PlantClassiNet: A Dual-Modal Fine-Tuning Framework for CNN-Based Plant Disease Classification

Abstract

Although Convolutional Neural Networks (CNNs) have delivered state-of-the-art accuracy in plant disease classification, their deployment is still hindered by data scarcity, computational cost, and architectural heterogeneity. Transfer learning from large-scale pre-trained datasets alleviates these issues, yet generic feature extraction suffers from domain shift, while indiscriminate fine-tuning risks over-fitting and elevated training budgets. To address the identified limitations, PlantClassiNet is implemented as a unified framework. This framework facilitates systematic comparative analysis of six CNN architectures, AlexNet, ResNet50, InceptionV3, MobileNetV3Small, DenseNet121 and EfficientNetB0, across three publicly available datasets: PlantVillage, PlantLeaves and Eggplant. Two alternative fine-tuning approaches are proposed: Adaptive Fine-tuning (AdapFitu), which adaptively determines the depth of unfrozen layers, learning rates, and reinitializes selected layers, and a fixed-parameter baseline, which trains only the newly added classifier while keeping the convolutional backbone frozen and unfreezes a fixed number of network layers for retraining. Extensive experiments demonstrate that large models AlexNet, ResNet50, and Inceptionv3 achieve test accuracy exceeding 98.74% on the sizable PlantVillage dataset, whereas lightweight counterparts MobileNetV3Small, DenseNet121, and EfficientNetB0 achieve high accuracy of 99.79% ± 0.21% on the smaller Eggplant collection after fine-tuning.

1. Introduction

Plant diseases are caused by various pathogens, including fungi, bacteria, viruses, and nematodes, leading to significant yield losses and economic impacts. Traditional methods for plant disease detection, such as visual inspection and laboratory testing, are time-consuming and labor-intensive. With the advent of deep learning, computer vision techniques have shown great potential in automating and enhancing plant disease detection.

CNNs have been widely used to classify plant diseases based on leaf images [1,2,3,4,5,6,7,8,9,10]. However, training deep CNNs from scratch requires a large amount of labeled data and computational resources.

Transfer learning, which leverages pre-trained models on large-scale datasets, has emerged as an effective solution to address these challenges [11,12,13,14]. Data scarcity in real-world settings hinders direct large-scale model training. Meanwhile, domain shifts from lab environments like PlantVillage’s clean images to field conditions drastically reduce model robustness. Moreover, existing approaches frequently suffer from limitations such as regional bias in training data, insufficient generalization across diverse environmental conditions, and computational inefficiency for real-time deployment.

To resolve these challenges, fine-tuning pre-trained models has become a promising paradigm. By leveraging knowledge transferred from large-scale generic datasets, models can achieve high accuracy with limited labeled data while improving generalization [15,16,17,18]. However, existing studies fall short in providing a systematic analysis of critical factors, such as the selection between pre-trained architectures ranging from lightweight to large-scale, adaptation strategies applied layer by layer, and the effects of domain-specific augmentation techniques.

To address these issues, we designed a unified framework, PlantClassiNet, that compares six pre-trained models systematically. The fine-tuning parameter configuration either adopts the novel AdapFitu or utilizes fixed empirical parameters. This presented approach leverages the rich feature representations learned by pre-trained models while adapting them to the specific characteristics of plant disease datasets and enhances feature extraction and classification accuracy while maintaining computational efficiency.

The main contributions are summarized as follows:

(1)

Comprehensive Performance Benchmarking: An extensive comparative analysis of six CNN models demonstrates that transfer learning with fine-tuning, where the last ten layers are unfrozen and retrained for 100 epochs, achieves exceptional mean classification accuracies of 99.35%, 99.19%, and 88.48% on the three datasets, highlighting the impact of dataset characteristics and model architecture on feature extraction and classification precision.

(2)

Lightweight Model Efficiency: MobileNetV3Small, DenseNet121, and EfficientNetB0 achieve competitive accuracy with reduced computational complexity, significantly accelerating inference time and making them ideal for real-time deployment on edge devices in resource-constrained agricultural environments.

(3)

Practical Implications: This study provides actionable guidance for selecting optimal deep learning models based on dataset size, disease complexity, and hardware limitations, advancing automated plant disease diagnostics and demonstrating the transformative potential of deep learning in enhancing plant health monitoring and fostering sustainable agricultural practices through rapid, scalable, and precise disease detection.

The remainder of this paper is organized as follows. Section 2 describes related work. Section 3 describes the datasets and the proposed methods used in this work. Section 4 explains and discusses the experimental setup and results. Section 5 summarizes the conclusions and future work.

2. Related Work

In recent years, deep learning techniques have gained prominence in plant disease detection and classification. CNNs have emerged as a powerful tool for vision-based tasks, demonstrating the ability to discern crucial features from images efficiently. Several studies have proposed CNN-based architectures for plant disease detection, achieving high accuracy and generalization capabilities.

For example, Shewale and Daruwala designed a CNN architecture for nine diseases classification of tomato plants [4]. The designed model incorporated advanced pre-processing techniques and hyperparameter tuning, outperforming existing models in computational efficiency and plant disease classification accuracy, achieving a maximum accuracy of 99.81%. Khaki et al. proposed a hybrid CNN-RNN model that effectively predicts crop yields, achieving a validation RMSE of 11.48 bushels per acre for corn and 3.85 for soybean, outperforming traditional models like LASSO and random forest in accuracy [2]. Argüeso et al. investigated few-shot learning for plant leaf disease classification using Triplet loss with a CNN on the PlantVillage dataset [3]. The Triplet loss achieved 94.0% accuracy with the full dataset and 90.0% with 80 samples per class, outperforming Contrastive loss. Most classes had F-scores above 85%, with four exceeding 90%, highlighting FSL’s effectiveness in reducing data requirements for plant disease detection. Bedi et al. developed the PDSE-Lite framework using CNN for detecting plant diseases and estimating severity from digital leaf images [19]. Experimental results showed 98.35% accuracy in disease detection and significant performance in severity estimation, with a 99% confidence interval from t-test analysis.

Despite progress, early CNN models suffered from data hunger, high computational costs, and architectural design challenges, limiting their field applicability.

To mitigate resource limitations, Mohanty et al. demonstrated that transfer learning from ImageNet to PlantVillage on AlexNet achieved an overall accuracy of 99.35% [1]. However, field testing revealed a significant performance drop to 31.4% accuracy, highlighting critical challenges in deploying deep learning models for practical plant disease diagnosis applications. Bedi and Gole applied five widely used predefined CNN architectures, AlexNet, LeNet5, GoogLeNet, VGG16, and ResNet50, to assess their performance in identifying bacterial spot disease in peach plants [20]. As per their paper, the AlexNet CNN model outperformed other models with 98.5% testing accuracy. In another work, Atila et al. employed EfficientNet for plant leaf disease classification, eliminating manual feature extraction while achieving high accuracy [12]. On PlantVillage, when comparing EfficientNet-B4/B5 with state-of-the-art models, their transfer learning approach utilizing fully trainable layers achieved 99.91% accuracy on the original dataset and 99.97% accuracy on the augmented dataset, showcasing superior performance. This work highlights scalable architectures’ potential in agricultural diagnostics, establishing a key benchmark.

However, feature extraction using pre-trained models avoids overfitting on small datasets by freezing convolutional weights but introduces new limitations, including poor domain adaptability and performance ceilings. Fine-tuning pre-trained models significantly improves domain adaptation through layered parameter optimization.

Iftikhar et al. introduced an enhanced CNN model for plant disease detection, achieving 98.17% accuracy in identifying fungal diseases [17]. The model is fine-tuned through hyperparameter optimization, adjusting learning rates, dropout rates, and batch sizes to maximize performance. It evaluated ResNet50, DenseNet121, AlexNet, VGG16, and MobileNetV2 pre-trained models, ensuring effective integration into a user-friendly mobile application for real-time disease management. Too et al. fine-tuned DenseNet, Inception-v4, VGG16, and ResNet for plant disease classification by replacing their final layers and training with SGD [15]. Results show that ResNet and DenseNet outperform others, achieving over 90% classification accuracy within 10 training epochs, demonstrating transfer learning’s effectiveness for agricultural applications. Prashanthi et al. employed ETPLDNet and LEViT models integrating CNN-ViT hybrid architectures, achieving 99.02% accuracy on 38 leaf disease classes through fine-tuned hyperparameter optimization and Grad-CAM interpretability [18]. The proposed approach surpasses VGG16 with 87.99% accuracy and InceptionV3 with 91.84% accuracy, delivering both high precision and explainability for agricultural diagnostics. Nag et al. proposed a tomato leaf disease detection system fine-tuning five CNN architectures: AlexNet, ResNet50, SqueezeNet, VGG19, and DenseNet121 [16]. Leveraging transfer learning and data augmentation on a dataset of 10 disease classes, their framework achieved over 99% classification accuracy using ResNet50, VGG19, and DenseNet121. The study also introduced a real-time mobile application with multilingual support, demonstrating its viability for field deployment in precision agriculture.

