通过数据生成策略提升深度学习模型性能：橄榄树病害分类

# Enhancing Deep Learning Model Performance through Data Generation Strategies: Classification of Olive Tree Diseases
Source: [https://arxiv.org/html/2610.35984](https://arxiv.org/html/2610.35984)
Hicham El AkhalEmail:[hi\.elakhal@edu\.umi\.ac\.ma](mailto:[email protected])Affiliation:Computer Science and Applications Laboratory, Faculty of Sciences, Moulay Ismail University of Meknes, PB 11201 Zitoune, Meknes, 50000, MoroccoAissa Ben YahyaEmail:[ai\.benyahya@edu\.umi\.ac\.ma](mailto:[email protected])Affiliation:Computer Science and Applications Laboratory, Faculty of Sciences, Moulay Ismail University of Meknes, PB 11201 Zitoune, Meknes, 50000, MoroccoAbdelbaki El Belrhiti El AlaouiEmail:[a\.elbelrhitielalaoui@umi\.ac\.ma](mailto:[email protected])Affiliation:Computer Science and Applications Laboratory, Faculty of Sciences, Moulay Ismail University of Meknes, PB 11201 Zitoune, Meknes, 50000, Morocco

###### Abstract

Deep learning models are highly effective in a wide range of domains, particularly in plant disease diagnosis\. However, their performance is often limited by challenges such as insufficient data and class imbalance\. In our work, we propose a comprehensive data generation strategy to overcome these limitations\. First, we apply various data generation techniques, including traditional augmentation methods, CutMix, MixUp, and synthetic image generation using Deep Convolutional Generative Adversarial Networks \(DCGANs\), to create new data samples\. Second, we exploit this generated data in three scenarios: \(S1\) balancing class distribution in the original dataset, \(S2\) increasing the number of samples per class, and \(S3\) expanding the dataset while maintaining balanced classes\. Third, we evaluate the impact of these data distributions \(S1, S2, S3\) on model performance using the EfficientNetV2B0 architecture with pre\-trained and fine\-tuned weights\. The experimental results show a substantial improvement in classification accuracy, emphasizing the critical role of data diversity and class balance\. These findings highlight the effectiveness and adaptability of the proposed strategy of data generation and distribution for enhancing dataset quality, offering promising applications in plant disease detection and precision agriculture\.

###### keywords

,Data Insufficiency, ,Data Imbalance, ,Data Generation, ,Deep Learning,,Classification

## 1Introduction

Plant diseases are a serious challenge to agricultural production, leading to significant crop losses and threatening global food security[Caparas et al\. \(2021\)](https://arxiv.org/html/2610.35984#bib.bib1)\. Traditional detection techniques, which mostly rely on visual inspection and manual evaluation, are frequently laborious, subjective, and prone to human error\. These issues are made worse in places where it is hard to reach agricultural specialists, as a delayed diagnosis can lead to a large decrease in productivity\.

In the field of agriculture, artificial intelligence \(AI\) has emerged as a revolutionary technology, particularly in the areas of plant disease detection and surveillance\. AI enables the early and precise identification of these diseases by leveraging extensive data, including sensor and image information[Liu \(2020\)](https://arxiv.org/html/2610.35984#bib.bib10);[Attri et al\. \(2023\)](https://arxiv.org/html/2610.35984#bib.bib4)\. Because it can automatically learn layered characteristics from raw data, deep learning, a subfield of artificial intelligence, plays a crucial role in this situation by reducing the need for intricate preprocessing or manually created features\. Using supervised and unsupervised learning techniques, advanced analysis models have been effectively used for tasks like finding sick areas, classifying plant diseases, and determining their severity[Hasan et al\. \(2020\)](https://arxiv.org/html/2610.35984#bib.bib11)\. These models have proved highly effective in detecting disease in a wide range of crops, including apple[Li et al\. \(2020a\)](https://arxiv.org/html/2610.35984#bib.bib12);[Liu et al\. \(2018\)](https://arxiv.org/html/2610.35984#bib.bib13), citrus[Khattak et al\. \(2021\)](https://arxiv.org/html/2610.35984#bib.bib14);[Çetiner \(2022\)](https://arxiv.org/html/2610.35984#bib.bib15), grapevine[Diana Andrushia et al\. \(2022\)](https://arxiv.org/html/2610.35984#bib.bib16), olive[Fazari et al\. \(2021\)](https://arxiv.org/html/2610.35984#bib.bib19);[Alshammari et al\. \(2022b\)](https://arxiv.org/html/2610.35984#bib.bib40);[El Akhal et al\. \(2023\)](https://arxiv.org/html/2610.35984#bib.bib21);[El Akhal et al\. \(2025\)](https://arxiv.org/html/2610.35984#bib.bib48), and tomato[Anandhakrishnan and Jaisakthi \(2022\)](https://arxiv.org/html/2610.35984#bib.bib17);[Karthik et al\. \(2020\)](https://arxiv.org/html/2610.35984#bib.bib18);[Kaur et al\. \(2022\)](https://arxiv.org/html/2610.35984#bib.bib22)\.

Despite these achievements, the lack of diverse and well\-balanced datasets remains a significant challenge\. The effectiveness of models can be negatively impacted by overfitting and limited generalization, which may arise from an imbalanced or biased data distribution\. However, techniques such as federated learning \(FL\), which combine decentralized data from multiple sources, seek to address this issue[Li et al\. \(2020b\)](https://arxiv.org/html/2610.35984#bib.bib38);[Mamba Kabala et al\. \(2023\)](https://arxiv.org/html/2610.35984#bib.bib37)\. This research adopts a different approach, focusing on the generation of new data\.

In this study, we generate enriched training data using a centralized single\-server infrastructure\. Our approach incorporates traditional data generation methods\.[Enkvetchakul and Surinta \(2022\)](https://arxiv.org/html/2610.35984#bib.bib20), as well as more advanced strategies such as CutMix[Pan et al\. \(2024\)](https://arxiv.org/html/2610.35984#bib.bib9), Mixup[Lewy and Mańdziuk \(2023\)](https://arxiv.org/html/2610.35984#bib.bib34), and deep convolutional generative adversarial networks \(DCGANs\)[Deshpande and Patidar \(2022\)](https://arxiv.org/html/2610.35984#bib.bib36);[Chen and Wu \(2023\)](https://arxiv.org/html/2610.35984#bib.bib35)\. These methods aim to improve the performance of convolutional neural networks \(CNNs\) in classifying olive tree diseases\.

The main goal of this research is to create a solid framework for data generation that will improve datasets used to categorize diseases of olive trees\. We focus on three specific scenarios:\(S1\)balancing the classes in the original dataset,\(S2\)increasing the number of samples per class, and\(S3\)expanding the dataset for all classes while maintaining a balanced class distribution\.

Through a generative diversity strategy addressing the limitation of having small datasets, the proposed model endeavors to enhance the precision and usability of models in diagnosing plant disease\. Unlike previous research that largely relies on primitive augmentation techniques, this paper presents a sophisticated pipeline that integrates traditional transformations with generative models, whose capacity can greatly improve the robustness of the models and facilitate their utilization in various applications of agriculture, especially in sparse data scenarios\. This is especially significant in low\-access\-data situations\.

