<![CDATA[This study aims to evaluate the performance of Convolutional Neural Network (CNN) models in detecting pests on caisim (Chinese mustard) plants using different architectural approaches, namely CNN from scratch, VGG16, and Xception. The dataset used consists of 1,000 images classified into several disease categories and a healthy class. Five experiments were conducted to compare the effectiveness of the models based on evaluation metrics such as accuracy, precision, recall, and F1-score.The results show that CNN models trained from scratch produced varying levels of performance. The first and second experiments experienced underfitting, with accuracies of 30.25% and 62.29%, respectively. A significant improvement was observed in the third experiment, achieving an accuracy of 83.73% and an F1-score of 0.82, indicating that the model began to better recognize data patterns. The best performance was achieved in the fourth (VGG16) and fifth (Xception) experiments, with accuracies of 91.41% and 92.12%, respectively, and balanced precision, recall, and F1-score values above 0.90.Factors contributing to model success include an optimal proportion of training data, appropriate architectural selection, hyperparameter tuning, and the use of callbacks such as early stopping and model checkpoint. This study demonstrates that selecting the appropriate CNN architecture can significantly improve the accuracy of image classification systems for pest detection in plants]]>
Copyrights © 2025
