Classification of early mechanical damage over time in pears based on hyperspectral imaging and transfer learning

Abstract

Mechanical damage of fresh fruit caused by compression and collision during harvesting and transportation is an urgent problem in the agricultural industry. The purpose of this work was to detect early mechanical damage of pears using hyperspectral imaging technology and advanced modeling techniques of transfer learning and convolutional neural networks. The visible/near-infrared hyperspectral imaging system was applied to obtain the intact and damaged pears at three time points (2, 12, and 24 h) after compression or collision damage. After the hyperspectral images were preprocessed and feature-extracted, ImageNet was used to pre-train ConvNeXt network, and then, transfer learning strategy was applied from compression damage to collision damage to build the T_ConvNeXt model for classification. The results showed that the test set accuracy of fine-tuned ConvNeXt model was 96.88% for compression damage time. For the classification of collision damage time, the test set accuracy of T_ConvNeXt network reached 96.61% and was 3.64% higher than the fine-tuned ConvNeXt network. The number of training samples was proportionally reduced to verify the superiority of the T_ConvNeXt model, and the model was compared with conventional machine learning algorithms. This study achieved the classification of mechanical damage over time and achieved a generalized model for different damage types. The accurate prediction of pear damage time is crucial for determining proper storage conditions and shelf-life time. PRACTICAL APPLICATION: The T_ConvNeXt model proposed in this paper transferred from compression damage to collision damage effectively promoted the generality of the damage time classification model. Guidelines for choosing an effective shelf life from a commercial aspect were presented.

Abstract from the paper’s OpenAlex record, shown for indexing and research. Copyright remains with the authors or the publisher. Read at the source ↗

The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.