The subsequent sections provide a detailed discussion in[section 5](https://arxiv.org/html/2610.35984#S5), present the results in[section 4](https://arxiv.org/html/2610.35984#S4), outline the experimental methodology in[section 3](https://arxiv.org/html/2610.35984#S3), and review the relevant literature in[section 2](https://arxiv.org/html/2610.35984#S2)\.[section 6](https://arxiv.org/html/2610.35984#S6), the study’s conclusion, offers possible avenues for further investigation\.

## 2Literature Review

### 2\.1Data Generation and Augmentation Strategies

This section examines recent developments in data creation and augmentation techniques intended to tackle issues like class imbalance and data scarcity in image\-based classification problems\. Their use in plant disease detection is highlighted in particular, since these methods have been crucial in enhancing models’ resilience, generality, and accuracy\.

To improve plant disease classification,[Tabbakh and Barpanda \(2023\)](https://arxiv.org/html/2610.35984#bib.bib26)suggested TLMViT, a hybrid model that combines Vision Transformers \(ViT\) and transfer learning\. By incorporating efficient data augmentation techniques, their model successfully mitigated overfitting and achieved 99\.86% accuracy on a wheat disease dataset\.

[Mhala et al\. \(2025\)](https://arxiv.org/html/2610.35984#bib.bib33)used data augmentation, L2 regularization, and transfer learning to address class imbalance in potato disease datasets\. Their final result was 81\.31%, with a 4\.17% increase in classification accuracy using DenseNet201\.

[Singh and Sanodiya \(2023\)](https://arxiv.org/html/2610.35984#bib.bib27)investigated domain generalization using a framework for zero\-shot transfer learning\. Even in cases where domain classes did not coincide, their approach showed strong generalization capabilities by fusing data augmentation with synthetic data synthesis\.

[Sathya et al\. \(2025\)](https://arxiv.org/html/2610.35984#bib.bib28)created artificial images of both healthy and ill leaves using conditional GANs\. An attention\-augmented ResNet combined with Faster R\-CNN was trained using these, yielding a classification accuracy of 98\.78%\.

[Wu et al\. \(2023\)](https://arxiv.org/html/2610.35984#bib.bib32)added attention mechanisms to a ConvNeXt\-based architecture by integrating random masking into their data augmentation pipeline\. Using this method, the average accuracy on a soybean disease classification task was 85\.42%\.

[Karthikeyan and Raja \(2023\)](https://arxiv.org/html/2610.35984#bib.bib31)created DTLDN\-CBIRA, a content\-based image retrieval system based on DenseNet\. With 81\.90% recall and an F1\-score of 89\.90%, their model reached 100% accuracy by augmenting limited training data with rotation and mirroring\.

[Haikal et al\. \(2024\)](https://arxiv.org/html/2610.35984#bib.bib30)presented a framework for composite augmentation that combines FMix, MixUp, Cutout, and CutMix\. When used to classify rice leaf diseases, the method produced accuracy rates of 99\.75% with MobileNetV3 and 91\.51% with ResNet34\.

[Chen et al\. \(2024\)](https://arxiv.org/html/2610.35984#bib.bib29)CNN and ViT modules were integrated to improve the FasterViT architecture\. Their method, which combined MixUp, Cutout, Triplet Attention, and Adaptive Average Pooling, showed better results on PlantVillage and small\-scale datasets\.

A summary of these approaches is presented in[Table 1](https://arxiv.org/html/2610.35984#S2.T1), which highlights the main data generation and augmentation strategies for classification tasks\.

Table 1:Summary of Related Work on Data Generation and Augmentation Strategies for Classification
### 2\.2Review of Olive Disease Classification Techniques

Deep learning has revolutionized artificial intelligence, offering powerful solutions for accurately and rapidly classifying plant diseases, including those affecting olive trees\. This section presents a detailed overview of recent advances in olive disease classification\.

[Cruz et al\. \(2017\)](https://arxiv.org/html/2610.35984#bib.bib39)created a deep learning framework that uses computer vision to identify OQDS, with a true positive rate of 98\.60 ± 1\.47%\.

[Uğuz and Uysal \(2021\)](https://arxiv.org/html/2610.35984#bib.bib23)carried out tests comparing models with and without data augmentation using hand\-gathered photos of olive leaves\. The accuracy of augmentation\-trained models was 95%\.

[Alshammari et al\. \(2022b\)](https://arxiv.org/html/2610.35984#bib.bib40)used a deep learning pipeline with a genetic algorithm for best results\. The accuracy of their model was 98% in binary classification and 96% in multi\-class tasks\.

[Alshammari et al\. \(2022a\)](https://arxiv.org/html/2610.35984#bib.bib41)suggested a hybrid architecture that combines CNNs and Vision Transformers, achieving 97% accuracy for binary classification and roughly 96% for multi\-class issues\.

[Ksibi et al\. \(2022\)](https://arxiv.org/html/2610.35984#bib.bib24)presented MobiRes\-Net, a ResNet50/MobileNet hybrid that was trained using 5,400 photos obtained by UAVs\. The accuracy of the model was 97\.08%\.

[Alshammari et al\. \(2023\)](https://arxiv.org/html/2610.35984#bib.bib42)created an enhanced ANN for feature selection and classification by utilizing the Whale Optimization Algorithm \(WOA\)\. All evaluation parameters, including F1\-score, precision, recall, and accuracy, showed gains with this approach\.

[Alsaeedi et al\. \(2023\)](https://arxiv.org/html/2610.35984#bib.bib43)offered a five\-step methodology that includes deep learning classification, feature extraction, clustering, and picture enhancement\. Their model achieved 98% accuracy with more than 3,200 photos, surpassing both VGG16 \(96% \) and AlexNet \(95%\)\.

[Mahmood and Alsalem \(2024\)](https://arxiv.org/html/2610.35984#bib.bib44)suggested a technique that combines pre\-trained networks with wavelet transforms to obtain highly discriminative features for precise illness identification\.

[Bruno et al\. \(2023\)](https://arxiv.org/html/2610.35984#bib.bib45)presented a two\-model ensemble technique utilizing EfficientNet\-b0\. The classification of olive diseases was made more accurate and resilient by this adaptive method\.

A summary of these approaches is provided in[Table 2](https://arxiv.org/html/2610.35984#S2.T2)\. While the reported models demonstrate high performance, they also highlight persistent challenges such as overfitting, limited dataset diversity, and computational constraints\.

Table 2:Summary of Studies on Olive Disease ClassificationAs shown in[Table 2](https://arxiv.org/html/2610.35984#S2.T2), the majority of previous studies rely on classical data generation techniques, or even no explicit generation at all\. These approaches often achieve high accuracy, but on relatively homogeneous datasets\. However, differences in accuracy can also be attributed to the diversity of data used in different studies\. For example, some studies used smaller or specific datasets, while others worked with larger and more diverse data\. The proposed work combines traditional and advanced data generation methods\-such as CutMix, MixUp, and DCGAN\-to simulate more varied distribution scenarios\. Although the accuracy obtained \(93\.48%\) is slightly lower than that of some conventional methods, this compromise allows greater robustness and generalizability in the face of real data variations\. The difference in accuracy reflects not only the choice of data generation techniques but also the inherent differences in the datasets used, which probably led to distinct model behavior and performance\.

