Integrating CO₂ laser photoacoustic spectroscopy with explainable CNNs for post-harvest quality assessment of Diospyros kaki.

Abstract

This study presents a non-invasive molecular spectroscopy approach for monitoring storage-related physiological changes in Diospyros kaki (Hachiya persimmon). CO₂ laser photoacoustic spectroscopy (CO₂LPAS) was applied to quantify ethylene, ethanol, and ammonia emissions from individual fruits over a six-week storage period. Early ethylene release triggered ripening processes, followed by ethanol accumulation indicative of low-oxygen metabolism and microbial activity. Progressive ammonia emissions reflected proteolytic activity and tissue degradation during late-stage spoilage. To correlate internal metabolic signatures with external phenotypic changes, visual inspection was combined with high-resolution RGB imaging and a convolutional neural network (CNN) classification into three physiological macro-states: Storage, Transition, and Spoilage. The CNN was trained using 5-fold stratified cross-validation with test-time augmentation, achieving a mean classification accuracy of 91.67 ± 3.73%. Gradient-weighted Class Activation Mapping (Grad-CAM) highlighted regions contributing to classification, showing a shift from specular highlights in early stages to tissue deformation and necrosis in later stages, confirming the link between molecular emissions and structural degradation. The combined use of CO₂LPAS-based molecular spectroscopy and image-based assessment provides a complementary framework for real-time, non-invasive evaluation of storage-related fruit quality. These results demonstrate that ethylene, ethanol, and ammonia serve as robust molecular markers of ripening, anaerobic metabolism, and spoilage, while complementary imaging captures macroscopic deterioration. This dual-method framework offers potential for predictive postharvest monitoring, quality control, and shelf-life assessment of climacteric fruits.