With the growing need for precision agriculture under Agriculture 5.0, reliable early detection of fruit spoilage has become crucial. Traditional visible-light imaging and manual inspection are limited in detecting internal decay and are highly sensitive to illumination changes. This study presents a multispectral deep-learning pipeline (510–920 nm, 16 bands) using MobileNetV2 for guava health classification. Experiments on small (83 images) and large (2000 images) datasets achieved 100% and 99% accuracy, confirming strong model robustness. To interpret decision behavior, Grad-CAM was employed to visualize spectral attention and analyze error patterns. Results show that the model consistently focuses on short-wavelength regions for rotten samples, whereas misclassifications stem from over-sensitivity to local curve fluctuations and high-contrast boundaries. These findings validate the feasibility of applying CNNs to curve-based multispectral data while emphasizing that visual explanations in non-spatial spectral domains require cautious interpretation.