Comparisons of Machine Learning Versus Deep Learning Architectures for Plant Disease Classification

The identification and classification of plant leaf diseases is an important aspect of agricultural management. It significantly affects crop yield and food safety. As global agricultural practices evolve, the importance of timely and accurate identification of diseases has increased [1]. The development and use of machine learning (ML) and deep learning (DL) methods in technology have become widespread. It has opened new avenues for automating this process. It has offered the potential to increase the accuracy of disease detection [2]. This comparative study examines various ML and DL approaches for identifying and classifying plant leaf diseases using the New Plant Disease dataset (https://www.kaggle.com/datasets/vipoooool/new-plant-diseases-dataset). By examining this dataset, the strengths and weaknesses of different methodologies have been highlighted. To this end, insights into the effectiveness of each approach have been provided. In addition to the ML and DL methods applied in this study, data preprocessing and cluster number determination processes were performed on the dataset. The results of this study were evaluated through performance metrics and comparative analysis of the datasets. A comprehensive understanding of plant disease recognition and classification is presented. Important insights were identified in this research, challenges encountered were highlighted, and recommendations for future research were provided.

The importance of automated systems for the detection and classification of plant diseases is emphasized in order to increase agricultural productivity and ensure food security in the face of growing global demands [3].

In the study conducted by Gupta et al. [4], ML is applied to the Pear and Watermelon datasets, and 2,994 features were used for classification. These features were generated by adding the mean and standard deviation of each channel RGB, as well as Gray-Level Co-occurrence Matrix (GLCM) and Histogram of Oriented Gradients (HOG) features. The study achieved a success rate of 93.4%. Additionally, class imbalance is addressed using the Synthetic Minority Over-sampling Technique (SMOTE).

Another study [5] presents a new plant leaf identification approach that combines shape and texture features. The proposed Multiscale Triangle Descriptor (MTD) is used to characterize the leaf's shape information, while the Local Binary Pattern Histogram Fourier (LBP-HF) is used as the texture feature. Shape and texture features are combined with weighted distance measurement; the distance is used for shape, and the Chi-square distance is used for texture. The method is tested on the Flavia, Swedish, and MEW2012 datasets. It achieved classification accuracies of 99.16%, 98.48%, and 95.64%, respectively. The results are better than or comparable to existing modern methods.

While recent studies have successfully applied architectures such as EfficientNet, SimCLR, and ConvNeXt to plant disease datasets, these works primarily focus on maximizing raw accuracy, often overlooking the deployment constraints of agricultural edge devices. Furthermore, existing literature rarely quantifies the inherent statistical complexity of the datasets used for evaluation. To address these gaps, the specific contributions of this study are threefold: (1) proposing a hybrid feature-fusion framework that combines deep ResNet-50 representations with handcrafted descriptors for accurate yet lightweight classification; (2) providing a statistical characterization of dataset complexity using GMM with the AIC and BIC criteria; and (3) establishing a comparative benchmark that explicitly measures the trade-off between classification accuracy and model storage size. This study provides a practical engineering perspective on selecting suitable architectures for resource-constrained agricultural systems by evaluating both ML workflows that combine features derived from deep feature extraction with manually crafted features, as well as modern DL models.

Gil et al. [6] tested various DL architectures on two different datasets. The proposed ConvNeXt architecture achieved accuracy rates of 99.47% and 93.83% on the Plant Village and Plant Pathology datasets, respectively, following hyperparameter optimization using the grid search method. Galabuzi et al. [7] achieved a 99.65% accuracy using the EfficientNet architecture with the New Plant Disease image dataset. Another study using the New Plant Disease dataset is the FL-EfficientNet model. The proposed FL-EfficientNet model solved the sample imbalance problem in plant disease classification using Focal Loss and reduced overfitting using DropConnect. It demonstrated better performance than other comparison networks with an average accuracy rate of 99.72% and fast convergence time. These features provide a strong foundation for deploying the model on mobile devices [8, 9]. Apleni et al. [10] proposed an ensemble model for the dataset. A success rate of 97.00% is achieved by utilizing the features of three different CNN architectures. Another study using this dataset examines the effectiveness of self-supervised learning (SSL) using contrastive learning of visual representations. Unsupervised pre-training scenarios on unlabeled plant images were investigated. The results showed that the performance of SSL in plant disease detection is comparable to that of traditional supervised learning approaches [11]. While DL models typically require massive datasets, integrating targeted segmentation techniques can mitigate this dependency. For instance, a recent study demonstrated that employing FCM-based segmentation to isolate diseased regions allowed a custom CNN architecture to achieve a 98% classification accuracy on apple leaf diseases without relying on a large-scale training set [12].

Relative studies and their characteristics are summarized in Table 1.