Fine-tuning pre-trained models significantly improves domain adaptation through layered parameter optimization; however, it triggers challenges like catastrophic forgetting and overfitting on limited data, prompting innovations such as progressive unfreezing and tiered learning rates. Current research in this field exhibits notable limitations, as most studies either focus on evaluating a limited number of models or emphasize computational efficiency at the expense of accuracy. More importantly, existing approaches fail to provide a comprehensive assessment, particularly in terms of systematically comparing more than five distinct architectures, including both large-scale and lightweight models with fine-tuning strategies across multiple crop varieties.

In this study, the proposed PlantClassiNet fine-tuning framework conducts plant leaf disease classification experiments on six pre-trained models of varying scales using three datasets of different sizes, thereby facilitating a comprehensive evaluation of the experimental results.

3. Materials and Methods

This section introduces the data sources and methods of the study. It covers three core datasets, their preprocessing, and two fine-tuning techniques, followed by the proposed PlantClassiNet architecture.

3.1. Dataset Description

The datasets used for the experiments in this study are PlantVillage [21], PlantLeaves [22] and Eggplant [23], which are widely selected as benchmark datasets available for academic research in the field of plant disease detection.

PlantVillage is a popular public dataset for plant disease classification research. 54,305 images of plant leaves were collected, which had 38 distinct classes. For each class label, the crop–disease pair is given for the image of the plant leaf, as shown in

Table 1. Plantvillage dataset description.

Category	Train	Validation	Test	# Sample
Apple Apple scab	441	94	95	630
Apple Black rot	434	94	93	621
Apple Cedar apple rust	192	41	42	275
Apple Healthy	1151	246	248	1645
Blueberry Healthy	1051	225	226	1502
Cherry Healthy	597	128	129	854
Cherry Powdery mildew	736	157	159	1052
Corn Gray leaf spot	359	76	78	513
Corn Common rust	834	178	180	1192
Corn Healthy	813	174	175	1162
Corn Northern leaf blight	689	147	149	985
Grape Black rot	826	177	177	1180
Grape Black measles	968	207	208	1383
Grape Isariopsis leaf spot	753	161	162	1076
Grape Healthy	296	63	64	423
Orange Citrus greening	3854	826	827	5507
Peach Bacterial spot	1607	344	346	2297
Peach Healthy	251	54	55	360
Bell pepper Bacterial spot	697	149	151	997
Bell pepper Healthy	1034	221	223	1478
Potato Healthy	106	22	24	152
Potato Early Blight	700	150	150	1000
Potato Late Blight	700	150	150	1000
Raspberry Healthy	259	55	57	371
Soybean Healthy	3563	763	764	5090
Squash Powdery mildew	1284	275	276	1835
Strawberry Healthy	319	68	69	456
Strawberry Leaf scorch	776	166	167	1109
Tomato Bacterial spot	1488	319	320	2127
Tomato Early blight	700	150	150	1000
Tomato Healthy	1113	238	240	1591
Tomato Late blight	1336	286	287	1909
Tomato Leaf mold	666	142	144	952
Tomato Septorial leaf spot	1239	265	267	1771
Tomato Two spotted spider mite	1173	251	252	1676
Tomato Target spot	982	210	212	1404
Tomato Mosaic Virus	261	55	57	373
Tomato Yellow leaf curl virus	3749	803	805	5357
Total				54,305

. Throughout this paper, the symbol # denotes the number of samples in all tables.

The PlantLeaves dataset consists of 2277 images of healthy leaves and 2225 images of diseased leaves, totaling 4502 images across 22 classes. Among them, eight images are used for prediction, while the remaining images are divided into training, testing, and validation sets, as seen in

Table 2. PlantLeaves dataset description.

Category	Train	Validation	Test	# Sample
Alstonia Scholaris diseased (P2a)	244	5	5	254
Alstonia Scholaris healthy (P2b)	168	5	5	178
Arjun diseased (P1a)	222	5	5	232
Arjun healthy (P1b)	210	5	5	220
Bael diseased (P4b)	107	5	5	117
Basil healthy (P8)	137	5	5	147
Chinar diseased (P11b)	110	5	5	120
Chinar healthy (P11a)	93	5	5	103
Gauva diseased (P3b)	131	5	5	141
Gauva healthy (P3a)	267	5	5	277
Jamun diseased (P5b)	335	5	5	345
Jamun healthy (P5a)	268	5	5	278
Jatropha diseased (P6b)	114	5	5	124
Jatropha healthy (P6a)	123	5	5	133
Lemon diseased (P10b)	67	5	5	77
Lemon healthy (P10a)	149	5	5	159
Mango diseased (P0b)	255	5	5	265
Mango healthy (P0a)	159	5	5	169
Pomegranate diseased (P9b)	261	5	5	271
Pomegranate healthy (P9a)	277	5	5	287
Pongamia Pinnata diseased (P7b)	265	5	5	275
Pongamia Pinnata healthy (P7a)	312	5	5	322
Total				4494

.

The Eggplant dataset initially included 3116 original high-resolution images; 392 original images were collected to be augmented, such as shifting, flipping, zooming, shearing, brightness enhancement, and rotation, resulting in a total set of 3159 augmented images. As shown in

Table 3. Eggplant dataset description.

Category	Train	Validation	Test	# Sample
Healthy Leaf	375	80	81	536
Insect Pest Disease	536	115	116	767
Leaf Spot Disease	627	134	135	896
Mosaic Virus Disease	201	43	44	288
Small Leaf Disease	78	16	18	112
White Mold Disease	44	9	11	64
Wilt Disease	347	74	75	496
Total				3159

, there are seven distinct kinds of eggplant diseases, including Healthy Leaf, Insect Pest Disease, Leaf Spot Disease, Mosaic Virus Disease, Small Leaf Disease, White Mold Disease, and Wilt Disease.

Table 4. Dataset Accession Description.

PlantVillage
Download URL	https://www.kaggle.com/datasets/abdallahalidev/plantvillage-dataset accessed on 5 July 2025
Citation Format	[21]
License Constraints	CC BY-NC-SA 4.0
Local Curation	These details are discussed in Section 3.2.
PlantLeaves
Download URL	https://www.kaggle.com/datasets/csafrit2/plant-leaves-for-image-classification accessed on 5 July 2025
Citation Format	[22]
License Constraints	Community Data License Agreement-Sharing-Version 1.0
Local Curation	These details are discussed in Section 3.2.
Eggplant
Download URL	https://www.kaggle.com/datasets/kamalmoha/eggplant-disease-recognition-dataset/data accessed on 5 July 2025
Citation Format	[23]
License Constraints	CC BY 4.0
Local Curation	These details are discussed in Section 3.2.

presents a comprehensive overview of key dataset accessibility details, including valid download links, recommended citation formats, and applicable license terms. This information is critical for researchers seeking to replicate or extend the study, particularly when working with images that underwent localized curation procedures. Additionally, the table outlines the specific additional curation measures applied to the dataset, such as normalization, rotation, shear, and other relevant preprocessing steps. These processes are essential for maintaining data integrity and ensuring reliability in computer vision applications.

3.2. Data Partitioning and Pre-Processing

The original datasets in Table 1, Table 2 and Table 3 were divided into train, validation, and test datasets. The training dataset was employed to train the models. The validation dataset was used exclusively to select the best-trained model from different epochs. The best model was subsequently evaluated using the test dataset. Since Dataset Plantleave has already been pre-partitioned into training, validation, and test subsets, we have retained its original division without further partitioning. For the PlantVillage and Eggplant datasets, a fixed partitioning strategy was employed, with 70% allocated to training, 15% to validation, and 15% to testing, rather than utilizing stratified sampling. Extensive data augmentation was applied to the training set through techniques such as rotation, translation, cropping, scaling, and horizontal flipping. Only normalization preprocessing was performed on the validation and test sets. Class weights were determined based on the class distribution of the training set. To ensure reproducibility, a fixed random seed was used with shuffle = False in the data loader. Class-imbalanced data distributions lead to inadequate representation of minority categories during training, thereby impairing the model’s generalization performance. Class weight parameters were calculated solely based on the training set partition, with the validation set and test set excluded from these calculations, to prevent data leakage. To address this imbalance, higher penalties are assigned to underrepresented categories, amplifying their influence on the training objective. Class-specific weights were computed according to Equation (1); the i class weight was computed as follows: number of samples divided by the product of the number of classes and number of samples for class i.

$$\mathrm{weight}\_\mathsf{i}=\#\mathrm{sample}/(\#\mathrm{class}\times \#\mathrm{sample}\_\mathsf{i})$$

(1)

To achieve optimal performance, the train data is preprocessed to align its distribution with the pre-trained distribution. Three normalization schemes were implemented across different models, as shown in

Table 5. Preprocessing Description.

