A meta-learning framework for leaf disease detection using vision transformer-based feature extraction, PCA, and tuned SVM classifier

Abstract

A hybrid meta-learning approach is proposed for effective leaf disease detection by integrating vision transformer (ViT), principal component analysis (PCA), and support vector classifier (SVC). The objective of this study is to accurately classify plant leaf conditions into three categories: healthy, angular leaf spot, and bean rust. The dataset consists of 1,167 labeled leaf images, divided into training (974 images), validation (133 images), and testing (60 images) sets. A pretrained ViT model is employed for feature extraction, producing a feature vector of shape (974, 64) for the training data. To mitigate the curse of dimensionality and improve classification performance, PCA is applied, reducing the features to 41 principal components while retaining 98% of the original variance and accuracy 97.85%. For the classification task, an SVC is used and fine-tuned using the Optuna hyperparameter optimization framework to enhance accuracy and generalization. A distributed deep learning strategy is incorporated to accelerate training and scale computation, while the tf.data API is utilized to construct an efficient and scalable data input pipeline. The hybrid model demonstrates strong classification performance on the test set, indicating that combining deep transformer-based feature extraction with dimensionality reduction and optimized classical machine learning classifiers is effective for plant disease identification. This approach offers a robust and computationally efficient solution for precision agriculture, enabling automated and accurate leaf disease diagnosis and supporting early intervention strategies in crop management.