Postharvest oranges are susceptible to decay, causing significant economic losses, while manual inspection remains subjective. This study proposes a nondestructive orange decay identification method combining multispectral reflectance curves with a lightweight MobileNetV2-based CNN. ROI reflectance curves (510–920 nm) from 50 orange samples were acquired via a Solomon™ multispectral camera and converted into standardized images for binary classification (healthy vs. rotten). The model achieved 96.6% accuracy, 99.5% healthy-sample Recall, and 99.4% rotten-sample Precision. Main errors involved 12 false-negatives caused by ROI placement near lesion boundaries. Grad-CAM analysis confirmed the model focused on the 510–650 nm region, learning spectrally meaningful features. These results demonstrate the feasibility of this approach for handheld inspection and smart-agriculture grading, though preprocessing inconsistencies between sets introduced a slight input-distribution shift. However, preprocessing inconsistency between the training and testing sets may have introduced a slight input-distribution shift.