	AlexNet [24]	ResNet50 [25]	InceptionV3 [26]	DenseNet121 [27]	EfficientNetB0 [28]	MobileNetV3Small [29]
Input size	$227\times 227$	$224\times 224$	$299\times 299$	$224\times 224$	$224\times 224$	$224\times 224$
Normalization	x-mean	x-mean	(x/127.5) − 1.0	x/255.0	(x/127.5) − 1.0	(x/127.5) − 1.0
Scaled	no	no	[−1, 1]	[0, 1]	[−1, 1]	[−1, 1]
Rotation	20°	20°	20°	20°	20°	20°
Width shift	0.2	0.2	0.2	0.2	0.2	0.2
Height shift	0.2	0.2	0.2	0.2	0.2	0.2
Shear	0.2	0.2	0.2	0.2	0.2	0.2
Zoom	0.2	0.2	0.2	0.2	0.2	0.2
Horizontal flip	Yes	Yes	Yes	Yes	Yes	Yes

. The first method linearly scales pixel values to the [0, 1] range, maintaining relative intensity relationships while simplifying computation for DenseNet121. The second approach redistributes values symmetrically into the [−1, 1] interval, particularly beneficial for InceptionV3, EfficientNetB0 and MobileNetV3Small models. Lastly, zero-centering was applied to AlexNet and ResNet50 by subtracting the ImageNet dataset mean (123.68, 116.779, and 103.939), which helps convolutional networks by removing inherent bias from the input distribution. These techniques systematically enhance training stability, accelerate convergence, and ensure consistent feature scaling regardless of the model architecture being used.

To further improve model generalization, data augmentation techniques were applied. The used approaches included random rotations with a tolerance of 20°, alterations in width and height of 20%, shear transformations of 20%, zoom modifications of 20%, and horizontal flips. These augmentations changed the size and orientation, which improved the model’s generalization ability for future agricultural application.

3.3. Research Approach

The PlantClassiNet methodology comprises two sequential phases:

(1)

Initial Feature Adaptation: The convolutional backbone of the pre-trained model remains frozen, while its original classification layer is substituted with a task-specific fully connected layer aligned with the new category dimensionality. Training exclusively targets the top two dense layers during this stage.

(2)

Dual-Modal Fine-tuning: The process bifurcates into two configurations. It utilizes the novel AdapFitu operator for parameter optimization and employs predefined training hyperparameters. The architectural framework incorporates essential components: ReLU activation functions for non-linear transformations, Batch Normalization to enhance training stability and accelerate convergence, and Pooling layers for feature dimensionality reduction, as illustrated in Figure 1.

3.3.1. The Adaptive Finetune Optimizer

The presented AdapFitu was designed for computing the unfreeze layers and learning rate (Algorithm 1).

Algorithm 1 The proposed AdapFitu unfreeze strategy

1: $\mathrm{Input}\mathrm{Pre-train}\mathrm{model}$   2: $\mathrm{total}\_\mathrm{layers}=\mathrm{len}(\mathrm{base}\_\mathrm{model}.\mathrm{layers})$   3: $\mathrm{total}\_\mathrm{params}=\mathrm{base}\_\mathrm{model}.\mathrm{count}\_\mathrm{params}\left(\right)\times {10}^{-6}$   4: $\mathrm{unfreeze}\_\mathrm{ratio}=\min (0.15+(\mathrm{total}\_\mathrm{params}/1000)\times 0.1,\mathrm{self}.\max \_\mathrm{unfreeze}\_\mathrm{ratio})$   5: $\mathrm{unfreeze}\_\mathrm{count}=\max (\mathrm{self}.\min \_\mathrm{unfreeze},\lfloor \mathrm{total}\_\mathrm{layers}\times \mathrm{unfreeze}\_\mathrm{ratio}\rfloor )$   6: $if\mathrm{total}\_\mathrm{params}>100:$   7: $\mathrm{lr}=1\times {10}^{-4}$   8: $elif\mathrm{total}\_\mathrm{params}>50:$   9: $\mathrm{lr}=3\times {10}^{-4}$ 10: $else:$ 11: $\mathrm{lr}=1\times {10}^{-3}$ 12: $\mathrm{Output}\mathrm{unfreeze}\mathrm{layers}\mathrm{and}\mathrm{learning}\mathrm{rate}$

The AdapFitu algorithm establishes a quantifiable relationship between model scale (parameter count and number of layers) and fine-tuning parameters (number of frozen layers and learning rate) through mathematical modeling. Its design integrates theoretical analysis with engineering practice: based on gradient sensitivity to model complexity, it dynamically controls the number of frozen layers according to parameter count and depth, thereby preventing training instability in large models caused by excessive unfreezing while ensuring sufficient knowledge transfer in smaller models; grounded in optimization theory’s gradient scaling effects, it adaptively adjusts the learning rate based on parameter magnitude, employing smaller learning rates for larger models to stabilize optimization and higher learning rates for smaller models to accelerate convergence. Through theoretical derivation and experimental validation, it calibrates constants (e.g., 0.15) and thresholds (e.g., 100 M) in the formulation. This approach achieves scale-driven adaptive fine-tuning, striking an optimal balance between computational resources and model performance.

The unfreeze layers and learning rate was computed by the AdapFitu optimizer, as shown in

Table 6. The computed fine-tune parameter by AdapFitu.

Models	#Layer	#Para	#Conv_Layer	#Unfreeze	lr
AlexNet	8	60 M	5	5	0.001
ResNet50	176	25.6 M	53	18	0.0001
InceptionV3	312	23.6 M	94	14	0.001
DenseNet121	428	7.04 M	120	22	0.001
EfficientNetB0	239	4.05 M	65	18	0.0001
MobileNetV3Small	230	0.94 M	41	27	0.0001

Note: #layer: total layers; #para: parameters; #conv_layer: convolutional layers; #unfreeze: unfreeze layers; lr: learning rate.

. In the fine-tune stage, all the computed unfreeze layers and dense layers were re-trained for the downstream plant disease classification task.

3.3.2. Adaptive and Fixed Training Strategies

The proposed PlantClassiNet framework incorporates two distinct training modalities: adaptive-parameter training and fixed-parameter training. Experimental implementation follows a two-phase procedure. During the initial phase, the convolutional base of a pre-trained model remains unchanged while replacing its original classification layer with a fully connected layer corresponding to the dimensionality of new task categories, with training exclusively applied to the top two fully connected layers. The subsequent phase diverges into the specified modalities: parameter configuration either adopts the novel AdapFitu operator or utilizes the predefined fixed training parameters enumerated in

Table 7. Experimental Environment Specification.

Software and Hardware Configuration
GPU	NVIDIA GeForce RTX 3080
GPU Memory Driver	550.163.01
CPU	AMD EPYC 7601
CPU Configuration	32 cores
Video Memory	20G
RAM	63G
Hard disk	70G
Programming language	Python 3.10
Deep Learning Framework	Tensorflow 2.13.0

. For detailed experimental configurations, refer to Section 4.1.

3.3.3. Evaluation Metrics

In plant disease classification where the class distribution is imbalanced (e.g., rare disease detection scenarios), conventional accuracy metrics may create misleading performance impressions. The confusion matrix and its derived metrics (Precision, Recall, and F1) enable precise identification of performance disparities across classes, thereby serving as a critical foundation for targeted model optimization strategies.

The confusion matrix cross-references predicted outcomes with true labels to generate four key metrics: True Positive (TP) for correctly identified positive instances, True Negative (TN) for correctly identified negative instances, False Positive (FP) for negative instances misclassified as positive (Class1 Error), and False Negative (FN) for positive instances misclassified as negative (Class2 Error). This visualization tool is particularly valuable for identifying classification biases in imbalanced datasets.

The derivation process for these metrics has been intentionally omitted from this section, as it has been comprehensively addressed in the experimental section (Equations (2)–(4)), maintaining methodological consistency while avoiding redundancy.

4. Results and Discussion

4.1. Experimental Section

The experimental configurations, including software and hardware settings, training parameters, and evaluation metrics, are based on the general methods described in the Materials and Methods, with specific parameters detailed below.

4.1.1. Software and Hardware Specifications

The experimental environment used in this study is outlined below, with detailed specifications provided in Table 7. The hardware includes a server equipped with an AMD EPYC processor, sufficient RAM, a high-performance GPU, and ample SSD storage. The software environment consists of an Ubuntu operating system, Python, and key libraries for data analysis and deep learning.

4.1.2. Training Hyperparameter Configuration

In our experiments, the various values of hyperparameters were selected for different training steps and strategies, as shown in

Table 8. Two-stage training: fixed parameters.

