• Keine Ergebnisse gefunden

Farmers and traders have traditionally made the first flax quality assessment by subjective judgment. The qualitative methods for fiber quality determination are slow and time consuming. New technologies for testing fiber quality faster and/or non destructively would be of great value to allow the rapid evaluation of flax samples [38].

4.3.1 Fiber quality determination at the ITV-Denkendorf

Some authors affirm that due to the irregularity of the flax fiber structure, the fiber fineness determination is operator dependent and microscopic methods should not be used. More adequate methods are those based on air permeability and are [71]:

For randomly oriented fibers the difference of pressure over 1.2 g of flax (compressed in a measuring chamber) of a certain air flow is measured. This method is suitable for fibers in the form of flocks.

In the reference method: the pressure drop over a bundle of parallel fibers is mea-sured. This method is suitable for fibers in the form of slivers.

The fiber fineness reported throughout this thesis was determined by analysis of mi-croscopy images at the ITV-Denkendorf (Section 2.17).

4.3.2 Fiber quality determination with IR methods

There are many examples in the literature where infrared-multivariate calibration methods were used successful in predicting properties of plants or other natural products [18, 19, 22, 54, 57, 60, 64, 106, 110]. It is important that the samples used for the calibration are representative and accurately measured. The quality of the calibration obtained depends entirely on the accuracy of the reference data (see Section 4.3.1).

A minimum of six spectra per regression factor is recommended for building an IR-multivariate calibration according to the ASTM guidelines. For seven factors, as used in the NIR model, a minimum of 42 samples is required. And for the ten factors used in the FTIR calibration, at least 70 samples are necessary. Both calibrations produced in this work, complied with this criterion. An increase in the number of factors used for the PLS analysis, will always decrease the error produced by the model in the calibration set. This presents the risk that unwanted variability in the data set, such as random noise, may also be considered by the model. The model is over-fitted, and shows excellent results for evaluating samples belonging to the calibration set but failing to predict samples from an external validation set [91]. Therefore the minimum necessary number of factors must be used.

The calibration models generated with the NIR-data (Section 3.3.2), were much better than those obtained using the FTIR-data. However, when tested against an independent validation data set; they proved to be not so robust. Causes that could have affected the IR analysis are described below.

Fiber heterogeneity: the majority of flax samples produced during the laboratory scale and pilot plant scale experiments were line flax (long flax fibers). They were sent to the ITV-Denkendorf for analysis where different parts of the sample were measured.

As mentioned in Section 1.4, the fiber structure varies at different parts of the stem.

It is therefore highly probable that the same sample although correctly measured by the ITV-Denkendorf and by IR spectroscopy, was too heterogeneous. This impeded to get a good correlation between both measurements.

Sampling window size: in the FTIR-ATR device, the fiber sample is pressed against the diamond window (∅1.8 mm). In a report from the literature, the quality of flax fiber determined by the airflow method was correlated with the reflectance Vis-NIRS data (400-2498 nm or 25 000-4000 cm−1). The sample device for scanning natural products of 21×5×4 cm; allowed to sample 20 g of fibers cut in 20 cm length.

A PLS regression method was used to establish relationships between fiber fineness and results from Vis-NIR spectra. This method produced a model of R2 of 0.97 and a standard error for calibration (SEC) of 1.69 dtex. The work was based on an equation developed with a population of 462 dew retted and water retted samples

with a fineness range of 18.3-65.4 dtex, analyzed in the range from 1100 to 2500 nm (or 25 000 to 4000 cm−1) [57, 100]. Additionally a model with thermogravimetric data was built, but the model was not so robust, because a larger data set of 100-250 of fiber samples was be necessary to generate a stable model [57].

The NIR sampling window diameter used in this thesis was 13 mm. In other work, NIR reflectance spectra of of a variety of cereal food products were acquired with a commercial dual diode-array (Si, InGaAs) spectrometer, which was able to measure a large area of the specimen surface of about 10 cm diameter. Interference from color, hydration effects, or irrelevant chemical variations was large and was not evenly distributed across the spectrum. It was found that selecting subsets of wavelength variables substantially improved the calibration performance. The calibration model obtained was able to determine the total dietary fiber accurately [19].

Therefore the scanning area needs to be larger to counteract the sample heterogene-ity and to allow more information to be collected.

Detection range: the FTIR range and the NIR range were used in this work. Other works have used the visible NIR range and obtained good calibrations for flax fiber quality [57].

Data: the samples required to produce a robust model must include the natural variabil-ity of property of interest (e.g. fiber qualvariabil-ity) whereas a number of samples uniformly distributed between the extremes values can give better results. The minimum range that a property or concentration value must vary in order to provide enough infor-mation for the calibration software has been recommended as ± 5 times (and not less than ± 3 times) the reproducibility of the reference method [91].

A larger data set of samples of fiber would be necessary to generate a stable model.

Besides this, the error of the reference method done at ITV-Denkendorf, should have been quantified and replicate sampling would have been needed. This was not possible, due to the long time required for the fiber analysis.

Regression type: A PLS regression presupposes a linear relationship between the spec-tral data and the concentration or other property value to be determined. If non-linearity (not easily accommodated by PCR and PLS) between the spectral data and the quantitative information of interest exists, “artificial neural networks” (ANN) have proved to be useful [91].

The models produced in this work, although not so accurate, can give a general idea of the sample quality in a fast way. But in order to use them for predictions, they need to be improved.

4.3.3 Fiber quality determination by measuring the uronic acid content

Meijer et al. found no clear relationship between uronic acid and fiber quality in flax retting experiments. Moreover, they found that water-retted fibers isolated from the stems having the lowest pectin content were the strongest and the high pectin samples of green flax fibers were the weakest [79]. In Section 3.1.2 the uronic acid content (pectin content) was compared to the fiber resolution and fineness, but there was no clear relationship between those values.

4.4 Isolation of the strains appearing during the flax fiber treatment

The bacterial culture of G. thermoglucosidasius PB94A was overgrown by other type of bacteria, when the culture was incubated with flax fibers. These new bacteria proba-bly originated from the flax fibers. The DGGE analysis of those broths confirmed the appearance of new strains, while the original pattern of G. thermoglucosidasius PB94A almost disappeared (see Sections 3.2.3 and 3.4). The strains were isolated selectively in pectin media. But no new strains were found in the DGGE analysis, only strain G.

thermoglucosidasius PB94A was detected in the experiments, some probable explanations are:

• A selection factor either in the sampling, or the PCR, or the DGGE was introduced, therefore isolating always G. thermoglucosidasius PB94A. This will be discussed in depth in Section 4.5.

• The other “new” bacteria were simply not cultivable with the employed method.

• The “new” bacteria were not degrading pectin, that is why they were not isolated in pectin media. The only pectin-degrading strain that could be detected wasG. ther-moglucosidasius PB94A, which was recurrently recovered from the broth of several experiments.

4.5 Analysis of the bacterial population by the DGGE