Detection of subsurface bruises in plums using spectral imaging and deep learning with wavelength selection

Plums are widely consumed, both fresh and processed. During harvest, handling, or transportation, they are exposed to static and dynamic compression forces exceeding the critical stress for tissue damage. Compressionrelated damage typically develops further as the fruit ripens and softens, facilitating the occurrence of rot, which might cause significant losses in the supply chain. Early detection of these damages is crucial to sorting the damaged fruit out and deviating it to processing, thus preventing food waste. However, early-stage bruises or damages on plums are not visible, especially not in dark-skin cultivars. Therefore, this study aimed to explore the potential of hyperspectral imaging in the 430 to 1 000 nm range and deep learning algorithms to detect these invisible bruises at an early stage. To this end, 'Presenta' plums were impacted at three different levels to simulate varying degrees of damage. Images of both bruised and non-bruised plums were taken immediately after bruising and 24 and 48 h after bruise induction. Three distinct CNNs were trained to analyze the images. Two of these networks were implemented using transfer learning (ResNet and HSCNN), while the third was customdesigned for this specific purpose. The most informative wavelengths were identified as inputs for the CNNs employing PCA. F1 scores over 81% were obtained in all cases, and almost 100% accuracy was obtained in classifying the bruised plums with the highest impact energy of 0.50 J. Thus, detecting and classifying bruised plums using only three wavelengths is possible, paving the way for in-line sorting with multispectral cameras in packing houses.