Analisis Variasi Arsitektur Convolutional Neural Network Model Xception untuk Klasifikasi Penyakit Daun Padi (Oryza sativa L)

Abstract

Convolutional Neural Network (CNN) has become one of the most reliable deep learning methods for image processing due to its strong ability to capture visual patterns. In rice leaf disease classification, CNN plays an important role because it can extract key features such as color variations, spots, and texture changes that are often difficult to identify manually. However, many previous studies still rely on standard CNN architectures without examining how architectural variations may improve classification performance. Therefore, this study compares two Xception-based CNN models: a baseline model and a variation model modified by adjusting the number of filters, input size, padding, and stride.The dataset consists of 17,800 rice leaf images from three sources, including four major disease classes. The research workflow includes data preprocessing, model training, evaluation using akurasi, precision, recall, F1-score, and loss, followed by integrating the best-performing model into a web-based system using Flask and TensorFlow Lite.The results show that the variation model performs better than the baseline, achieving 99,49% akurasi, 99,75% precision and recall, and a lower loss value. These improvements are largely influenced by the architectural modifications, which allow the model to extract more detailed features from the images. The variation model was chosen as the final model and successfully implemented into a web application capable of real-time classification.This study demonstrates that proper architectural adjustments can significantly enhance CNN performance in rice leaf disease classification.