The use of hyperspectral imaging to predict quality and characterize local varieties of artichoke

ABSTRACT Nowadays it is increasing the needs to know the quality and safety of the food products. These requirements call for on-line detection techniques which have the advantages of be assembled in the production line and take place under realistic environment, know early detection of possible failures, have permanent monitoring of the conditions and know assessment of conditions at any desired time. This study evaluated the feasibility of using a spectral scanner VIS-NIR (DV Srl, version 1.4., Italia) with a detector in the region between 400-1000 nm to predict quality and characterize local varieties of artichoke: “Violetto” and “Catanese” located respectively in the area of San Ferdinando di Puglia and Brindisi (Puglia Region, Italy). The samples were harvested during years 2009/10 from 20 plants for each field, randomly-chosen and labelled in order to reduce field variability among different harvest dates. Artichoke heads were harvested from December to May (7 harvest dates) for “Violetto foggiano” and from January to April for “Catanese” (4 harvest dates). Artichokes were processed and cut into quarters. One quarter for each artichoke was analyzed during storage at day 0, day 2, day 5 and day 7 acquiring hyperspectral images using a hyperspectral imaging system. Spectral data were analyzed using the Unscrambler packing software version X (CAMO ASA, Oslo, Norway) and PLS toolbox in Matlab (version 2014a). The data set included 736 samples (400 for “Violetto” and 336 for “Catanese”). All the reflectance measurements were firstly transformed to absorbance values using log(1/R) according to the law of Lambert-Beer. Classification models were built with the aim of discriminate among cultivars, harvest times and day after cut. Two methods were compared: SIMCA (Soft Independent Modelling of Class Analogy) and PLS-DA (Partial least squares discriminant analysis), defining a Training set of 308 samples for “Violetto” and 244 for “Catanese” and a Test set of 92 samples for “Violetto” and 76 for “Catanese”. In the classification by cultivar (“Violetto” and “Catanese”) the discriminant approach is superior to the classmodeling, mostly because of the two classes have a very similar general profile of the spectrum and one of them (“Violetto”) have an inner variability which encloses the one of the other class (“Catanese”). Forcing the discrimination, the differences between the two classes are exalted, and the classification is obtained with very interesting results. For the classification by harvest time, the SIMCA model was developed building individual PCA models for the spectra of each harvest time. Comparing the result coming from the analysis made with SIMCA and PLS model, it is evident how the PLS-DA is the most performing method for this application giving a “non error rate” of 80% on the external test set. For the classification by days of storage the PLS-DA model has for all the classes high value of specificity, and for some classes low values of sensibility. The results suggest that is possible to discriminate samples just cut from samples cut and stored for some days, but that is more difficult to exactly separate samples depending on the days of storage. Most likely this is not due to a low efficiency of the model but to the changing proprieties of the samples that are not so dissimilar between 2 and 7 days of storage, but becoming more evident with the passing of the time. For these analysis “Non Error Rate” values increased reducing the number of classed from 4 to 3: the model performance improved. Calibration model for phenols content and antioxidant activity was built analysing for day 0 several pretreatment (9 for antioxidant activity and 8 for phenols). Particularly the data of “Harvest Time 1” showed a different behavior compared to the remaining harvest times and for this reason the prediction models were tested on 3 classes: “All Harvest Time”, “Harvest Time 1” and “Other Harvest Time”. The efficiency of the model was always higher when using only sample from “Harvest Time 1”, suggesting that other sources of variation were included in the data set for the following samplings. The classes “All Harvest Times” and “Other Harvest Times”, for PLS-calibration model, had higher values of R2 C and R2 CV and low values of RMSEC and RMECV in the wavelength range of 400-1000 nm for both phenols content and antioxidant activity. “Harvest Time 1”, instead, carried out the best value for both (phenols content and antioxidant activity) in the range 650-1000nm (R2 pred 0.62 and RMSEP of 72 for phenols, and R2 pred 0.67, and RMSEP of 126 for antioxidant activity). Starting from this considerations and from obtained results it may be interesting to further investigate the effect of the harvest time on the phenolic and antioxidant activity prediction to try to improve prediction results. Moreover also the instrumental setting can be improved, trying to standardize as much as possible the acquisition conditions. Generally results of this thesis explored new area of research developing tools that may be used to increase the value of local productions, by mean of a better characterization and identification and by providing innovative non destructive-tools to be used online during the minimally processing operations for selecting raw material based also on its internal composition.