Evaluación de bandas espectrales en la clasificación de tipos de defectos de la manzana mediante aprendizaje profundo.

Abstract

This study addresses the early classification of defects in apples, a key factor in reducing economic losses and improving quality throughout the supply chain. It proposes an approach based on multispectral images combining four spectral bands: red (RE), blue (BL), green (GR), and infrared (IR), overcoming the limitations of previous studies focused solely on the visible spectrum. The methodology employs convolutional neural networks (DenseNet121, MobileNetV1, and MobileNetV2) with fine-tuning, data augmentation, and channel-wise attention (SE) to automatically highlight the most relevant spectral features. Additionally, a band ablation analysis was conducted to evaluate the contribution of each band to the detection of three types of defects (rot, bruises, and climate-stress spots) as well as to the anatomical classes of the fruit (stem and calyx) and healthy apples. Results showed that MobileNetV2 + SE achieved the best overall performance, with an accuracy of 95% and precision, recall, and F1-score values above 91 %. MobileNetV1 + SE reached 94 %, while DenseNet121 + SE achieved 89 %, significantly improving over the version without attention (84 %). Ablation analysis revealed that the green band (GR) was the most informative, providing greater contrast between healthy and affected areas. Furthermore, an interactive platform was developed, allowing users to upload multispectral apple images and automatically receive a report with the defect classification and the most relevant spectral band. Keywords: Apple defects, Multispectral imaging, Convolutional neural networks, Attention modules, Transfer learning, Spectral bands.