A Comparative Analysis of Convolutional Neural Network (CNN): MobileNetV2 and Xception for Butterfly Species Classification

Abstrak

This study aims to compare the effectiveness and efficiency of two convolutional neural network architectures, MobileNetV2 and Xception, for automated butterfly species classification. As biodiversity monitoring gains significance, effective species identification technologies are crucial for conservation. The research utilized a dataset of 100 butterfly species with 12,594 training images and 1,000 validation and test images. Transfer learning with pre-trained ImageNet weights was implemented, and both models were enhanced with custom classification layers. Data augmentation and class weighting mitigated dataset imbalance issues. Experimental results show Xception attained 93.40% test accuracy compared to MobileNetV2's 93.20%. These high accuracy rates were achieved through effective transfer learning that preserved general feature extraction capabilities, comprehensive class balancing techniques, and carefully tailored learning rate strategies for each architecture. Despite minimal performance difference, MobileNetV2 offers significant computational efficiency advantages with 4.15M parameters compared to Xception's 25.27M, while Xception provides marginally better classification. This study contributes to entomological research and highlights trade-offs between model complexity and performance in fine-grained classification tasks, supporting implementation decisions for butterfly identification systems in practical applications.