Muhammad Raihan, Cindi Wulandari, Rusdiyanto Rusdiyanto, Rudi Kurniawan

Klasifikasi Penyakit Tanaman Mangga Menggunakan Citra Daun Dengan Pendekatan Transfer Learning Efficientnet-B0

Introduction

Klasifikasi penyakit tanaman mangga menggunakan citra daun dengan pendekatan transfer learning efficientnet-b0. Klasifikasi penyakit daun mangga otomatis dengan citra menggunakan Transfer Learning EfficientNet-B0. Sistem identifikasi 8 kelas penyakit/sehat capai akurasi tinggi, bantu tingkatkan produktivitas panen.

Abstract

Mango plant disease is one of the factors that can reduce the quality and productivity of the harvest. Manual identification of mango leaf diseases still relies on visual observation, potentially requiring a long time and producing inconsistent diagnoses. This study aims to develop a mango plant disease classification system based on leaf images using a transfer learning approach with the EfficientNet-B0 architecture. The dataset used consists of eight classes, namely seven types of mango leaf diseases and one class of healthy leaves. EfficientNet-B0 is used as a feature extractor with pre-trained weights from ImageNet, then custom layers are added in the form of Batch Normalization, Dense, and Dropout to adjust to classification needs. The training process was carried out for 10 epochs by dividing the data into training, validation, and test data. The test results show that the model achieves maximum performance on the dataset used, indicated by very high accuracy, precision, recall, and f1-score values ??across all classes. Visualization of individual image predictions also shows a prediction confidence level of 0.91 for one of the disease classes. While these results demonstrate the potential of EfficientNet-B0 in mango leaf disease classification, the very high performance achieved in a limited number of epochs indicates the need for further evaluation of the model's generalization capabilities.

Review

This study presents a timely and relevant approach to classifying mango plant diseases using leaf imagery, aiming to improve upon the traditional, often inconsistent, manual identification methods. The authors propose a deep learning solution leveraging transfer learning with the EfficientNet-B0 architecture, which is a modern and efficient choice for image classification tasks. The system is designed to distinguish between seven specific mango leaf diseases and healthy leaves, utilizing a custom dataset. The initial results, indicating remarkably high accuracy, precision, recall, and f1-score across all classes within a mere 10 epochs of training, suggest a highly effective model within the confines of the used dataset. Methodologically, the choice of EfficientNet-B0 as a feature extractor, pre-trained on ImageNet, is a well-justified strategy for transfer learning, capitalizing on pre-existing knowledge to accelerate learning on a new, specific domain. The addition of custom Batch Normalization, Dense, and Dropout layers further demonstrates a thoughtful adaptation of the architecture for the specific classification needs, aiming to prevent overfitting and improve generalization. The reported high confidence levels for individual predictions also point towards a robust classification ability for the instances it has encountered, underscoring the potential for this approach to significantly aid in automated disease detection and management in mango cultivation. While the exceptionally high performance achieved in a limited number of epochs is encouraging, the abstract rightly highlights a critical area for further scrutiny: the model's generalization capabilities. Such rapid convergence often warrants a deeper investigation into the dataset's diversity, potential biases, or the difficulty of the classification task itself. Future work should rigorously validate the model on a significantly larger and more varied dataset, including images captured under different environmental conditions, camera settings, and geographical locations to ensure real-world applicability. Implementing robust cross-validation strategies or evaluating the model on an entirely independent, unseen dataset would provide a more reliable assessment of its robustness and practical utility beyond the confines of the current experimental setup.

Full Text

Comments

You need to be logged in to post a comment.