### 2\.3Research Gap Analysis and Contributions

Despite the promising advances discussed, most existing studies either focus on a single augmentation strategy or fail to explore the impact of data distribution scenarios on classification performance\. Additionally, few works systematically integrate both classical and generative generation techniques while evaluating their effects across multiple imbalance contexts\. To bridge this gap, the proposed work introduces a robust and comprehensive data generation strategy combining traditional augmentation, CutMix, MixUp, and DCGAN, evaluated under three distinct distribution scenarios \(S1, S2, S3\) for olive disease classification\. This approach aims to enhance both the generalization and the robustness of deep learning models under challenging real\-world constraints\.

### 2\.4Research Contributions

This study suggests a data generation technique to enhance the classification of diseases affecting olive trees\. The following is a summary of this work’s primary contributions:

1. 1\.Hybrid Data Generation Framework:We introduce a hybrid strategy that combines traditional augmentation techniques with MixUp, CutMix, and DCGAN\. This method improves classification performance, especially for underrepresented classes, by increasing the dataset’s diversity and representativeness\.
2. 2\.Mitigating Class Imbalance:The proposed framework explicitly addresses the issue of class imbalance by generating additional samples for minority classes and adjusting the overall data distribution\. This results in a more balanced training dataset, reduces model bias, and improves overall accuracy\.
3. 3\.Dataset Expansion Strategy:Beyond combining traditional augmentations, CutMix, MixUp, and DCGANs, we evaluate the framework under three distinct distribution scenarios \(S1, S2, S3\) for olive disease classification\. This strategy enhances the model’s generalizability and reduces the risk of overfitting, especially in data\-scarce contexts\.
4. 4\.Comprehensive Performance Evaluation:We carefully investigate the proposed data generation techniques to ascertain their impact on the classification accuracy model and robustness in identifying olive tree diseases\.

### 2\.5Problem Description

The impact of data generation methods on the categorization of diseases affecting olive trees is investigated in this study\. We combine traditional techniques \(such as geometric and color transformations\) with cutting\-edge strategies like Mixup, CutMix, and DCGAN to address problems of data scarcity and class imbalance\. Improving model generalization and accuracy, particularly for underrepresented classes, is the goal\.

Key challenges include:

- •Class Imbalance:Certain disease classes are under\-sampled, leading to biased models\.
- •Overfitting:Too much or unwarranted augmentation can add artefacts, heightening the danger of overlearning\.
- •Selecting Effective Augmentation:To enhance performance without sacrificing robustness, one needs to choose the right methods for a specific dataset and task\.

This study compares and evaluates different augmentation techniques to improve generalisation and avoid over\-fitting in a bid to improve plant disease detection models in low\-data scenarios\.

## 3Solution Approach

The primary phases of the proposed methodology, data preparation, model development, and performance evaluation, are explained in this section\. Two subsets of the data were initially prepared: Dataset A for training and Dataset B for evaluating generalization as a new test set\. To obtain a more balanced and diverse training dataset, certain data generation techniques, such as synthesis by DCGAN, MixUp, CutMix, and traditional transformations, were carried out on Dataset A\. Subsequently, to ensure an equal comparison across the different experimental conditions \(S1, S2, S3\), the original and generated samples were split into training and validation sets\. These datasets were used for training and fine\-tuning a deep learning architecture, i\.e\., EfficientNetV2B0, to develop the model\. Finally, the performance of the model was assessed on Dataset B, which was untouched during training, to get an unbiased estimation of each scenario’s performance in olive disease classification\. The entire simulation workflow is shown in[Figure 1](https://arxiv.org/html/2610.35984#S3.F1)\.

![Refer to caption](https://arxiv.org/html/2610.35984v1/fig1.png)Figure 1:The proposed block work\.### 3\.1Data Pereparation

#### 3\.1\.1Data Collection

Table 3:Image count per class in the dataset![Refer to caption](https://arxiv.org/html/2610.35984v1/fig2.png)Figure 2:Example images depicting various categories of diseases affecting olive leaves\.

### 3\.2Data Generation

Data generation techniques are becoming more and more important in modern deep learning applications, both to improve model generalization on heterogeneous and imbalanced datasets and to avoid overfitting\. This work uses a variety of data creation techniques, including as classical augmentation, CutMix, MixUp, DCGAN, and a thorough mix of these approaches, to improve the performance of the olive illness classification model\.

Traditional Augmentation:Several standard data augmentation methods, such as rotation, flipping vertically and horizontally, and zooming, were used to enlarge dataset A\. The training data became more varied and variable as a result of this procedure, which produced a dataset that was around four times larger than the original\.

CutMix:CutMix[Yun et al\. \(2019\)](https://arxiv.org/html/2610.35984#bib.bib8)is a regularization method that creates fresh training samples by merging the labels of various images proportionately and clipping and pasting patches between them\. After pasting a patch fromxBx\_\{B\}ontoxAx\_\{A\}given two input picturesxAx\_\{A\}andxBx\_\{B\}, the new imagex~\\tilde\{x\}and labely~\\tilde\{y\}are defined as follows:

x~=M⊙xA\+\(1−M\)⊙xB\(Eq\. 1\)\\tilde\{x\}=M\\odot x\_\{A\}\+\(1\-M\)\\odot x\_\{B\}\\quad\\text\{\(Eq\. 1\)\}\(1\)
y~=λ​yA\+\(1−λ\)​yB\(Eq\. 2\)\\tilde\{y\}=\\lambda y\_\{A\}\+\(1\-\\lambda\)y\_\{B\}\\quad\\text\{\(Eq\. 2\)\}\(2\)
whereMMis a binary mask indicating the patched region,⊙\\odotdenotes element\-wise multiplication, andλ\\lambdais sampled from a Beta distribution\.

MixUp:[Zhang et al\. \(2017\)](https://arxiv.org/html/2610.35984#bib.bib5)generates new training samples by linearly interpolating two random imagesxix\_\{i\}andxjx\_\{j\}, and their labelsyiy\_\{i\}andyjy\_\{j\}, using a mixing parameterλ∈\[0,1\]\\lambda\\in\[0,1\]:

x~=λ​xi\+\(1−λ\)​xj\(Eq\. 3\)\\tilde\{x\}=\\lambda x\_\{i\}\+\(1\-\\lambda\)x\_\{j\}\\quad\\text\{\(Eq\. 3\)\}\(3\)
y~=λ​yi\+\(1−λ\)​yj\(Eq\. 4\)\\tilde\{y\}=\\lambda y\_\{i\}\+\(1\-\\lambda\)y\_\{j\}\\quad\\text\{\(Eq\. 4\)\}\(4\)
This method encourages the model to behave linearly in\-between training examples, enhancing generalization and reducing overfitting\.

DCGAN:In this study, we utilized Deep Convolutional Generative Adversarial Networks \(DCGANs\)[Radford et al\. \(2015\)](https://arxiv.org/html/2610.35984#bib.bib6)to generate synthetic images of olive diseases\. A DCGAN consists of a generatorGGand a discriminatorDDtrained in a minimax game:

minG⁡maxD​𝔼x∼pdata​\(x\)​\[log⁡D⁡\(x\)\]\+𝔼z∼pz​\(z\)​\[log⁡\(1−D⁡\(G⁡\(z\)\)\)\]\(Eq\. 5\)\\min\_\{G\}\\max\_\{D\}\\;\\mathbb\{E\}\_\{x\\sim p\_\{\\text\{data\}\}\(x\)\}\[\\log D\(x\)\]\+\\mathbb\{E\}\_\{z\\sim p\_\{z\}\(z\)\}\[\\log\(1\-D\(G\(z\)\)\)\]\\quad\\text\{\(Eq\. 5\)\}\(5\)
wherezzis a random noise vector\. Through iterative training,GGlearns to generate realistic images, whileDDlearns to distinguish real from fake samples\.

