Using Hybrid Pre-trained Convolutional Neural Networks and SVM Based VGG16, ResNet50, and DeseNet201 for Identifying Plant Leaf Disease

Abstract

Purpose: Early and accurate detection of plant leaf diseases is vital for safe-
guarding crop yield and supporting sustainable agricultural practices. However,
practical deployment faces challenges such as inconsistent lighting conditions,
overlapping leaves, low-contrast early-stage symptoms, and noisy image data—all
of which hinder the reliability of deep learning models in field environments.
This study aims to develop a robust, scalable, and interpretable classification
framework capable of performing effectively under such real-world conditions.
Method: We propose a hybrid classification pipeline that integrates deep feature
extraction from three pre-trained Convolutional Neural Networks (CNNs) based
VGG16, ResNet50, and DenseNet201—with a linear Support Vector Machine
(SVM) classifier. To enhance robustness to varying illumination, all images
are converted from RGB to HSV colour space, enabling chromatic features to
be isolated from brightness fluctuations. Features are extracted from the final
global pooling layers of each CNN, then concatenated to construct a unified
high-dimensional feature vector. This vector is passed to the SVM classifier for
binary classification (healthy vs. diseased). The system was trained and validated
using a publicly available dataset comprising 4,503 labelled images, balanced
between healthy and diseased samples. A comprehensive data augmentation
strategy—including rotation, flipping, and scaling—was employed to improve
generalisation and mitigate overfitting.
Results: Among the evaluated configurations, the DenseNet201 + SVM model
achieved the highest accuracy of 95%, outperforming both standalone CNN mod-
els and other hybrid variants including, VGG16-SVM, and ResNet50-SVM. The
hybrid approach demonstrated enhanced generalisability, particularly in images
affected by noise and lighting inconsistencies. Precision, recall, F1-score, and
confusion matrix metrics confirmed the model’s strong performance across both
classes.
Conclusion: The proposed hybrid CNN-SVM framework offers a robust and
interpretable solution for real-world leaf disease detection. By leveraging HSV
colour space transformation and combining diverse CNN feature representations,
the model effectively addresses common challenges in agricultural image classi-
fication. This work presents a scalable pipeline with potential for deployment
in precision agriculture systems, including smartphone-based or drone-assisted
monitoring platforms.