Optimasi Mobilenetv2 Dengan Transfer Learning Untuk Klasifikasi Penyakit Daun Cabai

Abstract

This research effort seeks to establish a robust classification model for chili leaf diseases through optimization of the MobileNetV2 architecture using transfer learning methodology. The presence of diseases in chili plants is often a major barrier to agricultural productivity, necessitating the development of a rapid and accurate early detection system. The dataset used for this investigation includes six different leaf condition categories, specifically: Bacterial Spot, Cercospora Leaf Spot, Leaf Curl Virus, Healthy Leaf, Nutrient Deficiency, and White Spot. The investigation process begins with an image pre-processing phase and the application of data augmentation techniques, which aim to increase the variability of the training data while simultaneously reducing the risk of overfitting. Next, the model is trained using pre-trained weights from ImageNet, which are adjusted to align with the inherent visual characteristics of chili leaves. Model evaluation is conducted rigorously based on accuracy, precision, recall, and F1-score metrics. The experimental results demonstrate outstanding performance, with the model achieving an accuracy rate of 99%, an average F1-score of 0.99, and a validation loss of 0.07. These figures demonstrate the model's highly competent generalization ability when applied to new data. Analysis facilitated by a confusion matrix found a very low error rate, with only 11 images (0.73%) misclassified out of a total of 1,500 test images. These results support the assertion that MobileNetV2 optimization is highly efficient and accurate in identifying chili leaf diseases. This model has significant potential for integration into mobile devices or digital image-based smart farming systems, thereby assisting farmers in making informed decisions in real time.