Comparative Evaluation of Deep Features and Traditional Classifiers in the Detection of Potato Leaf Diseases

Abstract Early detection of potato leaf diseases is important for maintaining agricultural productivity and crop quality. This study comparatively evaluates different deep learning–based feature extraction methods and classifier models for potato leaf disease classification under both controlled and field conditions. Two separate datasets were used: the publicly available PlantVillage dataset and the Beunto dataset developed in this study. Each dataset contains 3000 images belonging to three classes: Potato_Early_blight, Potato_Late_blight, and Potato_healthy. Deep features extracted from pre-trained CNN models, namely VGG19, ResNet50, Xception, MobileNetV2, and DenseNet121, were evaluated using both traditional machine learning classifiers and DNN-based classifiers. Beyond a simple model comparison, the study offers a comparative framework for analysing the relationship between feature representation quality and classifier capacity across datasets with different image acquisition characteristics. Among all evaluated combinations, ResNet50 + DNN-3 achieved the best performance, with accuracy values of 98.56% on PlantVillage and 98.20% on Beunto. This comparison further revealed that classification performance was governed not only by the learning capacity of the classifier, but also by the discriminative quality of the extracted deep feature space. In particular, the consistent superiority of ResNet50-based features across different classifier settings indicates that backbone-level representation quality plays a primary role, whereas increasing classifier capacity mainly improves performance when the feature space is already sufficiently separable. Thus, the study provides scientific evidence that the interaction between feature representation and classifier complexity is not arbitrary, but structurally linked to the separability and transferability of disease-related visual patterns. Cross-dataset experiments further showed that the proposed framework retained meaningful generalization ability under different visual conditions. The results demonstrate that combining effective deep feature representations with properly configured classifiers provides a reliable benchmark for potato leaf disease classification and may support future research in agricultural image analysis.