Research on precise identification of wheat pests and diseases based on an enhanced VggNet16 model and transfer learning

• Addressing the issue of insufficient feature extraction in complex scene backgrounds within traditional convolutional neural networks. • The porous convolution and CBAM attention modules effectively enhance model performance. • The adoption of ImageNet pre-trained weights reduces reliance on labelled samples. • MDCVggNet16 achieves a classification accuracy of 97.17%. Wheat pests and diseases cause severe yield losses and threaten global food security. Traditional detection methods fail to adequately meet modern agricultural identification requirements due to issues such as high subjectivity, time-consuming processes, and complex field environments. This study proposes a wheat pest and disease recognition method based on transfer learning and deep convolutional neural networks, establishing an enhanced model named MDCVggNet16. Building upon the ImageNet-pre-trained VggNet16 architecture, this model adapts to wheat pest and disease classification through transfer learning. It incorporates multi-scale dilated convolutions (MD) to expand the receptive field, capturing multi-scale pest and disease features. Additionally, a convolutional block attention module (CBAM) is introduced to utilise channel and spatial attention mechanisms, thereby extracting key lesion features whilst suppressing complex field backgrounds. The study utilised 16,885 original wheat pest and disease images for model training and validation, categorised into nine classes: healthy wheat, powdery mildew, rust, aphid, spider mites, leaf blotch, root rot, wheat scab, and wheat smut. Experiments were conducted on a hybrid dataset combining PlantVillage and field-collected images, comparing results against ResNet50, DenseNet121, MobileNetv2, VggNet16, and VggNet19 models. Results demonstrated that MDCVggNet16 achieved an accuracy of 97.17%, surpassing other benchmark models. The model exhibited outstanding classification performance and robustness against complex field backgrounds, with recognition accuracy approaching 100% for categories such as healthy wheat and wheat smut. This efficiently meets field detection requirements, providing theoretical and technical support for identifying and diagnosing wheat leaf pests and diseases in complex environments.