Deep-learning-based identification of Halyomorpha halys using phenotypic image analysis

The identification of plant pests is crucial for effective phytosanitary monitoring and agricultural decision-making. The traditional identification methods often require expert knowledge and manual examinations, making them time-consuming and prone to inconsistencies. This study proposes a deep-learning-based approach to automating the identification of Halyomorpha halys.

Methods: The dataset comprised 10,000 labeled images of adult H. halys specimens. A Convolutional Neural Network (CNN) was implemented using TensorFlow and Keras. The data preprocessing included image augmentation, resizing, and normalization to improve the model generalization. The CNN architecture incorporated Global Average Pooling, dropout layers, and dense layers for robust identification. The model was trained on a stratified dataset using categorical cross-entropy loss and the Adam optimizer. The training progress was monitored through loss convergence and accuracy metrics.

Results: The trained model achieved a identification accuracy of 85-95% on the test set, with precision, recall, and F1-score values of 0.85-0.95, 0.80-0.93, and 0.82-0.94, respectively. The confusion matrix analysis revealed that the pest was correctly identified 85-95% of the time. The area under the receiver operating characteristic (ROC-AUC) curve was 0.90-0.98, indicating a high level of discriminatory capability. The validation loss and accuracy trends indicated minimal overfitting, suggesting a well-generalized model. Additionally, an error analysis was performed to point out misidentified instances, revealing potential areas for improvement through additional data augmentation or hyperparameter tuning.

Conclusion: This study validates the feasibility of CNN-based identification for adult plant pests using a phenotypic image analysis. The proposed method enhances the automated monitoring capabilities, reducing the need for manual inspections. Future work will focus on expanding the dataset and refining the model to improve the identification accuracy in real-world applications.