Comparison and Evaluation of Plant Pest and Disease Image Recognition Models Based on Tensorflow and CNN

DOI: https://doi.org/10.62517/jike.202604201

Author(s)

Jiner Li

Affiliation(s)

Intelligent Technology and Services, School of Data Science, City University of Macau, Macau, China

Abstract

This study addresses the potential for misjudgment in human visual identification of plant diseases, which is highly susceptible to subjective factors, and the variations in accuracy and F1 score among different image recognition models in the same environment. Utilizing comparative experimental methods and literature analysis, this paper explores the performance of two pre-trained models, MobileNetV3 and EfficientNetV2, in image recognition tasks, ultimately selecting the optimal model for agricultural pest and disease identification. Initially, a dataset was collected from the internet, encompassing 17 plant species. Tomatoes, cassava, corn, and apples each had no fewer than four types of pests and diseases, along with their healthy samples. The remaining plants each had at least one type of pest and disease, along with their healthy samples. After cleaning, removing ambiguous and incorrectly labeled samples, the dataset was divided into a training set, validation set, and test set in an 8:1:1 ratio. To address the scarcity of samples, additional samples were generated through random image augmentation techniques (flipping, rotating, brightness adjustment), ensuring each sample reached a total of 200. Subsequently, model construction was carried out. Based on the make_model function, MobileNetV3 Large and EfficientNetV2 B0 pre-trained models were flexibly constructed, configured with regularization, Dropout, and other overfitting strategies, and specified with the Adamax optimizer, cross-entropy loss, and F1 score evaluation metrics. Model training was then initiated, with settings for learning rate decay, early stopping callback, and interactive specification of training epochs, while monitoring convergence in real-time. Finally, the model convergence trend was analyzed through visualization of the training curve. Comprehensive evaluation of model performance, including accuracy and F1 score, was conducted using the test set, confusion matrix, and classification report, comparing the adaptability of the two pre-trained models. Experimental results indicated that the EfficientNetV2 B0 model achieved an accuracy of 93.41%, approaching an F1 score of 92.40%. The MobileNetV3 Large model reached a precision of 93.15%, nearing an F1 score of 92.16%. Evidently, the EfficientNetV2 B0 model outperforms the MobileNetV3 Large model in terms of accuracy and F1 score, particularly in identifying "confusable pests and diseases," making it suitable for scenarios demanding higher precision. Conversely, the MobileNetV3 Large model boasts a smaller parameter count and faster training/inference speed by approximately 15% to 20%, making it ideal for edge devices with limited computational power.

Keywords

Plant Pest and Disease Classification; Pre-Trained Model (EfficientNetV2 B0/MobileNetV3 Large); ACCURACY; F1 SCORE

References

[1] Bi, C., Xu, S., Hu, N., Zhang, S., Zhu, Z., & Yu, H. 2023. Identification method of corn leaf disease based on improved Mobilenetv3 model. https://scholar.google.com/scholar?hl=zh-CN&as_sdt=0%2C5&q=mobilenetv3+plant+disease&oq=MobileNetV3++plant [2] Devi, R. S., Kumar, V. R., & Sivakumar, P. 2023. EfficientNetV2 Model for Plant Disease Classification and Pest Recognition. https://scholar.google.com/scholar?hl=zh-CN&as_sdt=0%2C5&q=EfficientNetV2+plant+disease&oq= [3] Jiao Jiabao, Li Lingyi, Liu Yongjian, 2025. Research on Rice Pest and Disease Recognition System Based on Improved EfficientNet-V2.https://scholar.google.com/scholar?hl=zh-CN&as_sdt=0%2C5&q=%E7%84%A6%E4%BD%B3%E5%AE%9D&btnG= [4] Wang Jingjing, Zhang Wu, Liu Lianzhong, Huang Shuai, 2014. A Review of Crop Pest and Disease Image Recognition Technology. https://scholar.google.com/scholar?hl=zh-CN&as_sdt=0%2C5&q=%E6%B1%AA%E4%BA%AC%E4%BA%AC&btnG= [5] Rahman, C. R., Arko, P. S., Ali, M. E., Khan, M. A. I., Apon, S. H., Nowrin, F., & Wasif, A. (2020). Identification and recognition of rice diseases and pests using convolutional neural networks. Biosystems Engineering, 194, 112-120. [6] Shoaib M,Sadeghi-Niaraki A,Ali F, 2025. Leveraging deep learning for plant disease and pest detection: a comprehensive review and future directions. https://scholar.google.com/scholar?hl=zh-CN&as_sdt=0%2C5&q=.+Shoaib+M%E3%80%81Sadeghi-Niaraki+&btnG= [7] Llorca, C., Yares, M. E., & Maderazo, C. 2018,. Image-based pest and disease recognition of tomato plants using a convolutional neural network. https://scholar.google.com/scholar?hl=zh-CN&as_sdt=0%2C5&q=Tomato+pest+and+disease+image+recognition&btnG= [8] Prechelt, L. 1998. Automatic early stopping using cross validation: quantifying the criteria.https://scholar.google.com/scholar?hl=zh-CN&as_sdt=0%2C5&q=early+stopping+neural+network&btnG=&oq=early+stopping