Source-linked AI summary
Enhancing Deep Learning Model Performance through Data Generation Strategies: Classification of Olive Tree Diseases
Hicham El Akhal, Abderrahim Waga, Aissa Ben Yahya, Abdelbaki El Belrhiti El Alaoui
TL;DR
Insufficient, imbalanced, and insufficiently diverse data challenge deep-learning models for plant-disease classification. This study combines traditional augmentation, MixUp, CutMix, and DCGAN in three olive-disease data-distribution scenarios; the combined approach generally performs best, particularly in S1 and S2.
Problem
Insufficiently diverse and balanced datasets pose a challenge because they can contribute to overfitting and limited model generalization.
Source: §1 · p. 2Method
The study combines traditional transformations, MixUp, CutMix, and DCGAN, evaluating generated data in scenarios that balance classes, increase samples per class, or expand all classes while preserving balance.
Source: §1 · p. 2, §2.5 · p. 6Results
The combined traditional, MixUp, CutMix, and DCGAN approach generally gives the best overall performance, particularly in S1 and S2.
Source: §5 · p. 15Takeaways & Limitations
The combined strategies increase the quantity, variability, and representativeness of training data, according to the study's discussion.
Source: §5 · p. 15Performance depends on dataset quality and balance, while DCGAN resource demands may limit scalability and generated-data volume and hyperparameters require tuning.
Source: §6.3 · p. 17
Abstract
from arXiv · showhide
Deep learning models are highly effective in a wide range of domains, particu larly 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 Convolu tional 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 perfor mance using the EfficientNetV2B0 architecture with pre-trained and fine-tuned weights. The experimental results show a substantial improvement in classifica tion accuracy, emphasizing the critical role of data diversity and class balance. These findings highlight the effectiveness and adaptability of the proposed strat egy of data generation and distribution for enhancing dataset quality, offering promising applications in plant disease detection and precision agriculture.
1 Introduction
Plant disease diagnosis matters for agricultural production, but deep-learning classification is challenged by insufficient, imbalanced, and insufficiently diverse data. This study combines data-generation methods and evaluates them across three olive-disease dataset-distribution scenarios.
1 Introduction
- Plant diseases threaten crop production and food security, while manual visual diagnosis can be laborious, subjective, and error-prone.The introduction notes that delayed diagnosis can reduce productivity, particularly where agricultural specialists are difficult to reach.Source: §1 · p. 2
- Insufficiently diverse and balanced datasets can contribute to overfitting and limited model generalization.The study takes a data-generation approach to this challenge rather than the federated-learning approach mentioned in the introduction.Source: §1 · p. 2
- The study combines traditional data-generation methods with CutMix, MixUp, and DCGANs to improve CNN-based olive-disease classification.The introduction frames these techniques as means of generating enriched training data.Source: §1 · p. 2
- The framework evaluates balancing original classes, increasing samples per class, and expanding all classes while preserving balance.These are the study's three scenarios: S1, S2, and S3, respectively.Source: §1 · p. 2
- The proposed generative-diversity strategy targets small-data settings and aims to improve the precision and usability of plant-disease diagnosis models.The introduction presents the integrated pipeline as relevant to sparse-data agricultural applications.Source: §1 · p. 2
2 Literature Review
The literature review surveys augmentation and synthetic-data approaches for plant and olive disease classification, then identifies limited evaluation across data-distribution scenarios as a gap. The study responds with a hybrid generation framework and examines its use for olive-disease classification in low-data settings.
2.1 Data Generation and Augmentation Strategies
- Prior plant-disease studies combine augmentation with varied models, including transfer learning, GAN-generated images, and composite augmentation methods.Examples include wheat disease classification with TLMViT, leaf-image generation with conditional GANs [27], and rice-disease classification with composite augmentation [30].Source: §2.1 · p. 3
2.2 Review of Olive Disease Classification Techniques
- The review reports 93.48% accuracy for the proposed work and frames it as a robustness and generalizability trade-off against some conventional methods.It cautions that accuracy differences may also reflect differences in the datasets used.Source: §2.2 · p. 5, Table 2 · p. 5
2.3 Research Gap Analysis and Contributions
- The review identifies limited study of multiple data-distribution scenarios and systematic integration of classical and generative techniques as research gaps.The proposed approach combines traditional augmentation, CutMix, MixUp, and DCGAN across S1, S2, and S3.Source: §2.3 · pp. 5–6
2.4 Research Contributions
- The stated contributions include a hybrid generation framework, minority-class sample generation, dataset expansion, and evaluation across three distribution scenarios.The framework is presented as targeting diversity, representativeness, model bias, and overfitting in data-scarce contexts.Source: §2.4 · p. 6
2.5 Problem Description
- The problem description highlights class imbalance, overfitting risks from excessive or unwarranted augmentation, and the need to select methods for each dataset and task.The study compares augmentation techniques with the stated aim of improving generalization and avoiding overfitting in low-data scenarios.Source: §2.5 · pp. 6–7
3 Solution Approach
The approach generates and distributes olive-leaf training data across three scenarios, then trains and evaluates EfficientNetV2B0 using an untouched test set. It also specifies the data sources, evaluation procedures, and computational setup.
