Classification of Sugarcane Leaf Disease with AlexNet Model

Abstract

This study evaluates the influence of activation functions on the performance of the AlexNet deep learning model in classifying sugarcane diseases. Two popular activation functions, ReLU and LeakyReLU, were compared in terms of classification accuracy and computational efficiency. The ReLU function, known for its simplicity and speed, achieved an accuracy of 87.90% with a total training and testing time of 47 minutes. In contrast, LeakyReLU, which allows a small gradient when the input is negative and hence provides continuity in the learning process, obtained a higher accuracy of 90.67%, albeit at a higher computational cost, taking 54 minutes for the training and testing phase. These results highlight the trade-off between model accuracy and computational time in the deployment of deep learning models for agricultural applications. The study suggests that while LeakyReLU can lead to more accurate models, ReLU remains a competitive choice when efficiency is paramount. Future research should focus on optimizing the balance between accuracy and speed, potentially through the tuning of LeakyReLU parameters or the development of hybrid models.