ABSTRACT

The development of a rapid, precise and robust instrumental method to evaluate oil content in oilseeds is of major interest to growers, processors and oilseed breeders. Near infrared reflectance spectroscopy (NIRS) is routinely used for the prediction of oil content in canola crops at the Oilseed Analytical Laboratory at Horsham, Victoria. The laboratory analyses approximately 55,000 seed samples of canola cultivars and breeder's selections each year using NIRSystems, Model 5000 monochromator. This has allowed the establishment of a large database containing a broad range of spectral information.

Sample selection for inclusion in the calibration set was based entirely on the exploration of variation in the collected spectra. All calibrations were developed using full spectrum information with math treatment including correction for scatter and transformation into second derivative. The equations were calculated using a modified PLS algorithm and cross-validation technique.

Initial calibrations have shown a good agreement between the reference data and NIRS predicted values for oil content in canola seed (SEC = 0.69, R² = 0.94, SECV = 0.83, 1-VR = 0.91). These calibrations were specific to the environmental conditions under which the seed was grown and required continuous maintenance and expansion to account for the effects of crop location and time of sowing.

In the last three years spectrally unique information was systematically included in the calibration set of samples. The larger variance included in the calibration set has allowed the development of more robust calibrations for the prediction of oil content without notable loss of precision.

INTRODUCTION

Near Infrared Reflectance Spectroscopy (NIRS) is becoming the favoured analytical technique to determine oil content in canola seed. This is due to the relative ease of sample preparation and the flexibility of sample presentation. With the use of advanced spectrometrics and chemometrics software the selection of calibration samples is based on relevant spectral information; this is important when establishing robust calibrations. There is still, however, a need to conduct primary analyses on samples for the development, validation and maintenance of calibrations (Shenk and Westerhaus, 1995).

The two main techniques, recommended by the American Oil Chemists Society (AOCS), for analysis of oil content in canola are Soxhlet extraction and Nuclear Magnetic Resonance (NMR) (AOCS, 1996). The Soxhlet technique is the official method, according to AOCS standards, and involves solvent extraction. This technique includes modified versions of the Soxhlet extraction method: Soxflo (Scientific and Technical Supplies Ltd, England) and Soxtech (Tecator, Sweden). Broad Spectrum Nuclear Magnetic Resonance (NMR) is also recommended for determining oil content in canola using both pulse and continuous wave designs. The pulse wave NMR is the preferred method as it is non-destructive and can simultaneously measure oil and moisture content in whole seed canola (Rossell and Pritchard, 1991).

NIRS calibrations have shown to be sensitive to year, time of sowing, crop location and variety (Dardenne, 1996). To date, separate calibrations have been developed for each season, which have required systematic maintenance and expansion. During the past three years, the collection of a large library of spectrally unique samples has enabled the development of a more robust calibration for oil in whole seed canola.

This study was undertaken to determine if a combined season calibration set, with significant sample variation, could produce a more robust calibration to determine oil content without loss of precision.

MATERIALS AND METHODS

Collection of samples

Canola seed samples were collected across the southern part of Australia from the 1995, 1996, 1997 and 1998 seasons. Approximately 5000 canola seed samples were also obtained from Canada and Europe.

Collection of reference data

Analysis of oil content by Soxflo Seed samples were ground using a Model A-10 (Janke and Kunkel, Germany) water cooled sample mill.

A modified Soxhlet extraction apparatus, Soxflo (Scientific and Technical Supplies Ltd, England), was used to extract the oil from the prepared samples to determine oil content. It was established that a solvent to meal ratio of 10:1, (v/w), was required for complete extraction of lipids.

Analysis of oil content by NMR A broad spectrum, pulsed wave Model NMS 100 Minispec (Bruker Pty Ltd Scientific Instruments, Germany) NMR analyser was used as another method for the determination of oil content in the whole seed canola. The unit was calibrated using the data obtained by solvent extraction.

Collection of spectra

Clean whole seed samples were scanned using a Model 5000 monochromator (NIRSystems, USA) working in reflectance mode. Spectra were collected across the 1100 to 2500 nm spectral range in 2 nm wavelength increment.

