The prevention of the sugarcane leaf diseases is a significant milestone in maintaining the productivity of the sugarcane crop as well as in enhancing the practice of smart farming. The paper presents a solution to the problem, which relies on deep learning algorithms on a properly organized and labeled test dataset of five categories of datasets: Healthy, Red Rot, Rust, Mosaic and Yellow leaf disease. In order to enhance the classification robustness, an ensemble model, using DenseNet121, EfficientNetB0, and ResNet50 is created. Besides, the results of single standalone ResNeXt-101 models have been tested on their own and the highest classification rate is 98 %. To increase interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) is being used to create heatmaps which produce class-specific features that identify important areas on the visual, used to make a prediction. The models were extensively evaluated on the evaluation metrics of significance. Results show good predictive performance in both approaches, where ResNeXt-101 performs better than all the other methods in terms of ability to generalize and reliability. The results highlight the efficiency in ResNeXt-101 as an independent diagnostic technique for an early detection of disease in sugarcane leaves. Furthermore, integrating explainable AI helps to boost transparency and trust around model predictions. This research introduces a scalable and interpretable method to classification of plant diseases with a lot of potential to be implemented in real-time for precision agriculture and automatic monitoring devices in the field.