Hyperparameter Optimization in Transfer Learning for Improved Pathogen and Abiotic Plant Disease Classification

The application of machine learning, particularly through image-based analysis using computer vision techniques, has greatly improved the management of crop diseases in agriculture. This study explores the use of transfer learning to classify both spreadable and non-spreadable diseases affecting soybean, lettuce, and banana plants, with a special focus on various parts of the banana plant. In this research, 11 different transfer learning models were evaluated in Keras, with hyperparameters such as optimizers fine-tuned and models retrained to boost disease classification accuracy. Results showed enhanced detection capabilities, especially in models like VGG_19 and Xception, when optimized. The study also proposes a new approach by integrating an EfficientNetV2-style architecture with a custom-designed activation function and optimizer to improve model efficiency and accuracy. The custom activation function combines the advantages of ReLU and Tanh to optimize learning, while the hybrid optimizer merges feature of Adam and Stochastic Gradient Descent (SGD) to balance adaptive learning rates and generalization. This innovative approach achieved outstanding results, with an accuracy of 99.96% and an F1 score of 0.99 in distinguishing spreadable and non-spreadable plant diseases. The combination of these advanced methods marks a significant step forward in the use of machine learning for agricultural challenges, demonstrating the potential of customized neural network architectures and optimization strategies for accurate plant disease classification.