A dual oscillatory particle swarm optimization for tomato disease classification

The difficulties of tomato disease recognition

Although automated disease recognition technology shows great potential, its deployment in real-world, complex agricultural environments still faces numerous severe technical challenges:

Complex imaging conditions are the primary external factor affecting recognition accuracy. In natural settings such as fields or greenhouses, the image acquisition process is highly susceptible to issues like variations in lighting intensity, shadow interference, occlusion by healthy leaves, stems, or soil in the background, as well as image blur and noise caused by wind or camera shake. These factors significantly increase the difficulty of accurately extracting disease features from raw images (Albahli & Nawaz, 2022).

The inherent ambiguity and variability of disease symptoms constitute an intrinsic recognition challenge. On the one hand, different diseases may exhibit highly similar visual features (i.e., low inter-class variance), such as various fungal diseases initially appearing as small brown or yellow spots. On the other hand, the same disease may show significant variations in the size, color, and morphology of its lesions at different stages of infection, on different tomato varieties, or under different environmental conditions (i.e., high intra-class variance). This poses a high demand on the model’s generalization ability and robustness (Nawaz et al., 2022).

Limitations in data and computational resources are practical bottlenecks constraining model performance. In many real-world application scenarios, researchers find it difficult to obtain large-scale, accurately labeled, and class-balanced image datasets that cover all disease categories and developmental stages. This leads to problems such as small-sample learning (Liu & Zhang, 2023) and data imbalance (Zhao et al., 2022), which severely restrict the training effectiveness of deep learning models. Furthermore, many high-performance deep learning models have complex architectures and a massive number of parameters, requiring substantial computational resources. This hinders their deployment on resource-constrained mobile platforms like smartphones and drones, limiting the practical application value of the technology in the field (Ullah et al., 2023; Ahmed et al., 2022).

Traditional methods for tomato leaf disease detection

In the field of automatic plant disease recognition, the technological paradigm has shifted from traditional machine learning, which relies on manual feature engineering, to deep learning, which is dominated by end-to-end learning. This section provides an in-depth analysis of these two categories of methods, which constitute the current research baselines, and elucidates their respective advantages and remaining limitations (Thanammal Indu & Suja Priyadharsini, 2022).

Early research efforts were primarily focused on traditional machine learning methods. For instance, Kurmi & Gangwar (2022) proposed a context-based image classification algorithm that first localizes the leaf region using an Lab* color space transformation, then extracts handcrafted features including Bag of Visual Words and Fisher vectors, and finally uses classifiers like logistic regression or multilayer perceptrons for determination. Similarly, to achieve accurate detection of grape leaf diseases, Shantkumari & Uma (2023) concentrated on extracting high-quality histogram and extended histogram features to capture the fine-grained structure and boundary information of lesions, feeding them into their proposed Convolutional Neural Networks (CNN) and an improvised K-Nearest Neighbor (IKNN) model. The common thread in these methods is their core pipeline of “manual feature design + classic classifier.” However, this paradigm suffers from a fundamental flaw: it heavily relies on the researcher’s domain expertise, the development process is tedious, and the designed features exhibit poor generalization, being extremely sensitive to environmental variations like lighting and background, which can easily lead to a drop in classification accuracy due to irrelevant or redundant information (Rehman et al., 2023).

To overcome the limitations of traditional methods, deep learning approaches, particularly CNNs, have rapidly become the new, more powerful baseline, as they can automatically learn hierarchical and discriminative features directly from raw pixels (Yao et al., 2023). In this direction, Paymode and Malode (Paymode & Malode, 2022) applied the classic Visual Geometry Group (VGG) network architecture, using transfer learning to classify tomato and grape leaf diseases, thereby validating the effectiveness of pre-trained models in the agricultural domain. To further enhance model performance and adaptability to complex scenarios, researchers have made numerous improvements. For example, to address the problem of precise lesion localization, Nawaz et al. (2022) proposed a Faster-Region-based Convolutional Neural Network (RCNN) model based on ResNet-34 and innovatively integrated a Convolutional Block Attention Module (CBAM) into the feature extractor to enhance the model’s focus on key lesion features. Also aiming to improve feature discrimination, Stephen, Punitha & Chandrasekar (2023) introduced a self-attention mechanism into the ResNet architecture for the classification of rice leaf diseases.

Meanwhile, to tackle the high computational cost of deep models and their difficult deployment on mobile devices, Ahmed et al. (2022) proposed a solution based on the lightweight network MobileNetV2, significantly reducing model size and computational complexity while maintaining high accuracy. Ullah et al. (2023) went a step further by designing a custom lightweight network, DTomatoDNet, with only 19 learnable layers specifically for tomato disease recognition, intended for easy integration into mobile applications.

