Crop illness is life-threatening for agricultural productivity and global food availability. Detecting these diseases using traditional methods is frequently time-consuming, costly, and prone to human error. This paper proposes an advanced deep learning framework for accurate and interpretable plant leaf disease detection, specifically focusing on widely grown crops in Bangladesh. Our contributions include developing an ensemble deep learning model that achieves 97% accuracy across 21 distinct classes. We addressed the multiclass imbalance problem and handled diverse datasets through extensive data augmentation techniques. Our method uses five pre-trained deep learning models-EfficientNet-B3, DenseNet-121, MobileNet-V2, ResNet-50, and Inception-V3-to learn how to recognize leaf diseases. After training each model, we picked the three that performed best and combined their predictions with a softvoting ensemble, which boosted the overall accuracy. To make the system’s decisions understandable, we added an Explainable AI component using LIME (Local Interpretable Model-Agnostic Explanations), which highlights the parts of the leaf image that most influence each prediction, giving users clear insight into how the model reaches its conclusions. The proposed system is demonstrated through a React frontend web application, allowing users to upload leaf images from the mendeley leaf disease dataset for disease detection. The model was developed and tested using a custom dataset containing 12,786 images of both diseased and healthy leaves from six commonly grown vegetables in Bangladesh, including Bitter Gourd, Bottle Gourd, Tomatoes, Eggplants, Cauliflowers, and Cucumbers, with a 70% training split, 20% for testing, and 10% for validation.