<![CDATA[Cassava (Manihot esculenta) is a crucial commodity in tropical regions; yet, its output has markedly declined due to five primary categories of foliar diseases: Cassava Mosaic Disease (CMD), Cassava Brown Streak Disease (CBSD), Cassava Bacterial Blight (CBB), Cassava Green Mottle (CGM), and the healthy category. The constraints of deep learning models, which remain opaque, and data imbalances in the agriculture sector, pose significant hurdles to the development of precise and transparent diagnostic tools. This research seeks to establish an explainable deep learning framework utilizing ensemble transfer learning and adaptive augmentation to enhance the accuracy and interpretability of cassava leaf disease identification. The experimental investigation utilized 21,367 annotated photos from five disease categories within the Cassava Leaf Disease dataset. The dataset was first divided into training data (80%) and validation data (20%). The minority classes in the training data were augmented using Albumentations to rectify class distribution imbalances. As a result of this approach, an additional 51,280 photos were generated, resulting in a more balanced and representative class. The findings indicated that the ensemble averaging of ResNet50, DenseNet121, and EfficientNet-B0 achieved a validation accuracy of 86.75%, surpassing the performance of each individual model. Adaptive augmentation procedures enhance the model's generalization capabilities, while Gradient-weighted Class Activation Mapping (Grad-CAM) visualizes the leaf regions influencing classification judgments. The results validate that the use of adaptive augmentation, explainable AI, and ensemble transfer learning enhances the transparency and reliability of computer vision systems for detecting plant diseases. This study advances the creation of precise, interpretable, and pertinent AI models to enhance agricultural informatics.]]>
