Non-invasive detection of internal defects in fruit by using visible-shortwave NIR spectroscopy
thesisposted on 06.12.2017, 00:00 by Bed KhatiwadaBed Khatiwada
Non-invasive detection of three internal disorders of fruit of commercial relevance to Queensland horticulture was considered: (i) diffuse browning of apple fruit; (ii) gelling defect of mandarin fruit; and (iii) translucency of pineapple fruit. Visible - short wave near infrared spectroscopy (vis-SWNIRS) is in commercial use for non-invasive field and in line assessment of fruit dry matter and soluble solids content of mango and apple. Some claims exist for commercially available instrumentation for sorting of fruit internal defects, but no assessment of such systems exists in the scientific literature. Four vis-SWNIRS instruments were trialled, varying in optical geometry: (i) the Integrated spectronics’s ‘Nirvana’ handheld instrument, operating with an interactance optical geometry; (ii) a purpose built unit employing a 300W halogen illumination source in a partial transmittance geometry, ‘IDD0’; (iii) the MAF Roda Insight2 unit, employing a 150W halogen lamp and operated in a full transmission geometry, and (iv) the MAF Roda IDD2 unit, employing four near infrared light emitting diodes and operated in a full transmission geometry. A number of reference methods were assessed for scoring level of apple flesh browning, including visual assessment, image analysis (% cross section area affected), chromameter CIE Lab values (L* and a* value) and juice Abs420nm, of which visual scoring on a 5 point scale was recommended. Chlorophyll fluorescence and acoustic resonant frequency was poorly related to extent of defect, and thus these non-invasive techniques are not recommended. Apple flesh browning was best assessed using visible-shortwave NIRS in a transmission optical geometry, with a typical PLSR model R2cv = 0.83 and RMSECV = 0.63 (5 point visual scale). Of different binary (good and defect fruit) classification approaches trialled, the best result was achieved using PLS discriminant analysis (PLS-DA) method, followed by linear discriminant analysis. More than 95% of defect fruit were predicted as defect (true negative rate) at the expense of having 10-20% of good fruit falsely predicted as defect fruit (false negative rate), across six populations. A number of reference methods were also assessed for scoring level of granulation in mandarin fruit, including visual assessment (5 point scale), chromameter CIE Lab (L and colour index values) and % juice recovery, of which visual score and % juice recovery were recommended. Mandarin granulation, indexed by either visual score or % juice recovery, was best non-invasively assessed using vis-SWNIRS in a transmission optical geometry, with a typical PLSR model R2cv = 0.74 and RMSEP = 3.6 (% juice recovery). PLS-DAwas able to predict well for good fruit with up to 87 and 100% of good fruit as good fruit (true positive rate) using the IDD0 and MAF Roda Insight2 units, respectively. Defect fruit were wrongly predicted as good fruit (false positive) using both the machine with best result (97 % true negative rate) obtained with PLS-DA using IDD0 unit. Translucency in pineapple, indexed by either a (5 point) visual score or image analysis was assessed using vis-SWNIR spectroscopy. Typical PLSR calibration results for models developed using the range 700-1000 nm results were modest (R2cv = 0.58, RMSECV = 0.55 on 5 point scale), and prediction results were poor (R2p = 0.41, RMSEP = 0.93. For binary classification, PLS-DA was able to predict 98.7% of good fruit as good (true positive rate) while only 34% of defect fruit were predicted defect (true negative rate) based on visual translucent score. Sorting involves a trade-off between yield and quality. The use of a receiver operating characteristics curve (ROC) and a sorting optimisation curve (SOC) was explored for the comparison of binary classifiers and the optimisation of sorting set point. The need to adjust the sorting set point to maintain a desired quality specification (e.g. % of defect fruit in accepted class) as population mean and spread (SD) for the defect varies is explained. Internal defects of fruit under consideration are well detected and sorted for based under transmission optical geometry with visual defect score as a reference parameter.