Optimizing Potato Leaf Disease Classification with Channel Dropout Attention in EfficientNetB0

Potato is one of the major crops and is considered the fourth most consumed food worldwide. Due to its sensitive nature, a variety of diseases harm potato production, with most disease syndromes appearing on the potato leaf. A careful analysis of potato leaf disease classification at an early stage can reduce damage and ensure healthy food for people. Existing Deep Learning techniques perform well but still have room for improvement due to less accuracy and transparency of the model. Enhancing the detection model and providing explainability could usher in a new era in this domain. In this work, we propose a modification to the EfficientNetB0 model by incorporating a Channel Dropout Attention layer into its main architecture. This layer selectively drops channels while assigning weights to focus on important features. To evaluate our proposed model, we used a potato leaf disease detection dataset comprising 2,152 images across three classes. We assessed the model’s performance using two approaches: a single train-validation-test split and 5-fold cross-validation. Proposed model achieved 99.08% accuracy with the single train-validation-test split and an average accuracy of 99.07% in 5-fold cross-validation. Additionally, we compared our proposed model with pre-trained CNN architectures such as AlexNet, LeNet5, and ShuffleNet. We also employed explainable AI techniques, including LIME and Grad-Cam, to interpret and visualize the predictions of our model. These techniques help in understanding the model’s decisions, debugging issues, and building trust in its outputs.