From Handcrafted Features to Vision Transformers: A Study on Coconut Leaf Disease Classification
Аннотация
Coconut leaf diseases significantly impact crop productivity and economic sustainability, making early and accurate diagnosis crucial for effective agricultural management. Traditional visual inspection methods are labor-intensive, subjective, and often lead to delayed disease identification, motivating the need for automated image-based diagnostic systems. This study presents a unified comparative framework for six-class coconut leaf disease classification using classical machine learning, deep learning, and transformer-based approaches. The framework integrates dataset preprocessing, exploratory analysis, handcrafted feature extraction, and transfer learning with modern neural architectures. Classical models including k‑nearest neighbors (KNN), support vector machine (SVM), logistic regression, and random forest were evaluated using color and texture descriptors, while convolutional neural networks such as MobileNetV2, EfficientNet-B0, Inception-V3, and ResNet50 were applied through transfer learning. Additionally, a Vision Transformer (ViT-Tiny) model was employed to capture global contextual relationships in leaf images. Experimental results show a progressive improvement from feature-based methods to deep representation learning, with random forest and ResNet50 achieving 93.3% accuracy, while the transformer-based model achieved the best performance with 98.0% accuracy and an F1-score of 0.98. The results highlight the effectiveness of attention-based global feature modeling and establish strong performance baselines for automated coconut disease diagnosis in precision agriculture.