First Stage:		Second Stage:
Freeze Convolutional Base		Unfreeze Top 10 Layers
epochs	10	epochs	100
batch size	32	batch size	32
learning rate	0.001	learning rate	0.0001
retraining layers	Dense	retraining layers	Dense + BN + top

. The pre-trained convolutional base values were used for the first train stage, and the two dense layers were retrained. In the fine-tuning stage, the freeze layers and learning rate were the top ten layers and 0.0001 or computed by the developed AdapFitu optimizer (Table 6). For general methodological details, see “Materials and Methods” in Section 3.3.

4.1.3. Evaluation Metrics

Performance metrics commonly used in plant disease detection studies include accuracy, precision, recall, and F1-score. Accuracy measures the proportion of correctly classified samples, while precision and recall provide insights into the model’s ability to correctly identify positive samples. The F1-score combines precision and recall into a single metric, offering a balanced measure of performance.

Precision represents the accuracy of positive predictions, which is calculated as shown in Equation (2).

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(2)

where TP is true positives and FP is false positives.

Recall explains the ability of a model to find all relevant cases and is calculated as shown in Equation (3), where FN is false negatives.

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(3)

The F1-score is the harmonic mean of the precision and recall, where an F1-score reaches its best value at 1 and worst score at 0. The relative contributions of precision and recall to the F1-score are equal. The formula for the F1-score is shown in Equation (4).

$$F1={\displaystyle \frac{2\times TP}{2\times TP+FP+FN}}$$

(4)

The reported averages consist of the macro average, which averages the unweighted mean per label, weighted average, which averages the support-weighted mean per label, and sample average, applicable only for multilabel classification. Micro average, which averages the total true positives, false negatives, and false positives, is displayed solely for multi-label or multi-class scenarios involving a subset of classes.

4.2. Result

The proposed PlantClassiNet architecture was fine-tuned and trained for six pre-trained models, evaluated in precision, recall and F1-score. Each model was trained using training and validation datasets and then used to make predictions on testing data, which was separated from the training and validation datasets. The six models’ weighted average values on the three datasets are presented in

Table 9. Weighted average for six pre-trained models on PlantVillage dataset.

Model	Layers	Params	Precision	Recall	F1-Score
AlexNet	8	60 M	0.9938	0.9938	0.9938
ResNet50	50	25.5 M	0.9946	0.9945	0.9945
InceptionV3	46	23.6 M	0.9876	0.9875	0.9874
DenseNet121	121	7.03 M	0.9968	0.9967	0.9967
EfficientNetB0	237	4.05 M	0.9950	0.9950	0.9950
MobileNetV3Small	54	2.5 M	0.9937	0.9936	0.9936

, Table 16 and Table 23. The experiment results of each class are presented in

Table 10. The proposed PlantClassiNet for AlexNet on the Plantvillage dataset.

Plantvillage Category	Precision	Recall	F1-Score	# Test Sample
Apple Apple scab	0.9892	0.9684	0.9787	95
Apple Black rot	0.9895	1.0000	0.9947	94
Apple Cedar apple rust	1.0000	1.0000	1.0000	42
Apple healthy	0.9919	0.9919	0.9919	247
Blueberry healthy	0.9956	1.0000	0.9978	226
Cherry Powdery mildew	1.0000	0.9937	0.9968	158
Cherry healthy	0.9847	1.0000	0.9923	129
Corn Gray leaf spot	0.9359	0.9481	0.9419	77
Corn Common rust	1.0000	1.0000	1.0000	179
Corn Northern Leaf Blight	0.9730	0.9730	0.9730	148
Corn healthy	1.0000	1.0000	1.0000	175
Grape Black rot	1.0000	0.9887	0.9943	177
Grape Black Measles	0.9905	1.0000	0.9952	208
Grape Isariopsis Leaf Spot	1.0000	1.0000	1.0000	162
Grape healthy	1.0000	1.0000	1.0000	64
Orange Citrus greening	1.0000	0.9988	0.9994	827
Peach Bacterial spot	1.0000	1.0000	1.0000	345
Peach healthy	1.0000	1.0000	1.0000	54
Pepper bell Bacterial spot	1.0000	0.9867	0.9933	150
Pepper bell healthy	1.0000	0.9955	0.9977	222
Potato Early blight	1.0000	1.0000	1.0000	150
Potato Late blight	0.9804	1.0000	0.9901	150
Potato healthy	0.9200	1.0000	0.9583	23
Raspberry healthy	1.0000	1.0000	1.0000	56
Soybean healthy	1.0000	0.9961	0.9980	764
Squash Powdery mildew	1.0000	1.0000	1.0000	276
Strawberry Leaf scorch	1.0000	0.9940	0.9970	167
Strawberry healthy	1.0000	1.0000	1.0000	69
Tomato Bacterial spot	0.9969	0.9906	0.9937	320
Tomato Early blight	0.9548	0.9867	0.9705	150
Tomato Late blight	0.9965	0.9791	0.9877	287
Tomato Leaf Mold	0.9929	0.9720	0.9823	143
Tomato Septoria leaf spot	0.9925	1.0000	0.9963	266
Tomato Two spotted spider mite	0.9729	0.9960	0.9843	252
Tomato Target Spot	0.9902	0.9621	0.9760	211
Tomato Yellow Leaf Curl Virus	0.9988	0.9963	0.9975	804
Tomato mosaic virus	0.9825	1.0000	0.9912	56
Tomato healthy	0.9795	1.0000	0.9896	239
Macro Avg	0.9897	0.9926	0.9910

,

Table 11. The proposed PlantClassiNet for ResNet50 on the Plantvillage dataset.

Plantvillage Category	Precision	Recall	F1-Score	# Test Sample
Apple Apple scab	0.9894	0.9789	0.9841	95
Apple Black rot	1.0000	1.0000	1.0000	94
Apple Cedar apple rust	1.0000	1.0000	1.0000	42
Apple healthy	0.9919	0.9960	0.9939	247
Blueberry healthy	1.0000	1.0000	1.0000	226
Cherry Powdery mildew	1.0000	0.9873	0.9936	158
Cherry healthy	1.0000	1.0000	1.0000	129
Corn Gray leaf spot	0.9710	0.8701	0.9178	77
Corn Common rust	1.0000	0.9944	0.9972	179
Corn Northern Leaf Blight	0.9355	0.9797	0.9571	148
Corn healthy	1.0000	1.0000	1.0000	175
Grape Black rot	1.0000	1.0000	1.0000	177
Grape Black Measles	1.0000	1.0000	1.0000	208
Grape Isariopsis Leaf Spot	1.0000	1.0000	1.0000	162
Grape healthy	1.0000	1.0000	1.0000	64
Orange Citrus greening	1.0000	1.0000	1.0000	827
Peach Bacterial spot	1.0000	0.9971	0.9985	345
Peach healthy	0.9815	0.9815	0.9815	54
Pepper bell Bacterial spot	1.0000	0.9933	0.9967	150
Pepper bell healthy	0.9955	1.0000	0.9978	222
Potato Early blight	0.9934	1.0000	0.9967	150
Potato Late blight	1.0000	0.9933	0.9967	150
Potato healthy	0.9565	0.9565	0.9565	23
Raspberry healthy	1.0000	1.0000	1.0000	56
Soybean healthy	0.9987	1.0000	0.9993	764
Squash Powdery mildew	0.9964	1.0000	0.9982	276
Strawberry Leaf scorch	1.0000	1.0000	1.0000	167
Strawberry healthy	1.0000	1.0000	1.0000	69
Tomato Bacterial spot	1.0000	0.9938	0.9969	320
Tomato Early blight	0.9930	0.9467	0.9693	150
Tomato Late blight	0.9862	0.9930	0.9896	287
Tomato Leaf Mold	1.0000	0.9790	0.9894	143
Tomato Septoria leaf spot	0.9888	1.0000	0.9944	266
Tomato Two spotted spider mite	0.9763	0.9802	0.9782	252
Tomato Target Spot	0.9635	1.0000	0.9814	211
Tomato Yellow Leaf Curl Virus	1.0000	0.9988	0.9994	804
Tomato mosaic virus	1.0000	1.0000	1.0000	56
Tomato healthy	0.9917	1.0000	0.9958	239
Macro Avg	0.9923	0.9900	0.9911

,

Table 12. The proposed PlantClassiNet for InceptionV3 on the Plantvillage dataset.

