Hazelnuts classification by hyperspectral imaging coupled with variable selection methods

The increasing normative requirements and market competitiveness lead the agricultural sector and the food industry to constantly look for new fast and non-destructive classification logics that can be applied for product sorting applications and/or quality control actions. With reference to hazelnut production, the dried fruits must be sorted from unwanted foreign bodies or inedible hazelnuts that can negatively affect the quality of the final product. In this perspective, the utilization of HyperSpectral Imaging (HSI) can be applied to set-up a novel hazelnuts quality control. Hazelnuts and contaminants were acquired by a push-broom hyperspectral device working in the Short-Wave InfraRed (SWIR: 1000-2500 nm) region. A PLSDA model was set up in order to identify 3 classes of products (i.e. edible hazelnuts, hazelnut shells and rotten hazelnuts) with the highest level of efficiency in full spectrum mode (Precision = 0.92, Accuracy = 0.94, Efficiency = 0.94). Subsequently, different variable selection methods (i.e. Interval PLSDA, Selectivity Ratio and Variable Importance in Projection score methods) were adopted in order to identify the fundamental bands to recognize the 3 classes and evaluate which of the variable selection methods shows efficiency values close to the values obtained by the full spectrum mode. VIP score-based classification showed the best performance, with Precision, Accuracy and Efficiency values equal to those based on full spectrum PLSDA. Classification results suggest that this methodological approach can be powerful to develop and implement hazelnut sorting and/or quality control strategies. Moreover, the variable selection approach allows to increase processing speed , compared to that in full spectrum mode, making possible online applications directly at plant scale.