A comparative deep learning methodology for plant insect image classification: Assessment of CNN architectures and augmentation techniques

Deep learning has become a leading approach for agricultural image analysis and leveraging it for pest recognition has offered tangible value for crop protection. This work has presented a comparative methodology for plant-insect image classification on the BAU-Insectv2 dataset, emphasizing how augmentation choices and optimizers have shaped model behavior on small, field-collected data. We have evaluated four convolutional architectures (ResNet101V2, EfficientNet-B1, InceptionV3, InceptionResNetV1) under transfer learning, six single-factor augmentations, and three optimizers (Adam, SGD, RMSprop). Performance has been assessed with accuracy, precision, recall, and F1-score. Across settings, Adam has generally produced the most stable high accuracy on limited data; model-augmentation pairings have also mattered-e.g., EfficientNet-B1 with cropping has achieved near-perfect accuracy, while ResNet101V2 with rotation and InceptionV3 with brightness have remained competitive. The study has delivered a reproducible pipeline and augmentation-aware guidance that practitioners can adopt when data are scarce, enabling robust insect recognition for downstream agronomic decision support. • We have curated BAU-Insectv2 and designed six single-factor augmentations. • We have benchmarked four transfer-learned CNNs with three optimizers. • We have validated with standard metrics and optimizer-augmentation ablations.