HUKI
l Onel *1 oi / Talonie (2015 J S4-20
300 350 400 450 500 550 600
Emission [nm]
Fig. 2. («*) Projccnons of ihf inacf ramaln pjMPł iłur w*rr d*vnb*d by ihe EEMs in «h* spać* spjnn*d by PC 1 and PC Z Th* umpits from Southern Europ* jr* markrd with and the samples from PoLind ar* rrarked wuh •' and (b) a rolor imag* nf PC 1 rrfolded lojdmgs (For mrrrprclalior nf rh* r*f*rrnr*s to rolor m this figur*, th* v*ad#r U r*f*rr*d to thr w*b v*r%ion of tltis arilrl*.)
of che imago of the PC 1 loadings revealed the activity of fluorophores in the rangę of 350-500 nm with an excitation of around 360 nm. These fluorophores may be cafTek and ferulic acids. High levels of those fluorophores were found in the samples that originated from Southern Europę. This observation is in the agreement with the observations that tomato products that were cultivated under influence of increased UV radiation had 20% higher concentrations of caffeic and ferulic acids |35J. Differences in UV radiation conditions are also common for the samples in our study (i.e. tomato products from southem Europę compared with those from Poland) |36|. The PC analysis of the IR spectra did not reveal any interesting data trends.
The IR speara hołd information about the Chemical content of samples. Depending on the band that is analyzed. the water. amino acids or phenolic content can be evaluated. For the study of the total antioxidant capacity. the spectral region between 2000 and 900cm'1 is recommended in the literaturę 112.37j. This spearal rangę includes the stretch of phenyl groups and an asymmetrk deformation of methyl and carbonyl groups that are prereąuisite to the Chemical structure of antioxidants. That is why, in our study. the PLSR models for TAC and TPC were construaed using this particular spectral rangę. Moreover, these models were compared to the models that were obtained for the samples that were described by their entire spectra (see Fig. 3).
Fig. 3. Mpjn IR spectrum nf thf inua tomjio sjmpl**. Th* spectral region berwetn 2000 and 900 cm ft Is Indicated with a dashed Uo*.
in order to include all of the samples with unique characteristics into the model set. the Kennard and Stone algorithm [38] was used. Evaluation of the uniqueness of samples that were analyzed was performed on the basis of mean spectra of the replicates. Thus. the model set contained 70% of available samples (cach described by three replicates). while the remaining 30% of samples formed the test set The same routine was applied when modcling the TAC as a function of ether EEMs (of the intaa samples and the extracts of tomato pastę) or the IR spectra (in fuli and rcduced spectral ranges). The spectra of the same samples were selected for the model set in order to perform the comparison.
Both PLSR and N-PLS models were constructed for the difFerent sets of spectra: 1) EEMs registered for the extracts from tomato pastes. 2) EEMs registered for the intact tomato pastes using a fiber optics probe, 3) IR spectra of the intaa samples that were collected in the rangę of 4000-400 cm \ and 4) IR spectra of the intaa samples reduced to the spectral rangę of 2000-900 cm The complexity of the models that were construaed (evaluated using the leave-one cross validation procedurę [29|) and the figures of merit that were obtained (RMSE. RMSEP and D2) for TAC and TPC. respcaively. are listed in Table 3
The analysis of the figures of merit that are listed led to the conclusion that the PLSR models that were obtained from the tomato pastę samples in the reduced spectral rangę of IR (2000-900 cm ‘ 1) are similar (or better) to the models that were built for the model set that was described with the entire IR spectra.
The best result was obtained for the model of the ORAĆ parameter as a function of the EEMs (fit and prediction of the model were equal to 5.87% and 852%. respeahrely) for the extraas. Higher errors were obtained for the TAC prediction that was evaluated from the spectra of the intaa samples than from speara of extracts. The N-PLS model that was built to predia the TAC value from the EEMs had much better prediaion properties than any model built using the IR spectra (in the entire or the shortened rangę). In contr as t, the PLSR models that were built to predia the TPC from IR had Iower errors than both the PLSR and N-PLS that were built using the EEMs. This was also confirmed by the improved performance of the models that were construaed for the IR speara of the intaa samples. These IR models were characterized by two-fold larger values of fit and prediction than the models that were built for the other type of spectra. In the Supplcmcntary matcnal seaion Tabtcs SI and S2 include observed and prediaed TAC and TPC contents from PLS and N-PLS models for the studied model and test set samples.