In this work, we explored the utility of deep learning models (CNN) in detecting diseases on banana leaves which includes following groups Sigatoka, Cordana and Pestalotiopsis. The CNN model was able to reach up to 85.12% of accuracy, that making it capable enough for a recognition tasks about the infected materials exploiting complex patterns on large images dataset. The results indicate that state-of-the-art techniques might permit improved solutions to then classical methods for plant health and productivity. CNN Model performs better than other complex architectural models which is capturing the important features of images. Validation of the ability to detect diseases through metrics such as prevalence, recall/F1-score and Confusion Matrix. The use of Grad-CAM is also an added bonus, it provides a visual explanation on the image as to which part contributes toward classifying it under diseased category and thus makes our understanding more simpler towards the model development process. In conclusion, this study improves the current state of technology for plant disease detection and aids to food health. For further research, we may consider adopting hybrid models to enhance the accuracy and effectiveness of plant diseases detection which can be done using larger datasets in subsequent studies.