Klasifikasi Penyakit Daun Bawang Merah Menggunakan Efficientnetv2-B0 Berbasis Cbam dan Grad-Cam

Abstract

Disease diagnosis of shallot (Allium cepa var.aggregatum) leaves at the farmer level still relies on subjective manual visual inspection, which is particularly error-prone in distinguishing downy mildew (Peronospora destructor) from purple blotch (Alternaria porri) due to their similar chlorotic symptoms, resulting in inappropriate fungicide application. This study develops an automated deep learning classification system using the EfficientNetV2-B0 architecture reinforced with a Convolutional Block Attention Module (CBAM) to focus feature representation on pathological lesions, trained through a 4-stage progressive fine-tuning scheme combined with MixUp regularization (α = 0.2), and equipped with Gradient-Weighted Class Activation Mapping (Grad-CAM) for decision transparency. The dataset comprises 3,200 images distributed across four diagnostic classes. Evaluation on 483 independent test images shows an Overall Accuracy of 90.89% and a Macro F1-Score of 90.56%, with Grad-CAM visualizations qualitatively indicating that the model’s attention concentrates on lesion areas. The proposed model is significantly more accurate than VGG16, MobileNetV2, and ResNet-50 (McNemar test, p < 0.05). The system was further deployed into the ShallotGuard AI web-based laboratory platform equipped with preliminary botanical validation. The system remains a laboratory prototype that is promising as a decision-support tool for early diagnosis, but it has not yet been tested directly by farmers in the field, so field validation is still required.