Spectra were corrected for scatter and transformed into their respective second derivatives using the chemometrics software, Infrasoft International (ISI). The algorithm CENTER was used for the calculation of principle components and Mahalanobis distance. This information was used for the description of spectral boundaries and detection of outliers (Shenk and Westerhaus, 1995). Following centering of the data, the SELECT algorithm was used for grouping the spectra into clusters with neighbourhood H distance of 0.6. One representative spectrum from each cluster was selected for the arrangement of the calibration set of samples.

Calibration development

Selected spectra were matched with the reference data and calibration models were developed using PLS (partial least squares) algorithm and cross-validation technique. Calibration performance was assessed by SEC (standard error of calibration), R² (correlation coefficient) and SECV (standard error of cross-validation) (Shenk and Westerhaus, 1993).

RESULTS AND DISCUSSION

Comparison of Soxhlet and NMR for oil content analysis

Soxhlet extraction is the official method of determining oil content in oilseed crops (AOCS, 1996), however, the procedure is lengthy, laborious and relatively hazardous. With the expansion of canola breeding programs there is a need to analyse thousands of samples using rapid and reproducible techniques. The amount of seed available for Soxhlet extraction is also often limited. These factors have led to the replacement of solvent extraction, for the determination of oil content, with methods such as broad spectrum NMR and NIRS. These methods both use whole seed, which enables seed to be retained for further analysis and planting. NMR and NIRS are both classified as secondary analytical techniques, therefore, Soxhlet extraction is still used for the collection of reference data for calibration of NMR and NIRS instruments and validation of their performance.

Analytical data obtained by secondary techniques for the purpose of calibration can often be interpreted to generate inconsistent information. Secondary techniques have been successfully and widely implemented due to their speed and non-destructive nature (McGregor, 1990). Oil content was determined for a set of 10 randomly selected canola samples using modified Soxhlet extraction (Soxflo) and NMR. The results obtained from this experiment are presented in Table 1. The residuals indicate a good agreement between these two methods showing that both techniques can be used to collect reference data for calibration of the NIR instruments.

Table 1: Analysis of oil content in canola samples by Soxflo and NMR

Sample No.	Soxflo Oil Content	NMR Oil Content	Residuals
1	47.60	47.11	0.49
2	44.96	44.26	0.70
3	49.75	49.54	0.21
4	49.20	49.68	0.48
5	41.85	40.65	1.20
6	47.10	47.86	0.76
7	47.12	46.45	0.67
8	46.57	45.58	0.99
9	45.54	46.32	0.78
10	47.54	47.74	0.20

Bias of Residuals = 0.65; Std Dev. Residuals = 0.32.

Development of NIRS calibrations

Traditionally NIRS calibration sets have been selected on the basis of chemical data rather than spectral properties, which required samples to be analysed initially by wet chemistry (Harman et al, 1997). More recently, the implementation of advanced spectroscopic and chemometrics software has allowed for the selection of calibration samples based entirely on spectral information.

Calibration statistics listed in table 2, for individual and combined years, showed good agreement between reference and NIRS predicted values for oil content. Higher SEC values were recorded for the 1996 and 1997 calibration sets, which may have been due to the range of oil content in canola seed samples being broader. Samples from 1995, 1996 and 1997 calibration sets were combined for a total of 223 samples; from this, 91 samples were selected as being spectrally unique. These samples were used for the development of the combined calibration, which had an oil content ranging from 32.46 per cent to 50.64 per cent, SEC of 0.61 and R² of 0.98. The combined calibration showed that accuracy was maintained and there was a decrease in the values of standard error of calibration and cross validation. It should also be noted that by using less than half of the selected samples for the combined calibration, performance was improved by comparison with calibrations developed for individual years.

