Deteksi Penyakit Daun Padi Berbasis Deep Learning untuk Pertanian Presisi: Studi Komparatif ResNet-18 dan Arsitektur CNN

Abstract

Rice plant diseases can reduce agricultural productivity and yield quality, highlighting the need for rapid and accurate identification methods to support precision agriculture. This study evaluates four deep learning architectures—ResNet-18, VGG-16, MobileNetV2, and Inception V3—for rice leaf disease classification using digital images. The dataset underwent preprocessing and augmentation before training, while transfer learning with ImageNet pre-trained weights was applied. Models were trained using CrossEntropyLoss and Adam optimizer with a learning rate of 0.0001 for 10 epochs. Performance was evaluated using accuracy, precision, recall, F1-score, AUC, and confusion matrix. Results demonstrate that the proposed ResNet-18 achieved the best overall performance, obtaining 96.94% accuracy, 100% precision, 95.45% recall, 96.18% F1-score, and 1.00 AUC. Inception V3 and MobileNetV2 showed competitive performance, whereas VGG-16 achieved relatively lower recall. These findings indicate that ResNet-18 provides effective feature representation and stable learning, demonstrating strong potential for rice leaf disease classification and future integration into mobile, field-camera, and IoT-based precision agriculture systems.