Plantvillage Category	Precision	Recall	F1-Score	# Test Sample
Apple Apple scab	0.9891	0.9579	0.9733	95
Apple Black rot	0.9895	1.0000	0.9947	94
Apple Cedar apple rust	1.0000	0.9762	0.9880	42
Apple healthy	0.9960	0.9960	0.9960	247
Blueberry healthy	0.9956	1.0000	0.9978	226
Cherry Powdery mildew	1.0000	0.9937	0.9968	158
Cherry healthy	1.0000	1.0000	1.0000	129
Corn Gray leaf spot	0.8795	0.9481	0.9125	77
Corn Common rust	1.0000	1.0000	1.0000	179
Corn Northern Leaf Blight	0.9718	0.9324	0.9517	148
Corn healthy	1.0000	1.0000	1.0000	175
Grape Black rot	0.9944	1.0000	0.9972	177
Grape Black Measles	1.0000	0.9952	0.9976	208
Grape Isariopsis Leaf Spot	1.0000	1.0000	1.0000	162
Grape healthy	0.9844	0.9844	0.9844	64
Orange Citrus greening	1.0000	1.0000	1.0000	827
Peach Bacterial spot	0.9914	0.9971	0.9942	345
Peach healthy	0.9815	0.9815	0.9815	54
Pepper bell Bacterial spot	0.9933	0.9933	0.9933	150
Pepper bell healthy	0.9867	1.0000	0.9933	222
Potato Early blight	0.9934	1.0000	0.9967	150
Potato Late blight	0.9551	0.9933	0.9739	150
Potato healthy	1.0000	0.7391	0.8500	23
Raspberry healthy	1.0000	1.0000	1.0000	56
Soybean healthy	0.9948	0.9987	0.9967	764
Squash Powdery mildew	1.0000	0.9964	0.9982	276
Strawberry Leaf scorch	1.0000	0.9880	0.9940	167
Strawberry healthy	1.0000	1.0000	1.0000	69
Tomato Bacterial spot	0.9812	0.9781	0.9797	320
Tomato Early blight	0.9371	0.8933	0.9147	150
Tomato Late blight	0.9682	0.9547	0.9614	287
Tomato Leaf Mold	0.9720	0.9720	0.9720	143
Tomato Septoria leaf spot	0.9562	0.9850	0.9704	266
Tomato Two spotted spider mite	0.9760	0.9683	0.9721	252
Tomato Target Spot	0.9543	0.9905	0.9721	211
Tomato Yellow Leaf Curl Virus	0.9975	0.9876	0.9925	804
Tomato mosaic virus	0.9492	1.0000	0.9739	56
Tomato healthy	0.9958	0.9958	0.9958	239
Macro Avg	0.9838	0.9789	0.9807

,

Table 13. The proposed PlantClassiNet for DenseNet121 on the Plantvillage dataset.

Plantvillage Category	Precision	Recall	F1-Score	# Test Sample
Apple Apple scab	1.0000	0.9895	0.9947	95
Apple Black rot	0.9895	1.0000	0.9947	94
Apple Cedar apple rust	1.0000	1.0000	1.0000	42
Apple healthy	1.0000	1.0000	1.0000	247
Blueberry healthy	1.0000	1.0000	1.0000	226
Cherry Powdery mildew	1.0000	1.0000	1.0000	158
Cherry healthy	1.0000	1.0000	1.0000	129
Corn Gray leaf spot	0.8837	0.9870	0.9325	77
Corn Common rust	1.0000	1.0000	1.0000	179
Corn Northern Leaf Blight	0.9929	0.9392	0.9653	148
Corn healthy	1.0000	1.0000	1.0000	175
Grape Black rot	0.9944	1.0000	0.9972	177
Grape Black Measles	1.0000	1.0000	1.0000	208
Grape Isariopsis Leaf Spot	1.0000	0.9938	0.9969	162
Grape healthy	1.0000	1.0000	1.0000	64
Orange Citrus greening	1.0000	0.9988	0.9994	827
Peach Bacterial spot	0.9942	1.0000	0.9971	345
Peach healthy	1.0000	1.0000	1.0000	54
Pepper bell Bacterial spot	1.0000	0.9867	0.9933	150
Pepper bell healthy	0.9955	1.0000	0.9978	222
Potato Early blight	0.9934	1.0000	0.9967	150
Potato Late blight	1.0000	0.9933	0.9967	150
Potato healthy	1.0000	0.9565	0.9778	23
Raspberry healthy	1.0000	1.0000	1.0000	56
Soybean healthy	0.9987	1.0000	0.9993	764
Squash Powdery mildew	1.0000	1.0000	1.0000	276
Strawberry Leaf scorch	1.0000	0.9940	0.9970	167
Strawberry healthy	1.0000	1.0000	1.0000	69
Tomato Bacterial spot	0.9938	1.0000	0.9969	320
Tomato Early blight	0.9868	0.9933	0.9900	150
Tomato Late blight	0.9965	0.9930	0.9948	287
Tomato Leaf Mold	1.0000	0.9860	0.9930	143
Tomato Septoria leaf spot	0.9963	1.0000	0.9981	266
Tomato Two spotted spider mite	1.0000	0.9960	0.9980	252
Tomato Target Spot	0.9906	0.9953	0.9929	211
Tomato Yellow Leaf Curl Virus	1.0000	0.9975	0.9988	804
Tomato mosaic virus	1.0000	1.0000	1.0000	56
Tomato healthy	0.9958	1.0000	0.9979	239
Macro Avg	0.9948	0.9947	0.9946

,

Table 14. The proposed PlantClassiNet for EfficientNetB0 on the Plantvillage dataset.

Plantvillage Category	Precision	Recall	F1-Score	# Test Sample
Apple Apple scab	1.0000	0.9789	0.9894	95
Apple Black rot	1.0000	1.0000	1.0000	94
Apple Cedar apple rust	1.0000	1.0000	1.0000	42
Apple healthy	1.0000	0.9960	0.9980	247
Blueberry healthy	0.9956	1.0000	0.9978	226
Cherry Powdery mildew	1.0000	1.0000	1.0000	158
Cherry healthy	1.0000	1.0000	1.0000	129
Corn Gray leaf spot	0.9012	0.9481	0.9241	77
Corn Common rust	1.0000	1.0000	1.0000	179
Corn Northern Leaf Blight	0.9720	0.9392	0.9553	148
Corn healthy	1.0000	1.0000	1.0000	175
Grape Black rot	1.0000	1.0000	1.0000	177
Grape Black Measles	1.0000	1.0000	1.0000	208
Grape Isariopsis Leaf Spot	1.0000	1.0000	1.0000	162
Grape healthy	1.0000	1.0000	1.0000	64
Orange Citrus greening	1.0000	1.0000	1.0000	827
Peach Bacterial spot	0.9914	0.9971	0.9942	345
Peach healthy	0.9643	1.0000	0.9818	54
Pepper bell Bacterial spot	1.0000	1.0000	1.0000	150
Pepper bell healthy	0.9911	1.0000	0.9955	222
Potato Early blight	0.9934	1.0000	0.9967	150
Potato Late blight	0.9868	1.0000	0.9934	150
Potato healthy	1.0000	0.8261	0.9048	23
Raspberry healthy	1.0000	1.0000	1.0000	56
Soybean healthy	1.0000	0.9987	0.9993	764
Squash Powdery mildew	1.0000	1.0000	1.0000	276
Strawberry Leaf scorch	1.0000	0.9940	0.9970	167
Strawberry healthy	1.0000	1.0000	1.0000	69
Tomato Bacterial spot	1.0000	0.9969	0.9984	320
Tomato Early blight	1.0000	0.9667	0.9831	150
Tomato Late blight	0.9828	0.9930	0.9879	287
Tomato Leaf Mold	0.9861	0.9930	0.9895	143
Tomato Septoria leaf spot	0.9925	0.9925	0.9925	266
Tomato Two spotted spider mite	0.9920	0.9802	0.9860	252
Tomato Target Spot	0.9769	1.0000	0.9883	211
Tomato Yellow Leaf Curl Virus	1.0000	0.9975	0.9988	804
Tomato mosaic virus	1.0000	1.0000	1.0000	56
Tomato healthy	0.9917	1.0000	0.9958	239
Macro Avg	0.9926	0.9894	0.9907

,

Table 15. The proposed PlantClassiNet for MobileNetV3Small on the Plantvillage dataset.