3 Solution Approach
- Dataset A is expanded and split into training and validation sets across S1–S3; fine-tuned EfficientNetV2B0 is evaluated on untouched Dataset B.The workflow covers data preparation, model development, and evaluation.Source: §3 · p. 7, Figure 1 · p. 8
3.1.1 Data Collection
- Images from prior datasets and Moroccan farms (2020–2024) represent healthy leaves and disease categories; Table 3 summarizes image counts per class.Representative category images are shown in Figure 2.Source: §3.1.1 · p. 7, Table 3 · p. 7, Figure 2 · p. 9
3.2 Data Generation
- Traditional augmentation, CutMix, MixUp, and DCGAN generate olive-disease samples; FID measures distance between real and generated feature distributions.For DCGAN, the generator learns to produce realistic images while the discriminator distinguishes real from generated samples. Traditional transformations include rotation, flipping, and zooming.Source: §3.2 · pp. 8–10
3.3 Data Distribution
- S1 balances classes in the original dataset, S2 adds samples to each class, and S3 expands the dataset while balancing classes.The study compares these three data arrangements to examine their relationship with model performance.Source: §3.3 · p. 10
3.4 Model design
- The model uses pre-trained EfficientNetV2B0 with a custom four-layer fully connected classifier, Dropout after each layer, and a final softmax activation.The architecture is illustrated in Figure 3.Source: §3.4 · p. 10, Figure 3 · p. 11
3.5 Perfermence Evalution
- Evaluation includes accuracy, F1, precision, recall, MCC, and G-mean; accuracy is validated with five-fold cross-validation and strategies are grouped using Scott-Knott ESD.The study identifies accuracy as its main metric.Source: §3.5 · pp. 10–11
3.6 Computational Resources and Time Requirements
- Computational costs are presented for a Windows 10 and TensorFlow setup with an NVIDIA GeForce RTX 2060 GPU and Intel Xeon E5-2630 v3 CPU.The listed system has 46 GB RAM and 12 GB of dedicated GPU RAM.Source: Table 4 · p. 11, §3.6 · p. 11
4 Results Analysis
Results varied across the three data-distribution scenarios: the combined approach led in S2, while results in S1 and S3 favored the combined approach and CutMix, respectively. Table 6 reports particularly strong scores for the combined approach in imbalanced S2.
4 Results Analysis
- 93.09% accuracy was the S1 high for combined augmentations, a +2.10-point gain over baseline; all methods outperformed the 90.99% baseline.Source: Table 5 · p. 12, §4 · p. 12
- In S2, MixUp improved accuracy by +0.74 points, while traditional augmentation and CutMix underperformed by -0.23 and -0.28 points, respectively.Source: Table 5 · p. 12, §4 · p. 12
- In S3, CutMix led F1-score (92.28%) and MCC (89.48%), while combined augmentation yielded no accuracy gain (+0.00), suggesting saturation or redundancy.Source: §4 · pp. 12–13, Table 5 · p. 12
- In S2, the combined approach reached 95.99% accuracy, 93.32% F1-score, and 91.04% G-mean; DCGAN alone scored 93.35% F1-score.Source: §4 · pp. 12–13, Table 6 · p. 13
- Table 6 reports test-dataset scores for accuracy, recall, F1-score, G-mean, and MCC across olive-disease scenarios S1, S2, and S3 and generation methods.Source: §4 · p. 12, Table 6 · p. 13
- Figures 4–5 rank training datasets to compare their impact on model performance and accuracy.Source: §4 · pp. 13–14, Figure 4 · p. 14
5 Discussion
Combined data-generation techniques generally performed best in S1 and S2, while sensitivity to individual methods varied with dataset balance and volume. The discussion also cautions that artificial samples may be uninformative and describes quality checks for generated images.
5 Discussion
- Some artificial samples may be uninformative, so the effectiveness of advanced augmentation depends on the application scenario and justified use.Source: §5 · p. 15
- The combined use of traditional augmentation, MixUp, CutMix, and DCGAN generally gave the best overall performance, particularly in S1 and S2.The authors describe the techniques as complementary, increasing training-data quantity, variability, and representativeness.Source: §5 · p. 15
5.1 Sensitivity Analysis
- In S1, all augmentation methods showed significant gains, with CutMix at +0.91 and DCGAN at +1.25.Source: §5.1 · p. 16
- In S2, CutMix was less sensitive or negatively impacted at -0.28, while MixUp and DCGAN remained positive at +0.74 and +0.34, respectively.Source: §5.1 · p. 16
- In S3, gains were more moderate or zero for some combinations, although CutMix retained an advantage of +1.30.The discussion suggests saturation when balanced, voluminous data already provide sufficient diversity.Source: §5 · p. 15, §5.1 · p. 16
- The authors say method sensitivity depends on dataset volume, balance, and intra-class diversity; DCGAN and CutMix stood out as efficient and stable.Source: §5.1 · p. 16
5.2 Ethical Considerations in Data Generation Techniques
- For DCGAN images, the authors used FID to evaluate samples and retain the highest-quality ones, alongside visualizing generated data and monitoring class distributions.Source: §5.2 · p. 16
6 Conclusions, limitations, and future research
The study reports improved olive disease classification with its data-generation strategy, while noting dataset, computational, and parameter constraints. It proposes exploring additional generation methods and validating the approach in real agricultural settings.
6.1 Findings
- The proposed data-generation strategy improved performance by 2.48% compared with the no-augmentation baseline.Source: §6.1 · p. 17
6.2 Future Research
- Future work could explore more sophisticated generation methods and evaluate practical viability, robustness, and adaptability in real agricultural settings.Source: §6.2 · p. 17
6.3 Limitations of Research
- Performance depends on dataset quality and balance; DCGAN resource demands can limit scalability, and generated-data volume and model hyperparameters require tuning.Source: §6.3 · p. 17