Table 1. Studies using the New Plant Disease dataset and its features

The New Plant Disease image dataset (https://www.kaggle.com/datasets/vipoooool/new-plant-diseases-dataset) consists of 38 classes. This dataset is recreated using offline augmentation from the original dataset. It consists of about 87,000 RGB images of healthy and diseased plant leaves, divided into 38 different classes. To strictly prevent data leakage and ensure robust evaluation, the dataset split (70% training, 20% validation, and 10% test) was conducted at the original image level prior to the application of offline augmentation techniques. This protocol guarantees that augmented variants or near-duplicate samples of the same source leaf do not cross-contaminate the validation and test sets, thereby confirming that the high accuracy rates are a true reflection of the models' generalization capabilities rather than dataset memorization. The training set contains 61,502 images, the validation set contains 17,574 images, and the test set contains 8,791 images.

The dataset contains a new directory with 33 test images for prediction purposes. Additionally, all images are 256 × 256 × 3 in size. Healthy and diseased leaf images from different plant species in the New Plant Disease dataset are shown in Figure 1.

Figure 1. Healthy and diseased leaf images from different plant species in the New Plant Disease dataset

In this study, dataset complexity is approached through the statistical density and structural distribution of the data rather than simple visual background clutter. To statistically evaluate this distribution, Gaussian Mixture Models (GMMs) were utilized as an unsupervised density estimation tool. Instead of direct class separability, we determined the optimal number of components using the Akaike Information Criterion (AIC) [30] and Bayesian Information Criterion (BIC) [31], which serve as established model selection criteria. The criteria are defined as:

$A I C=2 k-2 \ln (L)$     (1)

$B I C=\ln (n) k-2 \ln (L)$     (2)

where, k is the number of parameters, n is the sample size, and L is the maximum value of the likelihood function. A steady decrease in these scores indicates an improved model fit to the underlying feature space without excessive penalization for complexity. As shown in Figure 2, the clear inflection point converging at 38 clusters aligns with the structural variance of the dataset, confirming that the latent feature space requires robust model architectures capable of handling high-dimensional representations.

For this purpose, one image is taken from each class and resized to 224 × 224. Subsequently, 38 images were added together (50176,3 each) to obtain a (1906688,3) dimensional matrix. To evaluate the structural distribution of the dataset, Gaussian Mixture Models (GMM) are employed as an unsupervised density estimation technique. The optimal number of Gaussian components is determined using the AIC and BIC scores, which strictly function as model-selection tools balancing goodness-of-fit against parameter complexity. The GMM is initialized using the k-means algorithm, with the number of components systematically tested across a range from 2 to 40. To ensure reproducibility and convergence stability, the model is configured with a full covariance matrix, a random seed of 123, and 10 random initializations (n_init = 10). The Expectation-Maximization (EM) optimization parameters included a convergence tolerance of 0.001, a regularization covariance of 1e-06, and a maximum of 100 iterations. It is important to note that AIC and BIC do not directly measure spatial class separability or simplicity; rather, a relative decline and subsequent stabilization of these scores indicate a suitable probabilistic representation of the underlying feature space density without excessive overfitting. It should be noted that this analysis is intentionally performed at the raw pixel-level as a lightweight, model-agnostic screening of the dataset’s global color-intensity distribution, rather than as a measure of disease-class separability. A more discriminative feature-space analysis — for example, applying the same GMM/AIC/BIC procedure to deep ResNet-50 embeddings — is left as a direction for future work. The breaks in the GMM created using AIC and BIC scores for the New Plant Disease image dataset are shown in Figure 2.

The New Plant Disease dataset has image complexity and computational complexity due to its big data characteristics and the large number of classes it contains. For these reasons, both ML and DL methods have been applied for object recognition in the image data.

The ML methods (i) XGBoost classifier [32], (ii) Decision Tree [33] and (iii) SVM [34] are utilized for the classification of the New Plant Disease image dataset.

DL models (iv) ResNet-152 [35], (v) YOLO11 [36], and (vi) DINOv2 [37] were trained for the classification of the New Plant Disease image dataset. Unlike traditional supervised learning, DINOv2 operates on a self-supervised distillation principle. It utilizes a student-teacher network architecture where the student network learns to predict the output of the teacher network. The teacher network is updated as an exponential moving average (EMA) of the student weights, ensuring stable training representations without the need for negative pairs used in contrastive learning.

The training hyperparameters of the models are described in the implementation details section. Performance metrics, including accuracy, precision, recall, and F1 score, were used to evaluate the effectiveness of each method across the datasets.

To provide a comprehensive evaluation across the 38-class problem and to strictly account for potential class imbalances, performance is reported utilizing macro-averaged metrics alongside weighted averages. Furthermore, to evaluate algorithmic stability, DL model training is repeated over 3 times, yielding mean ± standard deviation scores across key metrics. A detailed per-class precision and recall analysis, supported by the Confusion Matrix (see Figure 4), reveals that while overall accuracy is exceptionally high, the rare misclassifications primarily occur between visually homologous disease stages, such as the early necrotic lesions of 'Tomato Late Blight' and 'Tomato Leaf Mold'; for instance, only 1 of the 253 'Tomato Late Blight' samples was misclassified as 'Tomato Leaf Mold'.