Plantvillage Category	Precision	Recall	F1-Score	# Test Sample
Apple Apple scab	0.9895	0.9895	0.9895	95
Apple Black rot	0.9895	1.0000	0.9947	94
Apple Cedar apple rust	1.0000	1.0000	1.0000	42
Apple healthy	1.0000	0.9919	0.9959	247
Blueberry healthy	1.0000	1.0000	1.0000	226
Cherry Powdery mildew	1.0000	0.9937	0.9968	158
Cherry healthy	1.0000	1.0000	1.0000	129
Corn Gray leaf spot	0.9241	0.9481	0.9359	77
Corn Common rust	1.0000	1.0000	1.0000	179
Corn Northern Leaf Blight	0.9724	0.9527	0.9625	148
Corn healthy	1.0000	0.9943	0.9971	175
Grape Black rot	1.0000	0.9944	0.9972	177
Grape Black Measles	0.9905	1.0000	0.9952	208
Grape Isariopsis Leaf Spot	1.0000	1.0000	1.0000	162
Grape healthy	1.0000	1.0000	1.0000	64
Orange Citrus greening	0.9988	1.0000	0.9994	827
Peach Bacterial spot	0.9971	1.0000	0.9986	345
Peach healthy	1.0000	0.9815	0.9907	54
Pepper bell Bacterial spot	1.0000	1.0000	1.0000	150
Pepper bell healthy	0.9955	1.0000	0.9978	222
Potato Early blight	1.0000	1.0000	1.0000	150
Potato Late blight	1.0000	0.9933	0.9967	150
Potato healthy	1.0000	0.9565	0.9778	23
Raspberry healthy	1.0000	1.0000	1.0000	56
Soybean healthy	0.9987	1.0000	0.9993	764
Squash Powdery mildew	1.0000	1.0000	1.0000	276
Strawberry Leaf scorch	1.0000	0.9940	0.9970	167
Strawberry healthy	1.0000	0.9855	0.9927	69
Tomato Bacterial spot	0.9968	0.9875	0.9922	320
Tomato Early blight	0.9796	0.9600	0.9697	150
Tomato Late blight	0.9792	0.9861	0.9826	287
Tomato Leaf Mold	0.9929	0.9720	0.9823	143
Tomato Septoria leaf spot	0.9638	1.0000	0.9815	266
Tomato Two spotted spider mite	0.9842	0.9881	0.9861	252
Tomato Target Spot	0.9631	0.9905	0.9766	211
Tomato Yellow Leaf Curl Virus	1.0000	0.9925	0.9963	804
Tomato mosaic virus	1.0000	0.9821	0.9910	56
Tomato healthy	0.9958	1.0000	0.9979	239
Macro Avg	0.9924	0.9904	0.9913

,

Table 16. Weighted average for six pre-trained models on PlantLeaves dataset.

Model	Layers	Params	Precision	Recall	F1-Score
AlexNet	8	60 M	0.8715	0.8545	0.8277
ResNet50	50	25.6 M	0.9513	0.9364	0.9346
InceptionV3	46	23.6 M	0.9133	0.8727	0.8654
DenseNet121	121	7.03 M	0.9165	0.9000	0.8964
EfficientNetB0	237	4.05 M	0.8715	0.8727	0.8568
MobileNetV3Small	54	2.5 M	0.8932	0.8727	0.8632

,

Table 17. The proposed PlantClassiNet for AlexNet on the PlantLeaves dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Alstonia Scholaris diseased (P2a)	0.8333	1.0000	0.9091	5
Alstonia Scholaris healthy (P2b)	1.0000	1.0000	1.0000	5
Arjun diseased (P1a)	1.0000	1.0000	1.0000	5
Arjun healthy (P1b)	1.0000	1.0000	1.0000	5
Bael diseased (P4b)	1.0000	1.0000	1.0000	5
Basil healthy (P8)	1.0000	1.0000	1.0000	5
Chinar diseased (P11b)	1.0000	1.0000	1.0000	5
Chinar healthy (P11a)	1.0000	1.0000	1.0000	5
Gauva diseased (P3b)	1.0000	0.8000	0.8889	5
Gauva healthy (P3a)	0.7143	1.0000	0.8333	5
Jamun diseased (P5b)	0.5000	1.0000	0.6667	5
Jamun healthy (P5a)	1.0000	0.4000	0.5714	5
Jatropha diseased (P6b)	1.0000	0.2000	0.3333	5
Jatropha healthy (P6a)	0.5000	1.0000	0.6667	5
Lemon diseased (P10b)	1.0000	0.4000	0.5714	5
Lemon healthy (P10a)	0.6250	1.0000	0.7692	5
Mango diseased (P0b)	1.0000	1.0000	1.0000	5
Mango healthy (P0a)	1.0000	1.0000	1.0000	5
Pomegranate diseased (P9b)	1.0000	1.0000	1.0000	5
Pomegranate healthy (P9a)	1.0000	1.0000	1.0000	5
Pongamia Pinnata diseased (P7b)	0.0000	0.0000	0.0000	5
Pongamia Pinnata healthy (P7a)	1.0000	1.0000	1.0000	5
Macro Avg	0.8715	0.8545	0.8277

,

Table 18. The proposed PlantClassiNet for ResNet50 on the PlantLeaves dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Alstonia Scholaris diseased (P2a)	0.8333	1.0000	0.9091	5
Alstonia Scholaris healthy (P2b)	1.0000	0.8000	0.8889	5
Arjun diseased (P1a)	1.0000	0.8000	0.8889	5
Arjun healthy (P1b)	0.8333	1.0000	0.9091	5
Bael diseased (P4b)	1.0000	1.0000	1.0000	5
Basil healthy (P8)	1.0000	1.0000	1.0000	5
Chinar diseased (P11b)	1.0000	1.0000	1.0000	5
Chinar healthy (P11a)	1.0000	1.0000	1.0000	5
Gauva diseased (P3b)	1.0000	1.0000	1.0000	5
Gauva healthy (P3a)	1.0000	1.0000	1.0000	5
Jamun diseased (P5b)	0.7143	1.0000	0.8333	5
Jamun healthy (P5a)	1.0000	0.6000	0.7500	5
Jatropha diseased (P6b)	1.0000	0.8000	0.8889	5
Jatropha healthy (P6a)	0.8333	1.0000	0.9091	5
Lemon diseased (P10b)	1.0000	0.6000	0.7500	5
Lemon healthy (P10a)	0.7143	1.0000	0.8333	5
Mango diseased (P0b)	1.0000	1.0000	1.0000	5
Mango healthy (P0a)	1.0000	1.0000	1.0000	5
Pomegranate diseased (P9b)	1.0000	1.0000	1.0000	5
Pomegranate healthy (P9a)	1.0000	1.0000	1.0000	5
Pongamia Pinnata diseased (P7b)	1.0000	1.0000	1.0000	5
Pongamia Pinnata healthy (P7a)	1.0000	1.0000	1.0000	5
Macro Avg	0.9513	0.9364	0.9346

,

Table 19. The proposed PlantClassiNet for InceptionV3 on PlantLeaves dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Alstonia Scholaris diseased (P2a)	1.0000	1.0000	1.0000	5
Alstonia Scholaris healthy (P2b)	0.8333	1.0000	0.9091	5
Arjun diseased (P1a)	0.8000	0.8000	0.8000	5
Arjun healthy (P1b)	0.6667	0.8000	0.7273	5
Bael diseased (P4b)	1.0000	1.0000	1.0000	5
Basil healthy (P8)	1.0000	1.0000	1.0000	5
Chinar diseased (P11b)	1.0000	1.0000	1.0000	5
Chinar healthy (P11a)	1.0000	1.0000	1.0000	5
Gauva diseased (P3b)	1.0000	1.0000	1.0000	5
Gauva healthy (P3a)	1.0000	0.8000	0.8889	5
Jamun diseased (P5b)	0.5000	1.0000	0.6667	5
Jamun healthy (P5a)	1.0000	0.2000	0.3333	5
Jatropha diseased (P6b)	1.0000	0.8000	0.8889	5
Jatropha healthy (P6a)	1.0000	1.0000	1.0000	5
Lemon diseased (P10b)	1.0000	0.4000	0.5714	5
Lemon healthy (P10a)	0.6250	1.0000	0.7692	5
Mango diseased (P0b)	1.0000	0.8000	0.8889	5
Mango healthy (P0a)	0.8333	1.0000	0.9091	5
Pomegranate diseased (P9b)	1.0000	0.8000	0.8889	5
Pomegranate healthy (P9a)	1.0000	1.0000	1.0000	5
Pongamia Pinnata diseased (P7b)	1.0000	0.8000	0.8889	5
Pongamia Pinnata healthy (P7a)	0.8333	1.0000	0.9091	5
Macro Avg	0.9133	0.8727	0.8654

,

Table 20. The proposed PlantClassiNet for DenseNet121 on the PlantLeaves dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Alstonia Scholaris diseased (P2a)	1.0000	0.8000	0.8889	5
Alstonia Scholaris healthy (P2b)	1.0000	1.0000	1.0000	5
Arjun diseased (P1a)	0.6667	0.8000	0.7273	5
Arjun healthy (P1b)	0.8000	0.8000	0.8000	5
Bael diseased (P4b)	1.0000	1.0000	1.0000	5
Basil healthy (P8)	1.0000	1.0000	1.0000	5
Chinar diseased (P11b)	1.0000	1.0000	1.0000	5
Chinar healthy (P11a)	1.0000	1.0000	1.0000	5
Gauva diseased (P3b)	1.0000	0.8000	0.8889	5
Gauva healthy (P3a)	1.0000	1.0000	1.0000	5
Jamun diseased (P5b)	0.5714	0.8000	0.6667	5
Jamun healthy (P5a)	0.6667	0.4000	0.5000	5
Jatropha diseased (P6b)	1.0000	1.0000	1.0000	5
Jatropha healthy (P6a)	1.0000	1.0000	1.0000	5
Lemon diseased (P10b)	1.0000	0.4000	0.5714	5
Lemon healthy (P10a)	0.6250	1.0000	0.7692	5
Mango diseased (P0b)	0.8333	1.0000	0.9091	5
Mango healthy (P0a)	1.0000	1.0000	1.0000	5
Pomegranate diseased (P9b)	1.0000	1.0000	1.0000	5
Pomegranate healthy (P9a)	1.0000	1.0000	1.0000	5
Pongamia Pinnata diseased (P7b)	1.0000	1.0000	1.0000	5
Pongamia Pinnata healthy (P7a)	1.0000	1.0000	1.0000	5
Macro Avg	0.9165	0.9000	0.8964

