Citrus greening disease recognition algorithm based on classification network using TRL-GAN
Introduction
Citrus Greening Disease, known as citrus cancer, has caused serious damage to the citrus industry. It is extremely easy to spread by psylla - the disease’s infection source.Up to now,the most effective way to avoid the spread of citrus greening disease is to dig, burn and sterilize the diseased plants in time. The traditional detection methods for yellow dragon disease include field diagnosis, indicator crop identification, microscopic observation of the pathogen, serological identification, hybridization standards, DNA probe hybridization and PCR techniques. At present, the more accurate and reliable method is the PCR technique, which is time-consuming and costly. For field diagnosis, It is possible to be misdiagnosed by personal experience or diagnosis by phytopathologists (Deng et al., 2016). Citrus greening disease is diagnosed by phytopathologists mainly based on the appearance of a large number of mottled leaves (Garnier and Bové, 1993, do Carmo Teixeira et al., 2005, Varma et al., 1993, Teixeira et al., 2008). The mottled leaves display asymmetrical yellowing spots, which are distinct from the uniform yellowing caused by elemental deficiency, due to the gradual accumulation of Huanglong disease bacteria in citrus leaves caused by the necrosis of the leaf vein bast. Usually, untrained personnel can easily be confused with various types of citrus leaf yellowing when identifying mottled leaves. Therefore, at this stage, there is a need for a rapid and intelligent method to detect citrus yellow dragon disease to replace manual diagnosis.
In recent years, there have been a large number of experiments using machine learning to identify diseased leaves instead of personal experience or specialists. Traditional plant disease leaf recognition mainly extracts features through the corresponding filters designed for color, shape, texture, etc. and uses these features to classify them through machine learning methods such as random forests, support vector machines, neural networks, and others (Jiang et al., 2020, Manavalan, 2020, Yadav et al., 2019, Sun et al., 2019, Bischoff et al., 2021). There are some identification methods of citrus greening disease. For example, Xiaoling Deng et al. (2016) employed the C-SVC method to detect Huanglong diseased leaves, and Gómez-Flores et al. (2019) extracted features by producing diseased leaf data based on constant light brightness and then conducted random forest classification. These methods can solve the majority of the problems, but they involve human feature selection and cost a huge amount of time.
Because of the limited number of manually selected features, and the wide variety of crop diseases, manual extraction of features requires a lot of labor and time costs, resulting in the accuracy of traditional machine learning recognition is hard to achieve practical application standards. Nowadays, deep convolutional models can be used to learn features instead of manual feature extraction. With deep learning technology and increasingly sufficient plant disease image data, the traditional machine learning shows that the problems of extracting plant features and poor generalization ability are effectively solved, and many plant diseases can be recognized simultaneously by using deep convolutional networks instead of traditional manual features extraction (Chaudhary et al., 2020, Ma et al., 2018, Li et al., 2020, Zhou et al., 2021, Waheed et al., 2020, Zeng and Li, 2020, Zhang et al., 2019b, Zhang et al., 2019a, Zhang et al., 2021, Gao et al., 2021, Wang et al., 2021, Abade et al., 2021, Jiang et al., 2021, Zhong and Zhao, 2020, Zhang et al., 2019b, Zhang et al., 2019a, Gui et al., 2021). Likewise, deep learning has been made a certain achievement in the application of citrus greening disease at this point. For instance, Utpal Barman et al. (2020) applied MobileNet to recognize citrus diseased leaves, which include citrus green leaves, yellow dragon disease diseased leaves, and fast fading diseased leaves.However, once the plant is infected with the virus, it takes a long time for relevant phenotypic characteristics to emerge in reality. Citrus orchards are cleaned up as soon as the symptoms of the citrus greening disease appear, making data scarce and more difficult to collect and produce. Consequently, the amount of disease leaf data fails to meet the needs of the recognition algorithm data training of citrus greening disease.
To address the above problem of limited samples of plant diseases, Zhong et al. (2020) used a CVAE-like structure CAAE for citrus disease leaf data expansion, and used Zero- and few-shot learning for recognition based on these data. The approach can tackle the shortage of sample quantity and also gets effective results with a novel few-sample learning method for classification. However, the VAE series network is a display generative network structure, and there exists a more blurred generated image and a more concentrated distribution of generated data.
Since Zn-deficient leaves, Mg-deficient leaves, yellow veins leaves, and yellow diachyma leaves are easily confused with mottled leaves, this paper focuses on the phenotypes of mottled leaves, Zn-deficient leaves, Mg-deficient leaves, yellow veins leaves, and yellow diachyma leaves. This paper also improves CycleGAN, and takes advantage of the fact that it does not require a large amount of manual annotation, and thus proposes a Texture Reconstruction Loss CycleGAN (TRL-GAN) to expand the number of citrus yellow dragon disease leaves in realistic scenes and improve the quality of the generated images. In the following, the preparation of the dataset will be described, then the principle of the algorithm in this paper will be introduced, and finally the analytical experiments will be conducted to analyze the presence of techniques such as T-SNE in the experiments (Van der Maaten and Hinton, 2008).
