This study proposes an interpretable deep learning framework for the automatic detection of dragon fruit stem diseases using the DenseNet121 CNN architecture. Dragon fruit, a tropical crop of increasing economic importance, is vulnerable to a variety of stem diseases that can reduce yield as well as quality. To enable reliable and explainable disease classification, the proposed model integrates Gradient-weighted Class Activation Mapping (Grad-CAM), which generates heatmaps to visualize the regions of the stem images that influence model predictions. The system was trained and tested on a publicly available dataset curated from Roboflow, containing labeled images of various stem disease classes. Experimental results show that the proposed model achieves strong classification performance, as reflected in metrics like accuracy where the model obtained a test accuracy of 93.71% and test loss of 0.1108 on the held-out test dataset, other evaluations metrics such as F1-score, and confusion matrices are also presented. By integrating Grad-CAM, the model becomes more transparent, offering meaningful visual explanations that can help agronomists and farmers understand and trust its predictions. This work demonstrates how interpretable AI can play a valuable role in agricultural disease management and support the growing field of precision farming.