Metaheuristic (Guo et al., 2024a, 2022) algorithms are considered an important method for image classification using texture features (Guo et al., 2024b; Shi et al., 2022). The stochastic nature of metaheuristics enables effective exploration of non-convex optimization landscapes, avoiding local optima that frequently trap gradient-based methods when optimizing texture feature combinations (Guo et al., 2023a, 2023b). Furthermore, metaheuristic approaches demonstrate superior capability in simultaneously addressing multiple objectives inherent in texture classification tasks, including feature subset selection, classifier hyperparameter tuning, and computational efficiency optimization (Guo et al., 2025; Tian et al., 2022).

Furthermore, to address the prevalent issues of data scarcity and class imbalance in agriculture, Zhao et al. (2022) proposed a two-stage generative adversarial network called DoubleGAN to generate high-quality synthetic images of diseased leaves for augmenting and balancing the training set. In their hybrid deep learning method, Islam et al. (2022) also employed a Conditional GAN (CGAN) to generate synthetic images to cope with data imbalance and noisy labels.

Dual oscillatory particle swarm optimization (DOPSO)

The DOPSO algorithm is proposed in this part. DOPSO addresses the fundamental limitations of conventional PSO, particularly premature convergence and insufficient population diversity, by implementing a sophisticated multi-layered search strategy.

The term “Dual Oscillatory” reflects the two core mechanisms of this algorithm. “Dual” refers to the dual subgroup architecture, in which the entire particle swarm is evenly divided into two independent subgroups. Each subgroup maintains its own global best solution and conducts parallel search operations. This separation ensures that the two subgroups can explore different regions of the search space simultaneously, thereby reducing the risk of the entire population converging to a single local optimum. “Oscillatory” describes the anti-phase oscillatory search strategy applied to the two subgroups. Specifically, oscillatory coefficients based on trigonometric functions are introduced into the velocity update process, and these coefficients are designed to operate in opposite phases for the two subgroups. This complementary oscillation pattern maintains continuous population diversity throughout the optimization process.

The algorithm operates through three primary mechanisms: (1) dual subgroup parallel exploration, where the particle swarm is divided into two independent subgroups that maintain separate global best solutions; (2) anti-phase oscillatory operators that introduce controlled perturbations to particle position updates using trigonometric functions; and (3) periodic information exchange between subgroups to facilitate global knowledge sharing while preserving local search independence. The mathematical foundation of DOPSO extends the classical PSO velocity-position update equations by incorporating an oscillatory coefficient that varies dynamically with iteration count. This oscillatory behavior creates a natural balance between exploration and exploitation phases, enabling the algorithm to maintain population diversity in early iterations while converging to high-quality solutions in later stages. DOPSO demonstrates particular effectiveness in high-dimensional optimization problems and machine learning hyperparameter tuning, where traditional PSO often suffers from stagnation in local optima. The algorithm’s dual-layer architecture provides robustness against premature convergence while maintaining computational efficiency comparable to standard PSO implementations. The theoretical contributions of DOPSO include: enhanced global search capability through dual subgroup diversity maintenance, improved convergence characteristics via oscillatory perturbation mechanisms, and a novel information sharing paradigm that balances exploration and exploitation throughout the optimization process.

Dual subgroup architecture

The DOPSO algorithm partitions the swarm S of size $2N$ into two equal subgroups:

(1) $$S_1=\{P_1,P_2,\dots ,P_N\},S_2=\{P_{N+1},P_{N+2},\dots ,P_{2N}\},$$where S is the particle swarm, $2N$ is the total number of particles, $S_1$ and $S_2$ are the first and second subgroups each containing N particles, respectively, and $P_i$ represents the $i$th particle.

Each subgroup maintains an independent global best solution:

(2) $$g_{best}^{(1)}=\arg \underset{i\in S_1}{min}f(p_{best,i}),g_{best}^{(2)}=\arg \underset{i\in S_2}{min}f(p_{best,i}),$$where $g_{best}^{(1)}$ and $g_{best}^{(2)}$ are the global best positions of subgroups 1 and 2, respectively, $f(\cdot )$ is the objective function, and $p_{best,i}$ is the personal best position of particle $i$.

Oscillatory position update

The core innovation of DOPSO lies in the modified position update equation incorporating an oscillatory coefficient:

For Subgroup 1:

(3) $$x_i^{t+1}=x_i^t+{\alpha }_1^{(t)}\cdot v_i^{t+1},i\in S_1.$$

For Subgroup 2:

(4) $$x_j^{t+1}=x_j^t+{\alpha }_2^{(t)}\cdot v_j^{t+1},j\in S_2,$$where $x_i^t$ and $x_i^{t+1}$ are the position of particle $i$ at iterations $t$ and $t+1$, respectively, $v_i^{t+1}$ is the velocity of particle $i$ at iteration $t+1$, and ${\alpha }_1^{(t)}$ and ${\alpha }_2^{(t)}$ are the oscillatory coefficients for subgroups 1 and 2 at iteration $t$, respectively.