Frechet Inception Distance \(FID\):To measure the quality of generated images, we used the FID metric[Jayasumana et al\. \(2024\)](https://arxiv.org/html/2610.35984#bib.bib7)which computes the distance between the feature distributions of real and synthetic data:

FID=‖μr−μg‖2\+Tr​\(Σr\+Σg−2​\(Σr​Σg\)1/2\)\(Eq\. 6\)\\text\{FID\}=\\left\\\|\\mu\_\{r\}\-\\mu\_\{g\}\\right\\\|^\{2\}\+\\text\{Tr\}\\left\(\\Sigma\_\{r\}\+\\Sigma\_\{g\}\-2\(\\Sigma\_\{r\}\\Sigma\_\{g\}\)^\{1/2\}\\right\)\\quad\\text\{\(Eq\. 6\)\}\(6\)
where\(μr,Σr\)\(\\mu\_\{r\},\\Sigma\_\{r\}\)and\(μg,Σg\)\(\\mu\_\{g\},\\Sigma\_\{g\}\)are the mean and covariance of features extracted from real and generated images using an Inception network\.

In summary, all of these approaches were employed to generate new data samples, which were subsequently distributed over the originally collected dataset to improve class balance and enhance the model’s performance\.

### 3\.3Data Distribution

After collecting olive leaf disease samples and expanding the dataset using data augmentation techniques, we designed a clear and structured strategy to distribute the data\. We defined three main experimental setups, each with a different way of organizing the data\. For each setup, we created a specific distribution plan to better understand how different data arrangements influence the model’s performance\.

1. 1\.S1: Balancing class distributions in the original dataset:In this case, new data was used to balance the number of samples between classes in the original dataset\. This process was unavoidable for reducing model bias against majority classes and enhancing classification fairness\.
2. 2\.S2: Increasing the sample size of each class:In this case, the intention was to enhance the dataset by adding more samples per class using the created data\. Enhancing the learning and generalization of the model was the intention of this improvement\.
3. 3\.S3: Balancing class distributions while expanding the dataset size:In this instance, class balancing goals and dataset expansion goals were combined\. Not only did it increase the overall amount of data, but also gave a balanced view of all classes, which is essential in order to get the highest possible accuracy and potency of the model\.

### 3\.4Model design

This study uses leaf imagery to identify diseases in olive trees using the pre\-trained EfficientNetV2B0 model\. Acknowledged for its effectiveness and precision in object recognition tasks, this architecture gains from previous training on large datasets like ImageNet and ImageNet21k[Tan and Le \(2021\)](https://arxiv.org/html/2610.35984#bib.bib25)\. We modified the model to fit our classification task by substituting a custom classifier consisting of four fully connected layers, each followed by a Dropout layer, and ending with a softmax activation \(see[Figure 3](https://arxiv.org/html/2610.35984#S3.F3)\)\. Although several deep learning models were tested during the experiment, EfficientNetV2B0 continuously produced better results\. Existing research demonstrates that its selection was justified by its balance of high predictive performance, fast convergence, and computational efficiency\.

![Refer to caption](https://arxiv.org/html/2610.35984v1/fig3.png)Figure 3:The architecture proposed for the classification\.
### 3\.5Perfermence Evalution

Several evaluation metrics were used in this study to measure the performance of the olive disease classification model\. The indicators used include accuracy, F1 score, precision, recall, Matthews Correlation Coefficient \(MCC\), and geometric mean \(G\-mean\)[Chicco and Jurman \(2020\)](https://arxiv.org/html/2610.35984#bib.bib46);[de la Cruz Huayanay et al\. \(2024\)](https://arxiv.org/html/2610.35984#bib.bib47)\. For a more thorough examination, the Non\-Parametric Scott\-Knott ESD \(NPSK\) test\([Tantithamthavorn et al\., 2016](https://arxiv.org/html/2610.35984#bib.bib2);[Tantithamthavorn et al\., 2018](https://arxiv.org/html/2610.35984#bib.bib3)\)was also used\. The main metric was accuracy, which was validated using 5\-fold cross\-validation \(k = 5\)\. The Scott\-Knott ESD test uses hierarchical clustering to group methods and determine the best distribution strategy\. It is implemented using the ScottKnottESD package in R\.

### 3\.6Computational Resources and Time Requirements

The experimental configuration used to implement our proposed method and evaluate its computational cost is presented in[Table 4](https://arxiv.org/html/2610.35984#S3.T4)\.

•Operating System:Windows 10•Software:TensorFlow, configured to run on the above operating system•RAM:46 GB•GPU:NVIDIA GeForce RTX 2060 with 12 GB of dedicated RAM•CPU:Intel\(R\) Xeon\(R\) E5\-2630 v3, 2\.40 GHz

Table 4:Computational Costs Associated with the Proposed Methodology

## 4Results Analysis

[Table 5](https://arxiv.org/html/2610.35984#S4.T5)presents the classification accuracy obtained with different data generation techniques in three experimental scenarios \(S1, S2, S3\), compared to a baseline model without augmentation \(accuracy = 90\.99%\)\.

S1: Balanced classes

In this scenario, all the methods outperformed the baseline\. The most significant improvement was seen with the full combination of augmentations \(Traditional \+ MixUp \+ CutMix \+ DCGAN\), achieving an accuracy of 93\.09%, an improvement of \+2\.10 points on the baseline\. CutMix alone also improved significantly by \+0\.91 points, followed by DCGAN \(\+1\.25\) and MixUp \(\+0\.51\)\. This shows that data augmentation is particularly beneficial in balanced contexts\.

S2: Increased class volume \(imbalanced\)

The results were more mixed\. Only the MixUp method significantly improved performance with \+0\.74 points\. The full combination of augmentations maintains a slight improvement \(\+0\.34\), while the traditional \(\-0\.23\) and CutMix \(\-0\.28\) methods underperform\. This indicates that certain augmentations can accentuate the imbalance, whereas MixUp, with its linear interpolation, favours better regularisation\.

S3: Balanced classes \+ increased volume

In this mixed scenario, CutMix is the best performing method with \+1\.30 points, demonstrating its effectiveness with larger datasets\. DCGAN and MixUp show moderate gains \(\+0\.40 and \+0\.11, respectively\), while the full combination brings no net gain \(\+0\.00\), suggesting saturation or redundancy of effects\.

Table 5:Test accuracy on the olive disease dataset using various data generation techniques\.S1: Balanced classes S2: Increased class volume \(imbalanced\) S3: Balanced classes \+ increased volume

[Table 6](https://arxiv.org/html/2610.35984#S4.T6)presents the evaluation scores \(Accuracy, Recall, F1\-score, G\-mean, and MCC\) obtained on the olive disease dataset for different scenarios \(S1, S2, S3\) and generation methods\.

