Early disease detection is a challenging and important task for yield and quality increase. Guava, being a popular and widely cultivated fruit in India, requires early and accurate disease detection. Deep learning techniques have shown its effectiveness in this domain. This paper has developed a deep learning model by fine-tuning EfficientNetB3 to improve classification accuracy. The work further used explainable AI technique Grad-CAM to visually explain the model’s decision-making workflow by ploting the relevant parts of the input images. It helps in empowering data-driven decision-making in disease management. The developed model achieved significant accuracy in distinguishing between diseased and healthy guavas when built and validated on an open-source dataset from Mendeley. This work also incorporates methods such as data enrichment, early stopping ensures rigorous training, preventing overfitting. The fusion of EfficientNet-B3 and XAI not only enhances classification performance but also builds trust and understanding for stakeholders. The system’s effectiveness is evaluated through multiple indicator and exhibited superior outcomes in key classification metrics including the F1-score are 94%, 93%, and 94%. Correspondingly on test datasets. The proposed model achieved training and testing accuracy of 96.67% and 94%, respectively, while XAI methods enhanced transparency, facilitating trust and understanding for users. This work holds promise for real-world agricultural applications, enabling timely interventions and improved disease management in guava cultivation.