Comparative study on automatic identification and performance of cotton pests and diseases based on deep learning

Cotton yield is significantly affected by pests and diseases. This study proposes using three deep learning models—AlexNet, GoogLeNet, and ResNet18—to automatically identify cotton pests and diseases. The models were trained on a dataset containing multiple types of cotton pests and diseases, with normalization and standardization applied to the data, achieving recognition accuracies of $72.22 \%, 90.12 \%$, and 93.67%, respectively. Ablation experiments conducted on GoogLeNet and ResNet18 showed that the Inception module and residual blocks contributed 6.92% and 4.15% to the accuracy of their models, respectively. At the end of the article, an improved ResNet model was proposed, achieving a recognition accuracy of 95.68%. This study concludes that AlexNet is not well-suited for practical applications due to its lower accuracy. In contrast, the multi-scale feature extraction capabilities of GoogLeNet and the detailed feature extraction of ResNet18 render them highly effective and promising for identifying cotton pests and diseases.