Section snippets
Dataset collection
The data of this experiment mainly came from the public dataset PlantVillage and the photo collection of citrus diseased plants from the College of Plant Protection, South China Agricultural University. The data collection scene is shown in Fig. 1.
The device for collecting data of the citrus diseased plants in College of Plant Protection of South China Agricultural University was a OnePlus 8 T phone with a 48MP camera, 15 cm–30 cm shooting distance from diseased leaves, 35 mm equivalent focal
Results and discussion
To analyse the TRL-GAN results, this paper validates the effectiveness of the TRL-GAN network by using generative adversarial network evaluation indexes, subjective evaluation indexes, and classification networks.
Conclusions and outlooks
This paper mainly focuses on TRL-GAN for generating phenotype data of citrus greening disease, and to improve the performance of classification networks for citrus diseased leaf identification. Compared to the small dataset, this paper expands the dataset by Mask-RCNN to cull the background of the photographed leaves, followed by image synthesis using a more reasonable non-pairwise GAN technique. What’s more, in addition to introducing the original loss function of CycleGAN, it also adds
CRediT authorship contribution statement
Deqin Xiao: Conceptualization, Supervision, Investigation, Funding acquisition. Ruilin Zeng: Conceptualization, Methodology, Software, Validation, Investigation, Formal analysis, Data curation, Writing – original draft, Writing – review & editing. Youfu Liu: Conceptualization, Methodology, Data curation. Yigui Huang: Data curation. Junbing Liu: Data curation. Jianzhao Feng: Data curation. Xinglong Zhang: Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This work was supported by the Science and Technology Planning Project of Guangzhou (Air, space and ground integrated intelligent planting monitoring and pest early warning system, grant number 202206010116), and Agricultural Research Project and Agricultural Technology Promotion Project of Guangdong (grant number 2021KJ383).
References (38)
- et al.
Plant diseases recognition on images using convolutional neural networks: A systematic review
Comput. Electron. Agric.
(2021) - et al.
Comparison of convolution neural networks for smartphone image based real time classification of citrus leaf disease
Comput. Electron. Agric.
(2020) - et al.
Technological support for detection and prediction of plant diseases: A systematic mapping study
Comput. Electron. Agric.
(2021) - et al.
A particle swarm optimization based ensemble for vegetable crop disease recognition
Comput. Electron. Agric.
(2020) Enhancing two-view correspondence learning by local-global self-attention
Neurocomputing
(2021)- et al.
Citrus greening detection using visible spectrum imaging and C-SVC
Comput. Electron. Agric.
(2016) - et al.
Citrus huanglongbing in São Paulo State, Brazil: PCR detection of the ‘Candidatus’ Liberibacter species associated with the disease
Mol. Cell. Probes
(2005) - et al.
Dual-branch, efficient, channel attention-based crop disease identification
Comput. Electron. Agric.
(2021) - et al.
Detection of Huanglongbing disease based on intensity-invariant texture analysis of images in the visible spectrum
Comput. Electron. Agric.
(2019) - et al.
Towards automatic field plant disease recognition
Comput. Electron. Agric.
(2021)
Recognition of rice leaf diseases and wheat leaf diseases based on multi-task deep transfer learning
Comput. Electron. Agric.
Image recognition of four rice leaf diseases based on deep learning and support vector machine
Comput. Electron. Agric.
A solanaceae disease recognition model based on SE-Inception
Comput. Electron. Agric.
A recognition method for cucumber diseases using leaf symptom images based on deep convolutional neural network
Comput. Electron. Agric.
Automatic identification of diseases in grains crops through computational approaches: A review
Comput. Electron. Agric.
SLIC_SVM based leaf diseases saliency map extraction of tea plant
Comput. Electron. Agric.
Distribution and quantification of Candidatus Liberibacter americanus, agent of huanglongbing disease of citrus in São Paulo State, Brasil, in leaves of an affected sweet orange tree as determined by PCR
Mol. Cell. Probes
An optimized dense convolutional neural network model for disease recognition and classification in corn leaf
Comput. Electron. Agric.
Few-shot vegetable disease recognition model based on image text collaborative representation learning
Comput. Electron. Agric.
Cited by (11)
SugarcaneGAN: A novel dataset generating approach for sugarcane leaf diseases based on lightweight hybrid CNN-Transformer network
2024, Computers and Electronics in AgricultureDeep learning model to detect deceptive generative adversarial network generated images using multimedia forensic
2024, Computers and Electrical EngineeringPlant growth information measurement based on object detection and image fusion using a smart farm robot
2023, Computers and Electronics in AgricultureDetection of plant leaf diseases using deep convolutional neural network models
2024, Multimedia Tools and ApplicationsImplementing deep-learning techniques for accurate fruit disease identification
2023, Plant PathologySample Expansion and Classification Model of Maize Leaf Diseases Based on the Self-Attention CycleGAN
2023, Sustainability (Switzerland)