Publication

Abstract : Leaf classification is essential for advancing botanical research, agricultural monitoring, and biodiversity preservation. Convolutional Neural Networks (CNNs), particularly MobileNetV2, offer efficient solutions for automated leaf identification, especially on resource-constrained devices like mobile platforms. The choice of activation function in deep learning models significantly influences performance by introducing non-linearity and affecting gradient flow. This study systematically evaluates the impact of 11 activation functions—ReLU, Leaky ReLU, PReLU, ELU, SELU, Swish, GELU, Tanh, Sigmoid, Softplus, and Softmax—on a fine-tuned MobileNetV2 model for multiclass leaf classification. Using the dataset with 96 classes, we fine-tuned MobileNetV2 and assessed performance through training and validation accuracy, loss, precision, recall, and F1-score. Results show that GELU consistently outperform traditional functions like ReLU, while Sigmoid and Tanh yield suboptimal results due to saturation issues. This work provides valuable insights for optimizing activation functions in lightweight CNNs for leaf classification tasks.