An explainable hybrid feature aggregation network with residual inception positional encoding attention and EfficientNet for cassava leaf disease classification

Cassava is a tuberous edible plant native to the American tropics and is essential for its versatile applications including cassava flour, bread, tapioca, and laundry starch. Cassava leaf diseases reduce crop yields, elevate production costs, and disrupt market stability. This places significant burdens on farmers and economies while highlighting the need for effective management strategies. Traditional methods of manual disease diagnosis are costly, labor-intensive, and time-consuming. This research aims to address the challenge of accurate disease classification by overcoming the limitations of existing methods, which encounter difficulties with the complexity and variability of leaf disease symptoms. To the best of our knowledge, this is the first study to propose a novel dual-track feature aggregation architecture that integrates the Residual Inception Positional Encoding Attention (RIPEA) Network with EfficientNet for the classification of cassava leaf diseases. The proposed model employs a dual-track feature aggregation architecture which integrates the RIPEA Network with EfficientNet. The RIPEA track extracts significant features by leveraging residual connections for preserving gradients and uses multi-scale feature fusion for combining fine-grained details with broader patterns. It also incorporates Coordinate and Mixed Attention mechanisms which focus on cross-channel and long-range dependencies. The extracted features from both tracks are aggregated for classification. Furthermore, it incorporates an image augmentation method and a cosine decay learning rate schedule to improve model training. This improves the ability of the model to accurately differentiate between Cassava Bacterial Blight (CBB), Brown Streak Disease (CBSD), Green Mottle (CGM), Mosaic Disease (CMD), and healthy leaves, addressing both local textures and global structures. Additionally, to enhance the interpretability of the model, we apply Grad-CAM to provide visual explanations for the model's decision-making process, helping to understand which regions of the leaf images contribute to the classification results. The proposed network achieved a classification accuracy of 93.06%.