2016; 85: 165–175.

FAOSTAT . Crops and livestock Australia production quantity. 2023. https://www.fao.org/faostat/en/#data/QCL, accessed 20 3 2025.

Walker

Panozzo

. Near-infrared spectroscopy enables quality selection in wheat breeding. Cereal Chem 2023; 100: 1347–1356. https://doi.org/10.1002/cche.10717

Glencross

Bourne

Hawkins

, et al. Using near infrared reflectance spectroscopy (NIRS) to predict the protein and energy digestibility of lupin kernel meals when fed to rainbow trout, Oncorhynchus mykiss. Aquac Nutr 2015; 21: 54–62.

Jansen

Jugert

Ordon

. High-throughput screening for protein content in blue, yellow and white lupins. In: 64th Conference of the Vereinigung der Pflanzenzüchter und Saatgutkaufleute Österreichs, Raumberg-Gumpenstein, Austria. Mutagenese und Hochdurchsatz-Screening CORNET Projekte. 2014: 19–20.

Jiang

Keaney

Ryan-Salter

, et al. Developing NIRS calibrations to predict the nutritional value of russell lupin (Lupinus polyphyllus). In: 16th Australian near infrared spectroscopy group conference, Coffs Harbour, NSW, 2014.

Ferrari

Barzaghi

Annicchiarico

. Development of NIRS calibrations for seed content of lipids and proteins in contrasting white lupin germplasm. In: Chu

Guo

Huang

, et al. (eds). Sense the real change: proceedings of the 20th international conference on near infrared spectroscopy, Singapore, 09 August 2022: Springer Nature, pp. 132–136.

10.

Clements

Dracup

Galwey

. Effect of genotype and environment on proportion of seed hull and pod wall in lupin. Aust J Agric Res 2002; 53: 1147–1154.

11.

Miao

Fortune

Gallagher

. Anatomical structure and nutritive value of lupin seed coats. Aust J Agric Res 2001; 52: 985–993.

12.

Krogh Mortensen

Gislum

Jørgensen

, et al. The use of multispectral imaging and single seed and bulk near-infrared spectroscopy to characterize seed covering structures: methods and applications in seed testing and research. Agriculture 2021; 11: 301.

13.

Mera

Jerez

Rouanet

, et al. Seed coat specific weight in lupinus angustifolius: influence of genotype and environment and relationship with seed coat proportion. Aust J Agric Res 2004; 55: 1189–1195.

14.

Williams

Dardenne

Flinn

. Tutorial: items to be included in a report on a near infrared spectroscopy project. J Near Infrared Spectrosc 2017; 25: 85–90.

15.

AACC International . Methods: 46-30.01 (crude protein-combustion method), 44-15.02 (moisture-Air oven methods), 55-60.01 (guideline for determination of particle size distribution). Approved methods of analysis. 11th ed. Cereals & Grains Association, 1999.

16.

Assefa

Purcell

Salmeron

, et al. Assessing variation in US soybean seed composition (protein and oil). Front Plant Sci 2019; 10: 298.

17.

Nielsen

Munck

. Prediction of malt quality on whole grain and ground malt using near infrared spectroscopy and chemometrics. In: Near infrared spectroscopy: proceedings of the 9th international conference (eds Davies

AMC

Giangiacomo

), Chichester, Great Britain, 1999: NIR publications, pp. 709–713.