Original papersThe use of hyperspectral imaging to predict the distribution of internal constituents and to classify edible fennel heads based on the harvest time

- Prediction of internal constituents of fennels by NIR spectroscopy was attempted.
- Better prediction models were achieved using Vis-NIR than NIR spectral region.
- SSC, DPPH and phenolics were most satisfactorily predicted.
- Internal constituent concentration was mapped on hyperspectral images.
- A high non-error rate was obtained for classification based on harvest time.

The objective of this study was to use hyperspectral imaging to predict the internal concentration of soluble solids (SSC), individual sugars and organic acids, phenols, and antioxidant activity of fennel heads in relation to different sheath layers and harvest times. Thirty-five fennel heads were collected during 7 different harvest times over a period of 3 weeks. For each fennel VisNIR (400-1000 nm) and NIR (900-1700 nm) images of the perpendicular section were acquired. From the external to the internal part of the fennel chemical analysis of each leaf was done summing up to 160 samples. Similarly, for hyperspectral images three regions of interest (ROI) were extracted and averaged for each corresponding leaf. A calibration set of 105 samples and a validation set of 31 samples was used to develop the PLSR models, after removing 20 samples without correct reference values and 4 outlier spectra. Among the predicted parameters only SSC, DPPH and phenols could be predicted with satisfactory accuracy. Particularly, for SSC, mean centering gave an R2 of 0.87, 0.81, 0.77 for calibration, cross validation, and prediction, respectively (RMSEP of 0.515 over a range of values from 4 to 9%). First derivative combined with SNV applied for DPPH gave the same accuracy with R2 of 0.81, 0.76, 0.78 (RMSEP of 2.460 over a range of 20-250 mg kg−1). The best preprocessing technique for phenols was MSC (mean) yielding RMSEP of 3.042 (over a range from 50 to 350 mg kg−1). In addition it was possible to map the constituent concentrations on the hyperspectral images showing the increase of soluble solids, phenolics and antioxidant activity from the external to the internal leaves. As for classification of fennels according to harvest time using PLS-DA, all the classes were distinguished with a non-error rate of 89.29% in calibration 75.71% in cross validation and 88.57% in prediction. Except for some samples of class 5 in calibration and 2, 4 and 5 in case of cross validation, all others were nearly correctly classified. In conclusion results of this work showed the potentiality of hyperspectral imaging in the Vis-NIR spectral range to predict internal constituents and to classify fennel heads according to the harvest time.