In scenario S1 \(balanced classes\), the best overall performance is observed with the combined approach \(all augmentations\), which achieves an F1\-score of 91\.89%, a G\-mean of 90\.37% and an MCC of 90\.15% , outperforming the baseline \(F1\-score = 89\.63%\)\. Using DCGAN alone also shows a significant improvement with an F1\-score of 93\.14% , suggesting better generalisation through the generation of new synthetic images\.

In scenario S2 \(volume increase without balancing\), the results are even more significant\. The combined approach obtains the highest scores on almost all indicators, reaching 95\.99% for accuracy, 93\.32% for F1\-score and 91\.04% for G\-mean, indicating that the joint increase in data and its diversity are particularly beneficial in cases of class imbalance\. Data generation with DCGAN alone also performed very well \(F1\-score = 93\.35%\), demonstrating the ability of this generator to fill gaps in under\-represented classes\.

For scenario S3 \(balanced classes with more volume\), the trend is more nuanced\. Although CutMix provides the best results in terms of F1\-score \(92\.28%\) and MCC \(89\.48%\), the combined approach does not always outperform the isolated methods\. This could indicate redundancy or a saturation effect when several techniques are used simultaneously on a game that is already sufficiently rich and balanced\. In summary, the results highlight that generation techniques, particularly DCGAN and advanced combinations \(MixUp, CutMix\), significantly improve the model’s robustness, particularly in scenarios of imbalance or low data volume\.

Table 6:Evaluation scores on the test dataset \(%\)\.[Figure 4](https://arxiv.org/html/2610.35984#S4.F4)presents the results of a cross\-validation procedure applied to evaluate the ranking of the training data sets used in this study\. The figure visualizes the comparative performance of the different training datasets through a ranking system, highlighting the models’ responses when trained on these datasets\. This ranking helps identify which dataset configurations have the most significant impact on model performance, providing valuable insights into data selection and distribution strategies\.

![Refer to caption](https://arxiv.org/html/2610.35984v1/fig4.png)Figure 4:Cross\-validation Results\.[Figure 5](https://arxiv.org/html/2610.35984#S4.F5)illustrates another key aspect of model evaluation, focusing on the ranking of training data sets as assessed during the study\. The figure compares the impact of various datasets on model performance, with each set ranked based on its efficacy in improving the model’s accuracy\. By analyzing these rankings, the study aims to identify the most effective data subsets for optimizing model training\.

![Refer to caption](https://arxiv.org/html/2610.35984v1/fig5.png)Figure 5:Ranking of Training Data Distributions Utilized in this Study for Model Training\.
## 5Discussion

The experimental results obtained using different scenarios and data generation techniques reveal several key insights into improving the performance of classification models, particularly in the context of olive disease detection\. Initially, traditionel augmentation methods such as rotation and inversion offer a modest yet consistent enhancement over baseline performance, highlighting their value in typical scenarios\. Nevertheless, their standalone effectiveness is constrained, especially in cases involving class imbalance or substantial inter\-class resemblance\. These techniques introduce only limited variation to the original dataset, contributing minimal new information to the model, which accounts for the relatively minor performance gains observed\.

The integration of advanced augmentation methods, such as MixUp and CutMix, shows more competitive results, particularly in scenarios where the classes are balanced \(S1\) or enriched \(S3\)\. These techniques, based on the creation of interpolated or hybrid images, facilitate the learning of stronger and more generalizable representations by the model, which translates into better generalization over the test set\. However, these methods have certain limitations: in some cases, they can introduce artificial samples that are not very informative for learning\. Their effectiveness, therefore, depends heavily on the application scenario and their judicious use within a justified framework\.

The generation of synthetic images using DCGAN has proved remarkably effective, particularly in scenarios involving imbalance \(S2\)\. By enriching the minority classes with realistic artificial samples, DCGAN reduces the impact of the bias induced by the initial distribution of the data\. This ability is particularly highlighted by the high F1\-score and MCC in S2, where the synthetic data compensates for the lack of diversity\. The use of DCGAN has a positive influence on the performance of the classification model, as it makes it possible to generate diverse samples from the original data\. However, this approach also raises ethical constraints linked to the quality and validity of the data generated\. This is why, in this study, we used the Fréchet distance to select only high\-quality samples, so as not to disrupt the model learning process\.

The combined approach, combining all the techniques \(Traditional, MixUp, CutMix, DCGAN\), generally gives the best overall performance, particularly in scenarios S1 and S2\. This confirms that diversification of augmentation strategies can act in a complementary way, increasing the quantity, variability, and representativeness of the training data\.

However, in scenario S3 \(data already balanced and voluminous\), the effect of certain combinations seems to diminish\. This phenomenon can be explained by model saturation, where the excessive addition of modified or synthetic samples no longer provides useful information and may even introduce noise\. In summary, this study underscores the significance of tailoring data generation strategies to the unique attributes of a given dataset\. The thoughtful application of advanced techniques, especially DCGAN and well\-planned combinations, proves to be a crucial factor in enhancing computer vision performance, both in agricultural contexts and broader domains\.

### 5\.1Sensitivity Analysis

A sensitivity analysis was carried out by comparing the relative impact of each method of increase on the different scenarios:

S1 \(balanced classes\): All augmentation methods showed significant gains, with a more marked sensitivity for CutMix \(\+0\.91\) and DCGAN \(\+1\.25\), indicating their relevance even on well\-structured datasets\.

S2 \(augmented volume\): The performance of traditional augmentations and CutMix is less sensitive or even negatively impacted \(e\.g\., CutMix at \-0\.28\), while MixUp and DCGAN remain robust, with a positive of \+0\.74 and \+0\.34, respectively\. This shows that the robustness of the techniques strongly depends on the initial imbalance of the data\.

S3 \(balanced \+ voluminous\): The gains are more moderate, even zero for some combinations\. This suggests performance saturation, as the model has already been exposed to sufficient diversity\. However, CutMix retains an advantage \(\+1\.30\), indicating that it can still enrich the representations, even in well\-fed cases\.

In summary, the sensitivity of performance to augmentation methods depends closely on the characteristics of the dataset: volume, balance, and intra\-class diversity\. DCGAN and CutMix stand out as the most efficient and stable, while classical methods are to be preferred in simple cases or for low\-variance tasks\.

### 5\.2Ethical Considerations in Data Generation Techniques

Use of advanced generation methods like DCGAN, MixUp, and CutMix is of serious ethical concern, especially in sensitive areas like agricultural disease detection\.

To guarantee ethical integrity and minimize any potential biases stemming from our approach to generating data, the following actions were taken:

- •Ethical Practices:We promote ethical handling of data during the process of generating and collecting the data\. We are committed to preventing artificial bias or unfair results, especially by systematically visualizing the generated data to ensure that the resulting samples are unbiased and of high quality\.
- •Bias Mitigation Strategies:To reduce potential model bias, we carefully select data sources and actively monitor class distributions for imbalances\. This ensures that the generated data supports real samples in a representative and balanced manner\.
- •Transparency and Fairness:Transparency in the generation and selection of synthetic data is a key objective\. For images generated by DCGAN, we employ the Fréchet Inception Distance \(FID\) to evaluate and retain the highest\-quality samples\.

By integrating these ethical considerations into our solution, we aim to foster fair, transparent, and replicable AI practices in precision agriculture\.

