The effect of activation functions and the number of layers on the identification of citrus leaf diseases using CNN

The identification of citrus diseases is crucial for maintaining the well-being and efficiency of citrus cultivation. The present study aims to examine the efficacy of various activation functions in the identification of citrus diseases through the application of deep learning methodologies. The activation functions being evaluated comprise Sigmoid, tanh, and Rectified Linear Unit (ReLU). The present investigation employs the Citrus Leaves Data set, comprising 58 images of healthy citrus leaves and 551 images of ailing citrus leaves afflicted by different diseases, including Black Spot, Melanose, Canker, and Greening. The experimentation involves the utilisation of four distinct neural network architectures, namely a network with 2 convolution layers, a network with 4 convolution layers, a network with 6 convolution layers, and the VGG16 model. The experimental findings indicate that the utilisation of a neural network comprising four convolution layers and tanh activation function results in the highest level of accuracy. Specifically, this model achieved an accuracy rate of 71.35% in the detection of citrus diseases.