Comparison of NIRS techniques to determine caffeine content in tea

1st April 2013

Near-infrared spectroscopy (NIRS) is an analytical tool that is used in many different fields, and many applications have already been described in the areas of agriculture, foodstuffs, cosmetics and pharmaceuticals. Here, Angela Schmidt, Brian P Davies, Dr Grit Schulzki, and Dr Andrea Giehl describe how NIRS used to monitor caffeine levels in tea.

Tea is decaffeinated mainly for deliveries to the American market. Quality control requires that production batches are monitored continuously for their remaining caffeine content in order to immediately halt the expensive extraction process if deviations should arise. Thus, the emphasis is on the quantitative analysis of very low caffeine content.

Up to now, an high performance liquid liquid chromatography (HPLC) method has been used, which is not only time consuming but also requires chemical solvents. In this article it will be shown that caffeine at <1 per cent by weight in decaffeinated tea samples can be analysed by FT-NIR spectroscopy.

NIR-spectra were acquired with a Bruker Vector 22/N-C FT-NIR spectrometer with a standard sample compartment for transmission measurements and the possibility to attach external accessories to the right side of the instrument. In order to measure the tea samples in diffuse reflection mode a fibre-coupling unit as well as an external integrating sphere module were tested.

In order to obtain reproducible spectra from inhomogeneous tea matrices, 16 scans at a spectral resolution of 8cm-1 were acquired in 9s for both experiments. The solid sample was irradiated with modulated NIR-light from the interferometer in the Vector 22/N-C; however, the diffusely scattered light from the sample was collected differently, according to the accessory used.

For one set of experiments a powder probe is connected by a fibre optic bundle to the front of the unit. The fibre bundle consists of two sets of conducting fibres: one set is used for light emission, the other set for light collection; both types are distributed statistically in the bundle. The NIR-light is guided from the spectrometer to the sample and then, after interaction, back to the detector located in the fibre coupling unit. The advantage is that remote measurements can be performed directly in the sample container.

The integrating sphere module allows sampling without direct contact with the sample in a quartz cup. The sample is irradiated with NIR-light, and the reflected light is totally collected by the integrating sphere and becomes independent from the orientation of the individual sample particles. Finally, the light is detected by a detector element, which is part of the integrating sphere module. This method has the advantage of a larger sampling area compared with that viewed by the powder probe. Furthermore, measurements are enhanced through the use of a rotating sample cup placed off-centre over the sample area with a diameter larger than this area, resulting in improved reproducibility from complex or coarse sample matrices.

The calibration model was based on a set of 12 ground tea samples, containing 0.7­4.2 per cent caffeine by weight (Plantextrakt GmbH&Co.KG, Vestenbergsgreuth/Germany). The NIR-spectra from the mixtures of teas of different origins were recorded in diffuse reflection mode with the powder probe as well as with an integrating sphere module.

In order to increase the prediction accuracy for a caffeine content of <1 per cent by weight, an additional 133 production samples of decaffeinated tea (0.04 to 0.4 per cent caffeine) were included in the model. Each sample was measured six times in order to account for the influence of different packing densities of the tea particles when inserting the probe or filling the cup for the integrating sphere module. From these six spectra an average spectrum was calculated for each sample.

The PLS1 algorithm for multivariate calibration in the Opus/Quant-2 software package was applied. The HPLC data obtained for all samples (PhytoLab GmbH &Co.KG) were denoted as the atrue values' in the calibration table. The operating differences between the powder probe and the integrating sphere module resulted in spectral differences; therefore, different frequency ranges for the determination of caffeine were selected (see Table 1).

The quality of the models was improved by preprocessing of the original spectra. The first derivative and multiplicative scatter correction (MSC) were calculated for the spectra taken with the powder probe, and vector-normalisation was performed on data recorded with the integrating sphere module. Optimised cross-validation results based on the aleave-one-out' method from NIR-spectra from the same 145 calibration standards are compared for both models.

One can see in Table 2 that the correlation coefficient R2 between true and predicted caffeine content and the root mean square error of cross-validation (RMSECV) are significantly improved when the integrating sphere module is used.

After developing the method with iterative cross-validations, we performed the final calibration step, which normally produces better critical values for the correlation coefficient R2 and the root mean square error of estimation (RMSEE) as shown in Table 2.

Since the calibration model obtained with the data from the integrating sphere module had the better prediction accuracy, it was used to analyse unknown production batches of decaffeinated tea samples, (Table 3).

Two categories of decaffeinated teas are of practical importance: extraction of caffeine to reach specified limits of <0.4wt. per centage (category I) or <0.1wt. per centage (category II). Therefore, we examined 13 production batches for each category as test sample sets. These sets were measured with the integrating sphere module using the set up as described above.


From the results we conclude that there is good conformity between NIRS and the standard HPLC measurements for production samples of decaffeinated tea in category I (ca. 0.4wt. per centage). For category II (<0.1wt. per centage) we find that the relative deviation between HPLC and NIRS are somewhat larger, but in each case the NIRS results are accurate enough to correctly determine that the tea sample meets the specifications for this category.

Consequently, NIRS provides an easy and especially rapid method for checking the low caffeine specifications of decaffeinated tea samples for at-line production control. Moreover, NIR spectroscopy can be used for several other applications, eg the determination of moisture content in the original coarse tea.

Angela Schmidt is with Bruker Optik GmbH, Karlsruhe, Germany, Brian P Davies and Dr Grit Schulzki are with PhytoLab GmbH & Co KG, Vestenbergsgreuth, Germany and Dr Andrea Giehl , Bruker UK Ltd, Coventry.



Twitter Icon © Setform Limited