MoLLe: A Hybrid Model for Classifying Diseases in Chili Plants Using Leaf Images

Abstract

Purpose: Leaf diseases are often early indicators of problems in plants. More detailed image information with feature extraction on leaves can improve accuracy. However, MobileNetV2 tends to be less than optimal in capturing the fine texture characteristics of leaves. This research aims to propose a classification model for diseases in chili plants based on leaf images using MobileNetV2 with Local Binary Pattern (LBP), with three fully connected layers (220-120-60 neurons) using the ReLU activation function, referred to as MoLLe.

Methods: This research consists of six stages. It begins with a dataset collected from chili farms comprising 900 images, which are then preprocessed into 3,600 images. Next, LBP feature extraction is performed. After that, a comparison between the benchmark architecture and the proposed architecture is conducted. A softmax layer is used to perform three-class classification. The MoLLe model was tested with the MobileNetV2 and MobileNetV2+LBP benchmark architectures and evaluated using a confusion matrix.

Result: Based on the evaluation conducted, using batch size 32, learning rate 0.001, and 20 epochs, the MoLLe model experienced early stopping at epoch 11, achieving an accuracy of 0.97 training data, 0.84 validation data, and 0.91 testing data. The evaluation results showed consistent precision, recall, and F1-score values of 0.91, indicating the model's balanced ability to identify the three disease classes.

Novelty: The novelty of this research lies in the integration of MobileNetV2 and LBP with modifications to three fully connected layers, which not only reduces the number of training parameters but also accelerates the detection process. This research makes an essential contribution to the development of more efficient and effective plant disease detection systems, with experimental results showing that MoLLe outperforms the benchmark architecture.