Tsega Y. Melesse
Abstract The assessment of postharvest fruit quality under non-invasive and accurate methodologies plays a significant role in enhancing sorting efficiency and loss minimization under precision agriculture. In this study, a deep learning framework is introduced in which a light weight CNN (LW-CNN) and thermal imaging are combined for banana classification under four distinct quality groups: fresh, ripe, overripe, and rotten. A specific set of 4336 thermal images was captured using a FLIR ONE Gen 3 infrared camera operating under thermal-only mode under controlled ambient setups. The set was subjected to balancing and augmentation procedures for better generalization capabilities, and the CNN was trained under such inputs for the identification of thermal signatures corresponding to ripeness. The proposed model demonstrated a high level of accuracy and robust performance in varied measures, marking its ability to distinguish effectively across all categories of quality. Testing under a confusion matrix and a precision-recall curve also supported the effective performance of the classifier under differing confidence thresholds. These results encourage the combination of deep learning and thermal imaging as a feasible, economically viable, and non-invasive real-time postharvest quality evaluation methodology. The proposed methodology forms the basis for smart sorting infrastructure and decision-support systems under data-informed sustainable postharvest management.