## 6Conclusions, limitations, and future research

### 6\.1Findings

This paper introduces a new method for olive tree disease classification by combining traditional data augmentation methods with state\-of\-the\-art approaches such as MixUp, CutMix, and Deep Convolutional Generative Adversarial Networks \(DCGAN\)\. Experiments on a multi\-class dataset with seven classes, namely disease\-free olive leaves and other disease states, indicate that the proposed data generation strategy in this paper significantly enhances model accuracy and robustness\.

The results confirm that data generation plays a crucial role in improving classification performance, particularly when used in well\-defined scenarios rather than applied randomly\. The proposed approach achieved a performance improvement of 2\.48% compared to the baseline \(no augmentation\)\.

### 6\.2Future Research

Future research could explore the following avenues to build upon and extend the findings of this study:

- •Exploration of Additional Generation Techniques:To better address issues with intra\-class variability and inter\-class similarity, future research should look at more sophisticated data generation methods\.
- •Real\-World Deployment and Validation:Applying the proposed method in real agricultural settings would allow for the evaluation of its practical viability, robustness, and adaptability under real\-world conditions\.

### 6\.3Limitations of Research

Despite its promising results, this study has a few limitations:

- •Data Dependency:The quality and balance of the dataset have an impact on the model’s performance\. The effectiveness of the model’s generalization may be hampered by incomplete or skewed data\.
- •Computational Cost:Techniques like DCGAN require significant resources, which can limit scalability\.
- •Parameter Sensitivity:The approach is sensitive to the volume of generated data and key model hyperparameters, requiring careful tuning for optimal results\.

