Revolutionizing plant disease diagnosis: PDDD-PreTrain models outperform traditional methods

Diagnosing plant disease is essential to meet the world’s growing food demand, which is expected to increase with a population of 9.1 billion by 2050. Diseases can reduce crop yields by 20-40% , so early detection is critical . Traditional disease identification methods include expert analysis and machine learning-based image processing. However, the manual approach is inefficient and error-prone, while machine learning, particularly deep learning methods like Convolutional Neural Networks (CNNs), has revolutionized disease detection by extracting detailed image features. However, these models are often pre-trained on non-botanical datasets like ImageNet and lack specific plant diseases domain knowledge, resulting in limited accuracy. This gap highlights the need to develop pre-trained models with specialized knowledge of plant phenotypes and diseases to improve the accuracy of plant disease diagnosis.

In May 2023, Plant Phenomics published a research article entitled by "PDDD-PreTrain: A Series of Commonly Used Pre-Trained Models Support Image-Based Plant Disease Diagnosis".

In this study, the authors developed a series of pre-trained models for plant disease diagnosis using a comprehensive dataset  called PDDD (plant disease diagnosis dataset), which contains over 400,000 images of 40 plant species across 120 disease classes. Different model structures and parameters were explored to suit different diagnostic scenarios and devices. For evaluation, the researchers used methods like classification recognition accuracy and mean average precision (mAP), using datasets like Kaggle plant disease dataset and PlantDoc for testing. The results showed that the pre-trained models based on the PDDD and PlantVillage datasets significantly outperformed those trained  on ImageNet alone. This was evident in tasks like plant disease classification, where the hybrid model combining the PlantVillage and ImageNet datasets excelled. In plant disease detection, the Faster R-CNN model initialized with weights from PDDD and ImageNet showed improved detection accuracy and generalization capacity. Similarly, in plant disease segmentation, DeeplabV3 models pre-trained on PDDD and ImageNet achieved higher accuracy, highlighting the advantage of incorporating domain-specific knowledge into the models.

In summary, these results highlight the importance of using large-scale, domain-specific datasets for pre-training in plant disease diagnosis. By making these models open-source, the authors aim to aid further research in this field and provide a basis for advanced and efficient plant disease diagnostic methods. The success of these models marks a significant step in the use of deep learning for plant disease diagnosis, suggesting potential applications in precision agriculture and other related  fields.

###

References

Authors

Xinyu  Dong¹, Qi  Wang^1*, Qianding  Huang¹, Qinglong  Ge¹, Kejun  Zhao¹, Xingcai  Wu¹, Xue  Wu¹, Liang Lei², and Gefei  Hao^1,3*

Affiliations

¹State Key Laboratory of Public Big Data, College of Computer Science and Technology, Guizhou University, Guiyang 550025, China.

²The School of Physics and Optoelectronic Engineering, Guangdong University of Technology, Guangzhou 510006, China.

³National Key Laboratory of Green Pesticide, Key Laboratory of  Green  Pesticide  and  Agricultural  Bioengineering,  Ministry  of  Education,  Guizhou  University,  Guiyang  550025, China.

About Qi  Wang

He is a professor in the College of Computer Science and Technology at Guizhou University. His research interests include Artificial Intelligence; Smart Agriculture; Computer Vision; Plant Disease Image Analyze; Precision Agriculture; Plant Protection; Knowledge Discovery and Data Mining; Image analysis and understanding; Few-shot learning; Fine-grained analysis; Image captioning; Adversarial attack; Obiect detection; Image segmentation; Multi-view stereo; Depth estimation.