Interpretable Deep Learning for Banana Leaf Disease Detection Using Dense Convolutional Models

Banana cultivation often suffers from losses due to leaf diseases with similar visual symptoms, making timely and accurate classification difficult. This study applied a deep learning-based solution using the DenseNet201 architecture to classify four banana leaf conditions: Cordana, Sigatoka, Pestalotiopsis, and healthy. A total of 2,538 images were used, divided into augmented sets for training and validation, and a separate set of unseen original images for testing. Two model variants were developed: one using Global Average Pooling (GAP) and another using feature flattening. The GAP-based model achieved a test accuracy of 96.05%, while the version without GAP reached 97.23%. Despite the slight difference in accuracy, the GAP model demonstrated improved interpretability. Gradient-weighted Class Activation Mapping visualizations were used to observe class activation regions and understand the model’s focus during prediction. These explanations revealed that the GAP model attended to more relevant lesion areas, especially in cases where visual similarity between disease types caused confusion. The analysis of misclassifications and confusion matrices supported this observation. The results showed that integrating GAP helps the model generalize better and produce more transparent decisions. This approach strengthens the trustworthiness of deep learning models for agricultural disease diagnosis, particularly when applied to real-world, visually complex data.