## References

- Alsaeediet al\.\(2023\)A\. H\. Alsaeedi, A\. M\. Al\-juboori, H\. H\. R\. Al\-Mahmood, S\. M\. Hadi, H\. J\. Mohammed, M\. R\. Aziz, M\. Aljibawi, and R\. R\. NuiaaDynamic clustering strategies boosting deep learning in olive leaf disease diagnosis\.Sustainability15\(18\),pp\. 13723\.Note:[https://doi\.org/10\.3390/su151813723](https://doi.org/10.3390/su151813723)Cited by:[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p8.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.8.1.1.1)\.
- Alshammariet al\.\(2022a\)H\. Alshammari, K\. Gasmi, I\. Ben Ltaifa, M\. Krichen, L\. Ben Ammar, and M\. A\. MahmoodOlive disease classification based on vision transformer and cnn models\.Computational Intelligence and Neuroscience2022\.Note:[https://doi\.org/10\.1155/2022/3998193](https://doi.org/10.1155/2022/3998193)Cited by:[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p5.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.5.1.1.1)\.
- Alshammariet al\.\(2022b\)H\. Alshammari, K\. Gasmi, M\. Krichen, L\. B\. Ammar, M\. O\. Abdelhadi, A\. Boukrara, and M\. A\. MahmoodOptimal deep learning model for olive disease diagnosis based on an adaptive genetic algorithm\.Wireless Communications and Mobile Computing2022\.Note:[https://doi\.org/10\.1155/2022/8531213](https://doi.org/10.1155/2022/8531213)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1),[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p4.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.4.1.1.1)\.
- Alshammariet al\.\(2023\)H\. H\. Alshammari, A\. I\. Taloba, and O\. R\. ShahinIdentification of olive leaf disease through optimized deep learning approach\.Alexandria Engineering Journal72,pp\. 213–224\.Note:[https://doi\.org/10\.1016/j\.aej\.2023\.03\.081](https://doi.org/10.1016/j.aej.2023.03.081)Cited by:[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p7.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.7.1.1.1)\.
- Anandhakrishnan and Jaisakthi \(2022\)T\. Anandhakrishnan and S\.M\. JaisakthiDeep convolutional neural networks for image based tomato leaf disease detection\.Sustainable Chemistry and Pharmacy30,pp\. 100793\.Note:[https://doi\.org/10\.1016/j\.scp\.2022\.100793](https://doi.org/10.1016/j.scp.2022.100793)External Links:ISSN 2352\-5541,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.scp.2022.100793),[Link](https://www.sciencedirect.com/science/article/pii/S2352554122001978)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Attriet al\.\(2023\)I\. Attri, L\. K\. Awasthi, and T\. P\. SharmaMachine learning in agriculture: a review of crop management applications\.Multimedia Tools and Applications,pp\. 1–41\.Note:[https://doi\.org/10\.1007/s11042\-023\-16105\-2](https://doi.org/10.1007/s11042-023-16105-2)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Brunoet al\.\(2023\)A\. Bruno, D\. Moroni, and M\. MartinelliEfficient deep learning approach for olive disease classification\.In2023 18th Conference on Computer Science and Intelligence Systems \(FedCSIS\),pp\. 889–894\.Note:[https://doi\.org/110\.15439/2023F4794](https://doi.org/110.15439/2023F4794)Cited by:[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p10.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.10.1.1.1)\.
- Caparaset al\.\(2021\)M\. Caparas, Z\. Zobel, A\. D\.A\. Castanho, and C\. R\. SchwalmIncreasing risks of crop failure and water scarcity in global breadbaskets by 2030\.Environmental Research Letters16\(10\),pp\. 104013\.Note:[https://doi\.org/10\.1088/1748\-9326/ac22c1](https://doi.org/10.1088/1748-9326/ac22c1)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p1.1)\.
- Çetiner \(2022\)H\. ÇetinerCitrus disease detection and classification using based on convolution deep neural network\.Microprocessors and Microsystems95,pp\. 104687\.Note:[https://doi\.org/10\.1016/j\.micpro\.2022\.104687](https://doi.org/10.1016/j.micpro.2022.104687)External Links:ISSN 0141\-9331,[Link](https://www.sciencedirect.com/science/article/pii/S0141933122002174)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Chenet al\.\(2024\)J\. Chen, S\. Wang, J\. Guo, F\. Chen, Y\. Li, and H\. QiuImproved fastervit model for citrus disease diagnosis\.Heliyon10\(16\)\.Cited by:[§2\.1](https://arxiv.org/html/2610.35984#S2.SS1.p9.1),[Table 1](https://arxiv.org/html/2610.35984#S2.T1.2.9.1.1.1)\.
- Chen and Wu \(2023\)Y\. Chen and Q\. WuGrape leaf disease identification with sparse data via generative adversarial networks and convolutional neural networks\.Precision Agriculture24\(1\),pp\. 235–253\.Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p4.1)\.
- Chicco and Jurman \(2020\)D\. Chicco and G\. JurmanThe advantages of the matthews correlation coefficient \(mcc\) over f1 score and accuracy in binary classification evaluation\.BMC genomics21,pp\. 1–13\.Note:[https://doi\.org/10\.1186/s12864\-019\-6413\-7](https://doi.org/10.1186/s12864-019-6413-7), publisher=SpringerCited by:[§3\.5](https://arxiv.org/html/2610.35984#S3.SS5.p1.1)\.
- Cruzet al\.\(2017\)A\. C\. Cruz, A\. Luvisi, L\. De Bellis, and Y\. AmpatzidisX\-fido: an effective application for detecting olive quick decline syndrome with deep learning and data fusion\.Frontiers in plant science8,pp\. 1741\.Note:[https://doi\.org/10\.3389/fpls\.2017\.01741](https://doi.org/10.3389/fpls.2017.01741)Cited by:[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p2.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.2.1.1.1)\.
- de la Cruz Huayanayet al\.\(2024\)A\. de la Cruz Huayanay, J\. L\. Bazán, and C\. M\. RussoPerformance of evaluation metrics for classification in imbalanced data\.Computational Statistics,pp\. 1–27\.Note:[https://doi\.org/10\.1007/s00180\-024\-01539\-5](https://doi.org/10.1007/s00180-024-01539-5)Cited by:[§3\.5](https://arxiv.org/html/2610.35984#S3.SS5.p1.1)\.
- Deshpande and Patidar \(2022\)R\. Deshpande and H\. PatidarTomato plant leaf disease detection using generative adversarial network and deep convolutional neural network\.The Imaging Science Journal70\(1\),pp\. 1–9\.Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p4.1)\.
- Diana Andrushiaet al\.\(2022\)A\. Diana Andrushia, T\. Mary Neebha, A\. Trephena Patricia, S\. Umadevi, N\. Anand, and A\. VarshneyImage\-based disease classification in grape leaves using convolutional capsule network\.Soft Comput\.27\(3\),pp\. 1457–1470\.Note:[https://doi\.org/10\.1007/s00500\-022\-07446\-5](https://doi.org/10.1007/s00500-022-07446-5)External Links:ISSN 1432\-7643,[Link](https://doi.org/10.1007/s00500-022-07446-5)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- El Akhalet al\.\(2023\)H\. El Akhal, A\. Ben Yahya, N\. Moussa, and A\. El Belrhiti El AlaouiA novel approach for image\-based olive leaf diseases classification using a deep hybrid model\.Ecological Informatics77,pp\. 102276\.Note:[https://doi\.org/10\.1016/j\.ecoinf\.2023\.102276](https://doi.org/10.1016/j.ecoinf.2023.102276)External Links:ISSN 1574\-9541Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- El Akhalet al\.\(2025\)H\. El Akhal, A\. B\. Yahya, and A\. E\. B\. El AlaouiPositive discrimination of minority classes through data generation and distribution: a case study in olive disease classification\.Engineering Applications of Artificial Intelligence139,pp\. 109646\.Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Enkvetchakul and Surinta \(2022\)P\. Enkvetchakul and O\. SurintaEffective data augmentation and training techniques for improving deep learning in plant leaf disease recognition\.Applied Science and Engineering Progress15\(3\),pp\. 3810–3810\.Note:[https://doi\.org/10\.14416/j\.asep\.2021\.01\.003](https://doi.org/10.14416/j.asep.2021.01.003)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p4.1)\.
- Fazariet al\.\(2021\)A\. Fazari, O\. J\. Pellicer\-Valero, J\. Gómez\-Sanchıs, B\. Bernardi, S\. Cubero, S\. Benalia, G\. Zimbalatti, and J\. BlascoApplication of deep convolutional neural networks for the detection of anthracnose in olives using vis/nir hyperspectral images\.Computers and Electronics in Agriculture187,pp\. 106252\.Note:[https://doi\.org/10\.1016/j\.compag\.2021\.106252](https://doi.org/10.1016/j.compag.2021.106252)External Links:ISSN 0168\-1699,[Link](https://www.sciencedirect.com/science/article/pii/S0168169921002696)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Haikalet al\.\(2024\)A\. L\. A\. Haikal, N\. Yudistira, and A\. RidokComprehensive mixed\-based data augmentation for detection of rice leaf disease in the wild\.Crop Protection184,pp\. 106816\.Cited by:[§2\.1](https://arxiv.org/html/2610.35984#S2.SS1.p8.1),[Table 1](https://arxiv.org/html/2610.35984#S2.T1.2.8.1.1.1)\.