,

Table 21. The proposed PlantClassiNet for EfficientNetB0 on the PlantLeaves dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Alstonia Scholaris diseased (P2a)	1.0000	1.0000	1.0000	5
Alstonia Scholaris healthy (P2b)	0.8333	1.0000	0.9091	5
Arjun diseased (P1a)	1.0000	0.8000	0.8889	5
Arjun healthy (P1b)	0.8333	1.0000	0.9091	5
Bael diseased (P4b)	1.0000	0.8000	0.8889	5
Basil healthy (P8)	1.0000	1.0000	1.0000	5
Chinar diseased (P11b)	0.8333	1.0000	0.9091	5
Chinar healthy (P11a)	1.0000	0.8000	0.8889	5
Gauva diseased (P3b)	1.0000	0.6000	0.7500	5
Gauva healthy (P3a)	0.8333	1.0000	0.9091	5
Jamun diseased (P5b)	0.6250	1.0000	0.7692	5
Jamun healthy (P5a)	1.0000	0.6000	0.7500	5
Jatropha diseased (P6b)	0.7143	1.0000	0.8333	5
Jatropha healthy (P6a)	1.0000	0.8000	0.8889	5
Lemon diseased (P10b)	0.0000	0.0000	0.0000	5
Lemon healthy (P10a)	0.5000	1.0000	0.6667	5
Mango diseased (P0b)	1.0000	1.0000	1.0000	5
Mango healthy (P0a)	1.0000	1.0000	1.0000	5
Pomegranate diseased (P9b)	1.0000	1.0000	1.0000	5
Pomegranate healthy (P9a)	1.0000	0.8000	0.8889	5
Pongamia Pinnata diseased (P7b)	1.0000	1.0000	1.0000	5
Pongamia Pinnata healthy (P7a)	1.0000	1.0000	1.0000	5
Macro Avg	0.8715	0.8727	0.8568

,

Table 22. The proposed PlantClassiNet for MobileNetV3Small on the PlantLeaves dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Alstonia Scholaris diseased (P2a)	1.0000	1.0000	1.0000	5
Alstonia Scholaris healthy (P2b)	0.8333	1.0000	0.9091	5
Arjun diseased (P1a)	1.0000	0.8000	0.8889	5
Arjun healthy (P1b)	0.8333	1.0000	0.9091	5
Bael diseased (P4b)	1.0000	1.0000	1.0000	5
Basil healthy (P8)	1.0000	1.0000	1.0000	5
Chinar diseased (P11b)	1.0000	1.0000	1.0000	5
Chinar healthy (P11a)	1.0000	1.0000	1.0000	5
Gauva diseased (P3b)	1.0000	0.8000	0.8889	5
Gauva healthy (P3a)	1.0000	1.0000	1.0000	5
Jamun diseased (P5b)	0.4286	0.6000	0.5000	5
Jamun healthy (P5a)	0.3333	0.2000	0.2500	5
Jatropha diseased (P6b)	1.0000	0.8000	0.8889	5
Jatropha healthy (P6a)	0.8333	1.0000	0.9091	5
Lemon diseased (P10b)	1.0000	0.2000	0.3333	5
Lemon healthy (P10a)	0.5556	1.0000	0.7143	5
Mango diseased (P0b)	0.8333	1.0000	0.9091	5
Mango healthy (P0a)	1.0000	1.0000	1.0000	5
Pomegranate diseased (P9b)	1.0000	1.0000	1.0000	5
Pomegranate healthy (P9a)	1.0000	0.8000	0.8889	5
Pongamia Pinnata diseased (P7b)	1.0000	1.0000	1.0000	5
Pongamia Pinnata healthy (P7a)	1.0000	1.0000	1.0000	5
Macro Avg	0.8932	0.8727	0.8632

,

Table 23. Weighted average for six pre-trained models on Eggplant dataset.

Model	Layers	Params	Precision	Recall	F1-Score
AlexNet	8	60 M	0.9878	0.9874	0.9874
ResNet50	50	25.6 M	0.9939	0.9937	0.9937
InceptionV3	46	23.6 M	0.9779	0.9770	0.9771
DenseNet121	121	7.03 M	1.0000	1.0000	1.0000
EfficientNetB0	237	4.05 M	0.9979	0.9979	0.9979
MobileNetV3Small	54	2.5 M	0.9959	0.9958	0.9958

,

Table 24. The proposed PlantClassiNet for AlexNet on the Eggplant dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Healthy Leaf	0.9759	1.0000	0.9878	81
Insect Pest Disease	0.9667	1.0000	0.9831	116
Leaf Spot Disease	1.0000	0.9556	0.9773	135
Mosaic Virus Disease	1.0000	1.0000	1.0000	44
Small Leaf Disease	1.0000	1.0000	1.0000	17
White Mold Disease	1.0000	1.0000	1.0000	10
Wilt Disease	1.0000	1.0000	1.0000	75
Macro Avg	0.9918	0.9937	0.9926

,

Table 25. The proposed PlantClassiNet for ResNet50 on the Eggplant dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Healthy Leaf	1.0000	1.0000	1.0000	81
Insect Pest Disease	0.9748	1.0000	0.9872	116
Leaf Spot Disease	1.0000	0.9852	0.9925	135
Mosaic Virus Disease	1.0000	1.0000	1.0000	44
Small Leaf Disease	1.0000	1.0000	1.0000	17
White Mold Disease	1.0000	1.0000	1.0000	10
Wilt Disease	1.0000	0.9867	0.9933	75
Macro Avg	0.9964	0.9960	0.9962

,

Table 26. The proposed PlantClassiNet for InceptionV3 on the Eggplant dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Healthy Leaf	0.9878	1.0000	0.9939	81
Insect Pest Disease	1.0000	1.0000	1.0000	116
Leaf Spot Disease	1.0000	0.9926	0.9963	135
Mosaic Virus Disease	1.0000	1.0000	1.0000	44
Small Leaf Disease	1.0000	1.0000	1.0000	17
White Mold Disease	1.0000	1.0000	1.0000	10
Wilt Disease	1.0000	1.0000	1.0000	75
Macro Avg	0.9983	0.9989	0.9986

,

Table 27. The proposed PlantClassiNet for DenseNet121 on the Eggplant dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Healthy Leaf	1.0000	1.0000	1.0000	81
Insect Pest Disease	1.0000	1.0000	1.0000	116
Leaf Spot Disease	1.0000	1.0000	1.0000	135
Mosaic Virus Disease	1.0000	1.0000	1.0000	44
Small Leaf Disease	1.0000	1.0000	1.0000	17
White Mold Disease	1.0000	1.0000	1.0000	10
Wilt Disease	1.0000	1.0000	1.0000	75
Macro Avg	1.0000	1.0000	1.0000

,

Table 28. The proposed PlantClassiNet for EfficientNetB0 on the Eggplant dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Healthy Leaf	0.9195	0.9877	0.9524	81
Insect Pest Disease	0.9826	0.9741	0.9784	116
Leaf Spot Disease	0.9923	0.9556	0.9736	135
Mosaic Virus Disease	1.0000	1.0000	1.0000	44
Small Leaf Disease	1.0000	1.0000	1.0000	17
White Mold Disease	1.0000	0.9000	0.9474	10
Wilt Disease	0.9868	1.0000	0.9934	75
Macro Avg	0.9830	0.9739	0.9779

and

Table 29. The proposed PlantClassiNet for MobileNetV3Small on the Eggplant dataset.

