<![CDATA[Early detection of coffee leaf diseases such as leaf rust and Phoma is essential due to its direct impact on crop productivity and quality. Recent studies have shown that lightweight CNN architectures like SqueezeNet are effective for deployment on resource-constrained devices, though they still face limitations in classification accuracy for complex disease types. This study aims to improve the accuracy of coffee leaf disease classification by integrating the SqueezeNet architecture with the SimAM attention module, which enhances feature representation without significantly increasing model complexity. A quantitative experimental approach was used, employing an open-source dataset of coffee leaf images that was augmented and categorized into three classes: healthy leaves, leaf rust, and Phoma. The models were evaluated using accuracy, precision, recall, and F1-score metrics. Results show that integrating SimAM into SqueezeNet increased the model’s accuracy from 81% to 84%. The most significant improvements were observed in the leaf rust and Phoma classes, with F1-scores rising from 0.70 to 0.79 and from 0.73 to 0.76, respectively. Additionally, the AUC score improved to 0.91. These results demonstrate that SimAM integration effectively enhances classification performance, though challenges remain in distinguishing classes with visually similar features. Further research is recommended to implement more aggressive data augmentation and regularization techniques to improve model generalization.]]>
Copyrights © 2025