- Hasanet al\.\(2020\)R\. I\. Hasan, S\. M\. Yusuf, and L\. AlzubaidiReview of the state of the art of deep learning for plant diseases: a broad analysis and discussion\.Plants9\(10\),pp\. 1302\.Note:[https://doi\.org/10\.3390/plants9101302](https://doi.org/10.3390/plants9101302)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Jayasumanaet al\.\(2024\)S\. Jayasumana, S\. Ramalingam, A\. Veit, D\. Glasner, A\. Chakrabarti, and S\. KumarRethinking fid: towards a better evaluation metric for image generation\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,pp\. 9307–9315\.Cited by:[§3\.2](https://arxiv.org/html/2610.35984#S3.SS2.p14.1)\.
- Karthiket al\.\(2020\)R\. Karthik, M\. Hariharan, S\. Anand, P\. Mathikshara, A\. Johnson, and R\. MenakaAttention embedded residual cnn for disease detection in tomato leaves\.Applied Soft Computing86,pp\. 105933\.Note:[https://doi\.org/10\.1016/j\.scp\.2022\.100793](https://doi.org/10.1016/j.scp.2022.100793)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Karthikeyan and Raja \(2023\)M\. Karthikeyan and D\. RajaDeep transfer learning enabled densenet model for content based image retrieval in agricultural plant disease images\.Multimedia Tools and Applications82\(23\),pp\. 36067–36090\.Cited by:[§2\.1](https://arxiv.org/html/2610.35984#S2.SS1.p7.1),[Table 1](https://arxiv.org/html/2610.35984#S2.T1.2.7.1.1.1)\.
- Kauret al\.\(2022\)P\. Kaur, S\. Harnal, V\. Gautam, M\. P\. Singh, and S\. P\. SinghAn approach for characterization of infected area in tomato leaf disease based on deep learning and object detection technique\.Engineering Applications of Artificial Intelligence115,pp\. 105210\.Note:[https://doi\.org/10\.1016/j\.engappai\.2022\.105210](https://doi.org/10.1016/j.engappai.2022.105210)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Khattaket al\.\(2021\)A\. Khattak, M\. U\. Asghar, U\. Batool, M\. Z\. Asghar, H\. Ullah, M\. Al\-Rakhami, and A\. GumaeiAutomatic detection of citrus fruit and leaves diseases using deep neural network model\.IEEE Access9\(\),pp\. 112942–112954\.Note:[https://doi\.org/10\.1109/ACCESS\.2021\.3096895](https://doi.org/10.1109/ACCESS.2021.3096895)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Ksibiet al\.\(2022\)A\. Ksibi, M\. Ayadi, B\. O\. Soufiene, M\. M\. Jamjoom, and Z\. UllahMobiRes\-net: a hybrid deep learning model for detecting and classifying olive leaf diseases\.Applied Sciences12\(20\),pp\. 10278\.Note:[https://doi\.org/10\.3390/app122010278](https://doi.org/10.3390/app122010278)Cited by:[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p6.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.6.1.1.1)\.
- Lewy and Mańdziuk \(2023\)D\. Lewy and J\. MańdziukAn overview of mixing augmentation methods and augmentation strategies\.Artificial Intelligence Review56\(3\),pp\. 2111–2169\.Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p4.1)\.
- Liet al\.\(2020a\)J\. Li, S\. Xie, Z\. Chen, H\. Liu, J\. Kang, Z\. Fan, and W\. LiA shallow convolutional neural network for apple classification\.IEEE Access8\(\),pp\. 111683–111692\.Note:[https://doi\.org/10\.1109/ACCESS\.2020\.3002882](https://doi.org/10.1109/ACCESS.2020.3002882)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Liet al\.\(2020b\)L\. Li, Y\. Fan, M\. Tse, and K\. LinA review of applications in federated learning\.Computers & Industrial Engineering149,pp\. 106854\.Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p3.1)\.
- Liuet al\.\(2018\)B\. Liu, Y\. Zhang, D\. He, and Y\. LiIdentification of apple leaf diseases based on deep convolutional neural networks\.Symmetry10\(1\)\.Note:[https://doi\.org/10\.3390/sym10010011](https://doi.org/10.3390/sym10010011)External Links:[Link](https://www.mdpi.com/2073-8994/10/1/11),ISSN 2073\-8994Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Liu \(2020\)S\. Y\. LiuArtificial intelligence \(ai\) in agriculture\.IT Professional22\(3\),pp\. 14–15\.Note:[https://doi\.org/10\.1109/MITP\.2020\.2986121](https://doi.org/10.1109/MITP.2020.2986121)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p2.1)\.
- Mahmood and Alsalem \(2024\)M\. A\. Mahmood and K\. AlsalemOlive leaf disease detection via wavelet transform and feature fusion of pre\-trained deep learning models\.\.Computers, Materials & Continua78\(3\),pp\. 13723\.Note:[https://doi\.org/10\.32604/cmc\.2024\.047604](https://doi.org/10.32604/cmc.2024.047604)Cited by:[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p9.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.9.1.1.1)\.
- Mamba Kabalaet al\.\(2023\)D\. Mamba Kabala, A\. Hafiane, L\. Bobelin, and R\. CanalsImage\-based crop disease detection with federated learning\.Scientific Reports13\(1\),pp\. 19220\.Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p3.1)\.
- Mhalaet al\.\(2025\)P\. Mhala, A\. Bilandani, and S\. SharmaEnhancing crop productivity with fined\-tuned deep convolution neural network for potato leaf disease detection\.Expert Systems with Applications267,pp\. 126066\.Cited by:[§2\.1](https://arxiv.org/html/2610.35984#S2.SS1.p3.1),[Table 1](https://arxiv.org/html/2610.35984#S2.T1.2.3.1.1.1)\.
- Panet al\.\(2024\)H\. Pan, Y\. Guo, M\. Yu, and J\. ChenEnhanced long\-tailed recognition with contrastive cutmix augmentation\.IEEE Transactions on Image Processing33\(1\),pp\. 1–12\.Note:[https://doi\.org/10\.1109/TIP\.2024\.3425148](https://doi.org/10.1109/TIP.2024.3425148)Cited by:[§1](https://arxiv.org/html/2610.35984#S1.p4.1)\.
- Radfordet al\.\(2015\)A\. Radford, L\. Metz, and S\. ChintalaUnsupervised representation learning with deep convolutional generative adversarial networks\.arXiv preprint arXiv:1511\.06434\.Cited by:[§3\.2](https://arxiv.org/html/2610.35984#S3.SS2.p11.1)\.
- Sathyaet al\.\(2025\)K\. Sathya, A\. Balakrishnan, P\. Baskaran, and R\. A\. KumarEnhancing plant disease detection using attention\-augmented residual networks and faster region\-convolutional networks\.IEEE Access\.Cited by:[§2\.1](https://arxiv.org/html/2610.35984#S2.SS1.p5.1),[Table 1](https://arxiv.org/html/2610.35984#S2.T1.2.5.1.1.1)\.
- Singh and Sanodiya \(2023\)R\. S\. R\. Singh and R\. K\. SanodiyaZero\-shot transfer learning framework for plant leaf disease classification\.IEEE Access11,pp\. 143861–143880\.Cited by:[§2\.1](https://arxiv.org/html/2610.35984#S2.SS1.p4.1),[Table 1](https://arxiv.org/html/2610.35984#S2.T1.2.4.1.1.1)\.
- Tabbakh and Barpanda \(2023\)A\. Tabbakh and S\. S\. BarpandaA deep features extraction model based on the transfer learning model and vision transformer “tlmvit” for plant disease classification\.IEEE Access11,pp\. 45377–45392\.Cited by:[§2\.1](https://arxiv.org/html/2610.35984#S2.SS1.p2.1),[Table 1](https://arxiv.org/html/2610.35984#S2.T1.2.2.1.1.1)\.
- Tan and Le \(2021\)M\. Tan and Q\. V\. LeEfficientNetV2: smaller models and faster training\.CoRRabs/2104\.00298\.External Links:[Link](https://arxiv.org/abs/2104.00298),2104\.00298Cited by:[§3\.4](https://arxiv.org/html/2610.35984#S3.SS4.p1.1)\.
- Tantithamthavornet al\.\(2016\)C\. Tantithamthavorn, S\. McIntosh, A\. E\. Hassan, and K\. MatsumotoAn empirical comparison of model validation techniques for defect prediction models\.IEEE Transactions on Software Engineering43\(1\),pp\. 1–18\.Note:[https://doi\.org/10\.1016/j\.aci\.2018\.08\.003](https://doi.org/10.1016/j.aci.2018.08.003)Cited by:[§3\.5](https://arxiv.org/html/2610.35984#S3.SS5.p1.1)\.
- Tantithamthavornet al\.\(2018\)C\. Tantithamthavorn, S\. McIntosh, A\. E\. Hassan, and K\. MatsumotoThe impact of automated parameter optimization on defect prediction models\.IEEE Transactions on Software Engineering45\(7\),pp\. 683–711\.Note:[https://doi\.org/10\.1109/TSE\.2018\.2794977](https://doi.org/10.1109/TSE.2018.2794977)Cited by:[§3\.5](https://arxiv.org/html/2610.35984#S3.SS5.p1.1)\.
- Uğuz and Uysal \(2021\)S\. Uğuz and N\. UysalClassification of olive leaf diseases using deep convolutional neural networks\.Neural computing and applications33\(9\),pp\. 4133–4149\.Note:[https://doi\.org/10\.1007/s00521\-020\-05235\-5](https://doi.org/10.1007/s00521-020-05235-5)Cited by:[§2\.2](https://arxiv.org/html/2610.35984#S2.SS2.p3.1),[Table 2](https://arxiv.org/html/2610.35984#S2.T2.2.3.1.1.1)\.
- Wuet al\.\(2023\)Q\. Wu, X\. Ma, H\. Liu, C\. Bi, H\. Yu, M\. Liang, J\. Zhang, Q\. Li, Y\. Tang, and G\. YeA classification method for soybean leaf diseases based on an improved convnext model\.Scientific Reports13\(1\),pp\. 19141\.Cited by:[§2\.1](https://arxiv.org/html/2610.35984#S2.SS1.p6.1),[Table 1](https://arxiv.org/html/2610.35984#S2.T1.2.6.1.1.1)\.
- Yunet al\.\(2019\)S\. Yun, D\. Han, S\. J\. Oh, S\. Chun, J\. Choe, and Y\. YooCutmix: regularization strategy to train strong classifiers with localizable features\.InProceedings of the IEEE/CVF international conference on computer vision,pp\. 6023–6032\.Cited by:[§3\.2](https://arxiv.org/html/2610.35984#S3.SS2.p3.1)\.
- Zhanget al\.\(2017\)H\. Zhang, M\. Cisse, Y\. N\. Dauphin, and D\. Lopez\-PazMixup: beyond empirical risk minimization\.arXiv preprint arXiv:1710\.09412\.Cited by:[§3\.2](https://arxiv.org/html/2610.35984#S3.SS2.p7.1)\.