PlantLeaves Category	Precision	Recall	F1-Score	# Test Sample
Healthy Leaf	0.9759	1.0000	0.9878	81
Insect Pest Disease	1.0000	1.0000	1.0000	116
Leaf Spot Disease	1.0000	0.9852	0.9925	135
Mosaic Virus Disease	1.0000	1.0000	1.0000	44
Small Leaf Disease	1.0000	1.0000	1.0000	17
White Mold Disease	1.0000	1.0000	1.0000	10
Wilt Disease	1.0000	1.0000	1.0000	75
Macro Avg	0.9966	0.9979	0.9972

, excluding Table 16 and Table 23. The confusion matrices of each model on each dataset are illustrated in the even-numbered figures from Figure 2, Figure 3, Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9, Figure 10, Figure 11, Figure 12, Figure 13, Figure 14, Figure 15, Figure 16, Figure 17, Figure 18, Figure 19, Figure 20, Figure 21, Figure 22, Figure 23, Figure 24, Figure 25, Figure 26, Figure 27, Figure 28, Figure 29, Figure 30, Figure 31, Figure 32, Figure 33, Figure 34, Figure 35 and Figure 36. The train accuracy and loss for the six pre-trained models and three datasets are demonstrated in the odd-numbered figures from Figure 3, Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9, Figure 10, Figure 11, Figure 12, Figure 13, Figure 14, Figure 15, Figure 16, Figure 17, Figure 18, Figure 19, Figure 20, Figure 21, Figure 22, Figure 23, Figure 24, Figure 25, Figure 26, Figure 27, Figure 28, Figure 29, Figure 30, Figure 31, Figure 32, Figure 33, Figure 34, Figure 35, Figure 36 and Figure 37. Overall, the fine-tuned DenseNet121 outperformed other baseline models on all of the datasets for downstream plant disease classification.

4.2.1. PlantVillage Dataset

The experimental results of all models on PlantVillage Dataset are presented. Table 9 reports the weighted average performance of all pre-trained models under the proposed PlantClassiNet across all categories of PlantVillage. Among these, the DenseNet121 base model achieves the highest performance, with Precision: 0.9968, Recall: 0.9967, and F1-score: 0.9967.

For each model, the results includes two figures: a confusion matrix and training accuracy/loss curves, along with one table detailing Precision, Recall, and F1-score per class. These visuals are systematically organized in Figure 2, Figure 3, Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9, Figure 10, Figure 11, Figure 12 and Figure 13 and Table 10, Table 11, Table 12, Table 13, Table 14 and Table 15, covering six pre-trained models on the PlantVillage.

4.2.2. PlantLeaves Dataset

The experimental results of all models on PlantLeaves Dataset are presented. Table 16 reports the weighted average performance of all pre-trained models under the proposed PlantClassiNet across all categories of PlantLeaves. Among these, the ResNet50 base model achieves the highest performance, with Precision: 0.9513, Recall: 0.9364, and F1-score: 0.9346.

For each model, the results includes two figures: a confusion matrix and training accuracy/loss curves, along with one table detailing Precision, Recall, and F1-score per class. These visuals are systematically organized in Figure 14, Figure 15, Figure 16, Figure 17, Figure 18, Figure 19, Figure 20, Figure 21, Figure 22, Figure 23, Figure 24 and Figure 25 and Table 17, Table 18, Table 19, Table 20, Table 21 and Table 22, covering six pre-trained models on the PlantLeaves.

4.2.3. Eggplant Dataset

The experimental results of all models on Eggplant dataset are presented. Table 23 reports the weighted average performance of all pre-trained models under the proposed PlantClassiNet across all categories of Eggplant. Among these, the DenseNet121 base model achieves the highest performance, with Precision: 1.0000, Recall: 1.0000, and F1-score: 1.0000.

For each model, the results includes two figures: a confusion matrix and training accuracy/loss curves, along with one table detailing Precision, Recall, and F1-score per class. These visuals are systematically organized in Figure 26, Figure 27, Figure 28, Figure 29, Figure 30, Figure 31, Figure 32, Figure 33, Figure 34, Figure 35, Figure 36 and Figure 37 and Table 24, Table 25, Table 26, Table 27, Table 28 and Table 29, covering six pre-trained models on the Eggplant.

4.3. Discussion

The proposed AdapFitu adaptive strategy enables rapid training because only a small subset of layers are updated and the number of epochs is limited. Nevertheless, preliminary experiments on three benchmark datasets exhibit high variance and sub-optimal classification accuracy. It is hypothesized that, although the strategy determines fine-tuning parameters by jointly considering model size, the number of frozen pre-trained layers, network capacity, and the scale of the downstream dataset, the individual or combined influence of these factors on fine-tuning effectiveness has not been rigorously validated. Therefore, a fixed-parameter protocol was adopted, in which multiple models are fine-tuned with deeper layers, a larger number of epochs, and consequently longer training times. This protocol yields markedly improved and stable results, achieving classification accuracies for the testing dataset that surpass those reported in most existing studies, as shown in

Table 30. Comparison of found plant leaf classification methods for PlantVillage.

Model	Testing Accuracy	Proposed PlantClassiNet
AlexNet [30]	0.9906	0.9938
AlexNet [31]	0.9948
MobileNetV3 [32]	0.9950	0.9937
MobileNetV3 [33]	0.9259
MobileNetV3 [34]	0.946
MobileNetV3 [35]	0.9927
DenseNet121 [36]	0.8620	0.9968
DenseNet121 [15]	0.9975
DenseNet121 [34]	0.791
ResNet50 [15]	0.9959	0.9946
ResNet50 [37]	0.982
ResNet50 [38]	0.9538
InceptionV3 [39]	0.9800	0.9876
InceptionV3 [35]	0.9803
EfficientNetB0 [34]	0.947	0.9950
EfficientNetB0 [35]	0.9803

and Figure 38.

Our study explicitly compares results with prior work, demonstrating superior classification accuracy over existing studies. Most prior work reports isolated results for single architectures and uses standard fine-tuning, while our approach integrates empirical validation for stability.

As illustrated in Figure 38, both ResNet50 and DenseNet121 consistently surpass 90% accuracy across all three benchmark datasets. More importantly, fine-tuning with frozen backbone parameters proves highly effective for large-scale architectures coupled with extensive data: AlexNet, ResNet50, and InceptionV3 all exceed 98% testing accuracy on the Plantvillage dataset. In contrast, lightweight models fine-tuned on smaller datasets exhibit comparable efficacy; MobileNetV3, DenseNet121, and EfficientNetB0 attain near-perfect classification performance (approximately 100%) on the Eggplant dataset.

Additionally, we can identify that some extreme values of 0 in the PlantLeaves dataset arise primarily due to the limited number of test samples for the corresponding categories, thereby possibly affecting the robustness of the metrics. These observations raise the hypothesis that large models are best suited to large datasets, whereas smaller datasets may benefit from fine-tuning compact architectures. Preliminary experiments with plant-domain adaptive optimization yielded inconsistent results due to data variability. We thus adopted a fixed-parameter fine-tuning approach, which proved more robust and reproducible. Fixed parameters may not generalize to all plant species or datasets with extreme imbalances.

In future work, we will systematically investigate hybrid optimization strategies (e.g., combining adaptive techniques with domain-specific constraints), explore multi-modal data integration approaches, and analyze the intricate interplay between dataset scale, model capacity, and their respective computational and memory resource demands.

Furthermore, the influence of key image attributes on model accuracy will be quantitatively assessed through analysis of noise levels, contrast variations, and object scale disparities. Subsequently, this evaluation will serve as the foundation for designing specialized architectures and customized training protocols, specifically adapted to address variations in image characteristics. This comprehensive analysis aims to identify optimal deployment strategies for practical applications.

5. Conclusions

This study leverages convolutional bases of computer vision pre-trained models and computes fine-tuning parameters through the proposed AdapFitu strategy. The performance of AdapFitu across all evaluated models demonstrated significant deviation from theoretical expectations, highlighting the intricate gap between conceptual design and practical implementation. While its dynamic adjustment mechanism is theoretically advantageous, empirical training processes revealed susceptibility to unaccounted variables—particularly gradient propagation characteristics and model structural coupling effects. These findings underscore the need for future research to develop a more comprehensive evaluation framework, integrating advanced metrics such as gradient sensitivity analysis and parameter coupling assessment into theoretical modeling.

Due to unsatisfactory results in subsequent exploratory experiments, the adaptive parameter strategy was revised to fixed-parameter fine-tuning training in subsequent research. The proposed PlantClassiNet with fix parameters was applied to PlantVillage, PlantLeaves, and Eggplant using AlexNet, ResNet50, InceptionV3, DenseNet121, EfficientNetB0, and MobileNetV3Small pre-trained models. During the initial training phase, the pre-trained convolutional base parameters remained fixed, while only the two dense layers underwent retraining. For the fine-tuning stage, the configuration involved freezing the top ten layers and applying a learning rate of 0.0001. After transfer learning and fine-tuning, the best model was used for inference on testing data. The experimental results achieved promising outcomes, effectively extracting leaf disease features and enabling multi-class classification of plant leaf diseases. Even though the performance of proposed PlantClassiNet with fixed train parameters is good, further research needs to be done to improve on the proposed AdapFitu adaptive strategy.

Furthermore, we plan to conduct dedicated experiments on domain adaptation and cross-dataset generalization in our future work to more comprehensively assess the method’s practical application potential. Future studies will systematically examine the relationship between dataset scale, model architecture, and their corresponding computational or memory costs to identify optimal deployment approaches in real-world applications.