Table 2: Calibration statistics for oil content in canola samples for individual and combined years

Year	N	Mean	Range (% oil)	SEC	R2	SECV	1-VR
1995	52	44.24	35.10-50.37	0.69	0.94	0.83	0.91
1996	49	43.31	29.52-50.82	0.86	0.98	0.93	0.98
1997	122	41.62	28.56-48.66	0.75	0.95	0.88	0.94
1995/6/7	91	43.37	32.46-50.64	0.61	0.98	0.77	0.96

N = number of samples; Mean = average oil content; SEC = standard error of calibration; R² = coefficient of determination of calibration; SECV = standard error of cross validation; 1-VR = coefficient of determination of cross validation; 1995/6/7 = Combined calibration.

Figures 1 to 4 show a graphical presentation of the relationship between reference values and NIR predictions of oil content in canola for the individual (1995, 1996, 1997) and combined calibration sets.

Figure 1: Correlation between reference and NIR predicted values for oil content in 1995 season canola.

Figure 2: Correlation between reference and NIR predicted values for oil content in 1996 season canola.

Figure 3: Correlation between reference and NIR predicted values for oil content in 1997 season canola.

Figure 4: Correlation between reference and NIR predicted values for oil content in 1995, 1996 and 1997 season canola.

Performance of the combined calibration

The combined calibration was validated using randomly selected canola samples from 1998. Table 3 shows a good agreement between values for oil content analysed by NMR and predicted data using the combined calibration. Calculated bias and standard deviation of residuals, 0.37 and 0.17 respectively, indicate that calibration accuracy is maintained for the prediction of oil content for the 1998 season.

Table 3: Comparison of NIR predicted and reference (NMR) values of oil content in 1998 canola samples

Sample No.	Predicted Oil Content	NMR Oil Content	Residuals
1	41.34	41.91	0.57
2	36.65	36.45	0.20
3	36.27	36.50	0.23
4	43.73	43.25	0.48
5	39.53	38.94	0.59
6	36.55	36.98	0.43
7	38.82	38.99	0.17
8	40.00	39.73	0.27
9	41.43	40.87	0.56
10	46.56	46.76	0.20

Bias of Residuals = 0.37; Std Dev. Residuals = 0.17.

CONCLUSION

Performance of the combined calibration developed with samples from 1995, 1996 and 1997 demonstrates that a universal calibration for the prediction of oil content in canola is possible. This would ultimately result in a reduction of wet chemistry necessary for development and maintenance of calibrations. Further investigation, however, is required to determine whether the combined calibration can be used to predict oil content in canola for subsequent years. The large diversity in sample populations from one season to another may introduce further modification of variance, which can often affect the stability of calibrations.

REFERENCES

1. American Oil Chemists Society (1996) AOCS official methods and recommended practices, Method No. Ak3-94 and Ba3-38.

2. Dardenne, P. (1996) Stability of NIR spectroscopy equations. NIR News. v.7, no.5, pp.8-9.

3. Harman, C.F., Filip, M.L. and Golebiowski, T. (1997) An Investigation into the Possible Application of Near Infrared Spectroscopy at Delivery Points. In Proceedings of the AOCS Australasian Section Workshop. Canberra, Australia, February 27-28.

4. McGregor, D.I. (1990) Application of Near Infrared to the Analysis of Oil, Protein, Chlorophyll and Glucosinolates in Canola/Rapeseed. Canola and Rapeseed – Production, Chemistry, Nutrition and Processing Technology. Chapter 13, pp.220-228.

5. Rossell, J.B. and Pritchard, J.L.R. (1991) Analysis of Oilseeds, Fats and Fatty Foods. Elsevier Science Publishers Ltd: London (Chapter 2, pp.48-53).

6. Shenk, J.S. and Westerhaus, M.O. (1993) Analysis of Agriculture and Food Products by Near Infrared Reflectance Spectroscopy. Intrasoft International: Port Matilda, pp 63-66.

7. Shenk, J.S. and Westerhaus, M.O. (1995) NIRS 3, Version 3.10 Routine Operation, Calibration Development and Network System Management Software for Near Infrared Instruments. Intrasoft International: Port Matilda, pp 100, 112, 251-257, 279.