NIR hyperspectral imaging to identify damage caused by Halyomorpha halys on pears: Automated identification of Regions of Interest related to punctured areas

Halyomorpha halys, commonly known as the Brown Marmorated Stink Bug (BMSB), is an emerging pest in pear orchards determining major economic losses. BMSB feeding on fruits close to harvest ripening cause internal damage invisible to the naked eye, therefore undetectable using RGB image acquisition systems. To face this issue, in the present work Near-Infrared Hyperspectral Imaging (NIR-HSI) is proposed as a non-destructive technique to automatically discard damaged fruits in post-harvest sorting lines. In this context, the identification of Regions of Interest (ROIs) ascribable to the punctures is a crucial step affecting the outcomes of supervised classification models. Due to irregular shapes and blurred edges between sound and punctured areas, most popular thresholding techniques are not able to automatically detect the ROIs while, on the other hand, manual thresholding is arbitrary and time consuming on large hyperspectral image datasets. This paper provides an innovative method for the automated ROIs selection based on image data dimensionality reduction (DDR) and image-level classification coupled with spatial feature selection. To this aim, the hyperspectral images were compressed into Common Space Hyperspectrograms (CSH), signals summarising both spatial and spectral information of the original images. The CSH features highly correlated with the presence of BMSB punctures and more frequently selected by interval Partial Least Squares – Discriminant Analysis (iPLS-DA) models allowed the identification of ROIs of punctured areas. Indeed, the reconstruction of the selected features back into the original image domain led to a successful identification of ROIs ascribable to BMSB punctures in an automated and objective way.