CNN Backbones with PCA and Gaussian Membership Function Pipelines for Maize Pest Identification: A Comparative Accuracy-Efficiency Study

Abstract Identifying morphologically similar maize pests is important because fall armyworm, African armyworm, and maize stalk borer require different management responses. This study examined whether lightweight CNN backbones can retain high predictive performance while reducing computational requirements relative to VGG16. VGG16, MobileNetV2, EfficientNet-B0, and MobileNetV3-Large were evaluated across backbone-only, PCA, and PCA-GMF pipelines using ten predefined random seeds. Performance was assessed using accuracy, precision, recall, and F1-score, while efficiency was measured using parameter count, FLOPs, and inference latency. VGG16-PCA-GMF achieved the highest accuracy (98.89±0.20%) and F1-score (98.88±0.20%) but required 15.267 million parameters and 30.714 GFLOPs. EfficientNet-B0-PCA-GMF achieved 98.14 ± 0.38% accuracy and F1-score, with only a 0.75 percentage-point reduction in accuracy while reducing parameters by 65.3% and FLOPs by 97.4%. MobileNetV3-Large and MobileNetV2 produced greater computational savings but larger performance reductions. The findings identify EfficientNet-B0-PCA-GMF as the strongest balance between predictive performance and computational efficiency. They can guide finegrained image-classification research and support future accessible pestidentification tools for farmers and agricultural extension practitioners.