Optimasi Model Deep Learning EfficientNet Berbasis Citra Digital untuk Deteksi Penyakit Padi

Abstract

This study optimized the EfficientNet deep learning model based on digital images for rice disease detection, focusing on two main classes: Brown Spot and Leaf Scald, which are constraints to farmer productivity in Pagar Alam City. The dataset consisted of 780 images (476 Brown Spot, 304 Leaf Scald) processed through 224×224 resizing, normalization, data cleaning, and augmentation (rotation, flip, shear, shift, zoom) to improve generalization and reduce overfitting. The model was initialized with transfer learning from ImageNet, trained and fine-tuned at the final layer, and then evaluated using accuracy, precision, recall, and F1-score metrics. EfficientNet B0 showed a high training accuracy of up to 95% with a validation accuracy of around 80%, indicating good detection performance although there are still symptoms of overfitting that need further optimization. The model was then integrated into a web-based expert system for automatic diagnosis from leaf images and presentation of knowledge-based treatment recommendations, thereby accelerating early identification and supporting decision-making in the field. These results confirm EfficientNet's potential as the foundation for a practical, accurate, and applicable rice disease diagnosis system for local agriculture.