Plant Disease Recognition using Two-Dimensional Convolutional Neural Networks

This study explores the application of a two-dimensional Convolutional Neural Network (2D-CNN) in the classification of potato leaf diseases, with a particular emphasis on Early Blight, Late Blight, and healthy leaves. A dataset comprising 208 labeled images was compiled and preprocessed through resizing and normalization to ensure better model generalization. The proposed CNN architecture consists of sequential convolutional layers with ReLU activation, max-pooling operations for dimensionality reduction, and fully connected layers for classification. The training process was guided by categorical cross-entropy and accuracy metrics, while dropout and early stopping techniques were employed to mitigate overfitting. The optimized model, which incorporated four convolutional stages (32, 64, 128, and 256 filters), was trained for multiple epoch settings to ascertain the optimal generalization balance. The model was evaluated on a held-out test set containing unseen images from each class. The experimental results demonstrate that the 2D-CNN achieved an overall accuracy of 71.7%, with F1-scores of 0.80 for Early Blight, 0.70 for Healthy, and 0.63 for Late Blight. These quantitative outcomes indicate the CNN's robustness in distinguishing between visually similar foliar conditions, even under small dataset constraints. Among the epoch trials, 10 epochs were identified as the most optimal setting, providing a trade-off between convergence and the prevention of overfitting. The findings indicate the viability of 2D-CNNs for efficient, automated detection of potato leaf blight, offering valuable tools for timely agricultural disease management and contributing to reduced yield loss and enhanced sustainability.