METHOD FOR PLANT LEAF PATHOLOGY CLASSIFICATION BASED ON CONVOLUTIONAL NEURAL NETWORKS USING DISTRIBUTED PARALLEL TRAINING TECHNOLOGIES

Abstract

The effective operation of automated plant disease diagnosis systems is a critical factor in modern precision agriculture, requiring robust solutions for early detection of pathologies. Traditional diagnostic methods based on visual inspection are labor-intensive, subjective, and difficult to scale. This paper presents an improved method for classifying pathologies of agricultural plant leaves based on deep learning technologies, specifically aimed at increasing diagnostic accuracy and optimizing computational performance in high-load environments.

A modified five-block Convolutional Neural Network (CNN) architecture, derived from the VGG16 baseline, is proposed. The key architectural innovation involves the deep integration of batch normalization mechanisms after convolutional layers and dropout regularization in the fully connected layers. These modifications successfully addressed the issue of overfitting on limited datasets, ensuring the model's robustness to variations in input data and improving feature extraction capabilities for complex disease patterns.

To ensure the efficiency of experimental studies involving large-scale image datasets, a distributed parallel training technology was implemented. This approach relies on the principle of data parallelism with synchronous gradient updates across multiple computing nodes. The implementation allowed for a significant reduction in model training time and provided horizontal scalability of the system, making it suitable for processing big data. Furthermore, the paper describes the technological aspects of software creation, emphasizing the use of declarative configuration and a comprehensive versioning system (following MLOps principles). This approach guarantees the full reproducibility of experiments, systematic documentation of the development process, and reliability of the obtained results.

Experimental studies were conducted using a representative dataset of bean leaf images classified into four distinct categories. The results established that the proposed method achieves a classification accuracy of 91.2%, outperforming the baseline model by 4%. The study also proved the critical impact of data augmentation techniques on the model's generalization ability, particularly under conditions of variable lighting and diverse shooting angles. The obtained results confirm the practical value of the method for developing scalable automated diagnostic systems.