Scientific research project

Abstract

Oranges are a highly demanded agricultural product due to their nutritional value and adaptability to various climates. However, orange cultivation faces significant challenges from diseases like bacterial ulcers and yellow leaf disease, which can severely impact yield and quality. In provinces like Ha Giang and Vinh Long, diseases have led to large-scale crop damage, with up to 20% of orange trees affected in some areas. Early disease detection is crucial to minimize losses, reduce pesticide use, and improve crop health. Applying deep learning techniques for detecting and classifying orange diseases can help farmers identify problems early, reduce reliance on chemicals, improve crop performance, and promote sustainable agriculture. Thus, the proposed study aims to develop a deep learning-based method to support farmers in accurately detecting and managing diseases in orange crops, ultimately maximizing benefits for growers.

Introduction

Fruits are among the high-demand agricultural products in the market, with oranges being a focus for development in Vietnam until 2025 and 2030, according to the Ministry of Agriculture and Rural Development [1]. In Vinh Long province, the orange cultivation area reached 15,458 hectares by March 2022, leading the fruit category in the province with an annual yield of over 630,000 tons [2]. Oranges are favored due to their compatibility with the local climate and soil, high yield, and nutritional benefits such as being rich in vitamin C, fiber, folate, and antioxidants while being low in calories and sugar. However, orange cultivation faces significant challenges from pests and diseases, including bacterial canker, huanglongbing (greening disease), and black spot disease, which can severely affect yield and fruit quality. For instance, in Ha Giang province, over 1,560 hectares of orange orchards were affected by diseases in the 2021-2022 season, with some regions reporting over 20% of orchards affected [4]. Similarly, in Vinh Long, diseases such as root rot and greening have impacted nearly 5,000 hectares, with severe infections affecting over 3,000 hectares [5]. Given these challenges, I propose the study "Detection and Classification of Orange Diseases Based on Deep Learning Techniques" to help farmers accurately identify diseases in oranges and save time, allowing for timely interventions and maximizing benefits for orange growers.

Figure 1. Orange diseases dataset

Proposed Method

To address the problem, this study proposes using a general model as shown in Figure 3.2. The model consists of two phases: the training phase and the testing phase.

Preprocessing: The dataset is already pre-labeled; however, for the YOLOv8 network model, in the preprocessing stage, I performed object localization and labeling on the proposed Orange Diseases dataset using the online tool makesense.ai. Then, I normalized the images to a size of 224 x 224 for the EfficientNet-B7 and MobileNetV2 models, 299 x 299 for the Inception-V3 model, and 640 x 640 for the YOLOv8 model.

Feature Extraction and Training: In this phase, as outlined in the proposed section, I applied deep learning networks including Inception-V3, EfficientNet-B7, MobileNetV2, and YOLOv8 for detection and classification. The input dataset, containing color images, is fed into deep neural networks to extract features, and the output is feature vectors corresponding to the four phases. These feature vectors are then passed into a classification network to output class probabilities and generate final bounding boxes.

Figure 2. Proposed model

To conduct an experiment for the proposed model, this study implemented 4 scenarios with the proposed datasets as shown in Table 1.

Table 1. Proposed Scenarios

Results

To ensure the experimental results are objective and comprehensive for all three types of diseases and one type of healthy fruit image, I randomly selected images with small diseased areas or subtle, hard-to-detect lesions to include in the four experimental models for evaluation, as presented in this section.

Figure 3. Some classification results on the Alzheimer's MRI dataset

Throughout the training and testing process, training parameters were adjusted to suit each proposed scenario. A comparison and evaluation were conducted, and the model evaluation criteria are summarized in Table 2.

Table 2. Table of Results Comparing Proposed Scenarios

Conclusions

The study investigated the symptoms of various diseases on oranges, including black spot disease, bacterial blotch, and leaf yellowing with green veins. It examined the use of EfficientNet-B7, Inception-V3, MobileNetV2, and YOLOv8 networks for detecting and classifying these diseases. The successful development of a model for detecting and classifying diseases on oranges was accomplished using the aforementioned network models, with each model—EfficientNet-B7, MobileNetV2, Inception-V3, and YOLOv8—achieving an accuracy rate above 97%. This high-accuracy detection and classification enable early and accurate treatment of diseases, helping farmers to minimize damage caused by pests and diseases, thereby contributing to the sustainable development of agriculture in our country.

References