Figure 4. Confusion matrix of ResNet-152+CBAM with seed 123 best checkpoint

While the XGBoost classifier achieved a respectable 91.59% accuracy using only structural and textural features, the integration of deep semantic representations (ResNet-50) paired with an SVM significantly elevated the diagnostic accuracy to 96.16%, highlighting the necessity of deep feature fusion for resolving visually ambiguous disease classes like Tomato Septoria Leaf Spot.

A detailed analysis conducted on a class-by-class basis reveals the limitations of a single-tree architecture in a high-dimensional feature space. Although the model demonstrated exceptional performance in visually distinct categories such as Corn Common Rust (F1 score: 94.30%) and Early Blight of Potato (F1 score: 87.61%), its predictive ability significantly decreased in classes with subtle and overlapping symptoms. In particular, the F1 scores for tomato Septoria leaf spot, pepper healthy, and apple scab dropped to 41.79%, 43.84%, and 45.07%, respectively. These key findings demonstrate that while a Decision Tree alone can capture the basic data distribution, it lacks the nonlinear classification capacity required for robust and large-scale plant pathology diagnosis. The optimized decision tree model is evaluated on an unseen test set (8,791 examples). As a weak learner on its own, the classifier achieved a moderate overall accuracy of 68.34%, while the macro-average and weighted-average F1 scores were 68.39% and 68.26%, respectively.

When evaluated on an unseen test set (8,791 examples), the optimized SVM model achieved an overall accuracy of 88.68%. The macro-average and weighted-average F1 scores were consistent at 88.62% and 88.63%, respectively. This represents a significant improvement over the baseline independent Decision Tree (68.34%) and demonstrates that morphological and textural features are better distinguished by hyperplanes that maximize the margin.

The Model, Accuracy, Weighted Average Precision, Weighted Average Recall, Weighted Average F1 Score, Macro Average Precision, Macro Average Recall and Macro Average F1 score values of ML and DL models in the classification of the New Plant Disease image dataset are provided in Table 4. Also, the best model’s confusion matrix is given in Figure 4.

While state-of-the-art DL architectures such as YOLO11 and ResNet-152+CBAM achieve substantially higher accuracy than the proposed hybrid model, this study does not claim equivalent performance. On the 8,791-image test set, the deep learning models reached 99.89% (YOLO11), 99.68% (ResNet-152+CBAM), and 99.13% (Dinov2), exceeding the hybrid model (96.16%) by approximately 3.0 to 4.7 percentage points, and they therefore provide clearly superior classification accuracy. However, these models operate with tens of millions of parameters (e.g., YOLO11x with over 28 million parameters) and require substantial GPU memory and computational resources. By contrast, the proposed hybrid PCA-SVM architecture requires only 8,072,928 trainable parameters and occupies a mere 68.96 MB of physical storage. The advantage of the proposed framework therefore lies not in matching the accuracy of these deep networks, but in its compact model design, its feature-fusion strategy and its markedly lower resource requirements, which indicate potential for deployment on resource-constrained agricultural systems, subject to future on-device benchmarking.

Table 4. Machine learning (ML) and deep learning (DL) model values for the classification of the New Plant Disease image dataset: Model, Accuracy, Weighted Average Precision, Weighted Average Recall, Weighted Average F1 Score, Macro Average Precision, Macro Average Recall and Macro Average F1 Score

Although manual feature extraction initially entailed a high computational cost (approximately 1.8 hours), reducing the dimensionality to 256 using PCA made the model extremely fast during the testing/inference phase and suitable for real-time applications.

While this study establishes a robust comparative baseline regarding feature quality and theoretical computational cost, future research is fundamentally required to validate these findings on physical agricultural embedded systems. Subsequent studies will focus on deploying these optimized models—particularly the YOLO11 architecture—on hardware accelerators such as NVIDIA Jetson edge nodes to rigorously benchmark real-time inference latency Frame Per Second (FPS), thermal behavior, and power consumption in dynamic field environments. External field-condition validation is acknowledged as an important next step and is planned for future work, as no suitable in-domain field dataset was available within the current revision scope.

The authors thank the editor and two of the anonymous referees for carefully reading the manuscript and making some valuable comments which had greatly improved the earlier draft of the manuscript. This work did not receive any specific grant from funding agencies in the public, commercial or not-for-profit sectors. During the preparation of this manuscript, the authors used a generative artificial intelligence (AI)-based language tool solely for language editing and readability improvement; the authors reviewed and edited the content and take full responsibility for it.