The oscillatory coefficients are defined as:

(5) $${\alpha }_1^{(t)}=(-1{)}^t\cdot \cos \left(\frac{\pi }{t}\right)$$

(6) $${\alpha }_2^{(t)}=(-1{)}^{t+1}\cdot \cos \left(\frac{\pi }{t}\right),$$where $t$ is the current iteration number, $(-1{)}^t$ provides phase alternation, and $\cos (\pi /t)$ controls the amplitude decay over iterations.

Velocity update with dual global guidance

The velocity update equations for each subgroup are:

For Subgroup 1:

(7) $$v_i^{t+1}=w\cdot v_i^t+c_1\cdot r_1\cdot (p_{best,i}-x_i^t)+c_2\cdot r_2\cdot (g_{best}^{(1)}-x_i^t).$$

For Subgroup 2:

(8) $$v_j^{t+1}=w\cdot v_j^t+c_1\cdot r_1\cdot (p_{best,j}-x_j^t)+c_2\cdot r_2\cdot (g_{best}^{(2)}-x_j^t),$$where $w$ is the inertia weight, $c_1$ and $c_2$ are the cognitive and social acceleration coefficients, respectively, $r_1$ and $r_2$ are random numbers uniformly distributed in $[0,1]$, $p_{best,i}$ is the personal best position of particle $i$, and $g_{best}^{(1)}$ and $g_{best}^{(2)}$ are the respective global best positions of each subgroup.

Information exchange mechanism

At predetermined intervals, the global best solutions are exchanged between subgroups:

(9) $$\mathrm{i}\mathrm{f}t\mathrm{m}\mathrm{o}\mathrm{d}{T}_{exchange}=0:g_{best}^{(1)}\leftrightarrow g_{best}^{(2)},$$where $T_{exchange}=\lfloor T/10\rfloor$, T is the maximum iteration count, and $\lfloor \cdot \rfloor$ denotes the floor function.

Oscillatory coefficient properties

The oscillatory coefficients exhibit the following mathematical characteristics:

Phase alternation:

(10) $${\alpha }_1^{(t)}\cdot {\alpha }_2^{(t)}=-{\cos }^2\left(\frac{\pi }{t}\right)\le 0.$$

Amplitude evolution:

(11) $$\left|{\alpha }_1^{(t)}\right|=\left|{\alpha }_2^{(t)}\right|=\left|\cos \left(\frac{\pi }{t}\right)\right|.$$

Convergence behavior:

(12) $$\underset{t\to \mathrm{\infty }}{lim}{\alpha }_1^{(t)}=\pm 1,\underset{t\to \mathrm{\infty }}{lim}{\alpha }_2^{(t)}=\mp 1,$$where the amplitude $| \cos (\pi /t)|$ decreases monotonically as $t$ increases, ensuring strong exploration in early iterations and refined exploitation in later stages.

DOPSO-LightGBM

LightGBM is a highly efficient gradient boosting framework that utilizes tree-based learning algorithms, developed by Microsoft Research. It employs novel techniques such as Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB) to significantly reduce computation time and memory usage while maintaining high accuracy. LightGBM constructs trees leaf-wise rather than level-wise, enabling faster training speed and lower memory consumption compared to traditional gradient boosting methods. The framework supports parallel and GPU learning, handles categorical features natively, and provides excellent performance in various machine learning competitions and real-world applications, particularly excelling in scenarios with large datasets and high-dimensional feature spaces. DOPSO-LightGBM represents an innovative hybrid approach that combines the advanced hyperparameter optimization capabilities of DOPSO with the efficient gradient boosting framework of LightGBM. This integration addresses the critical challenge of optimal hyperparameter configuration in LightGBM by employing DOPSO’s dual subgroup architecture to simultaneously optimize seven key parameters: learning rate, maximum depth, number of estimators, subsample ratio, feature sampling ratio, minimum split gain, and minimum child samples. The anti-phase oscillatory mechanism of DOPSO ensures comprehensive exploration of the hyperparameter space while preventing premature convergence to suboptimal configurations. Through periodic information exchange between subgroups and controlled perturbation dynamics, DOPSO-LightGBM achieves superior parameter tuning performance compared to conventional optimization methods, resulting in enhanced model accuracy and robustness for complex classification tasks such as plant disease recognition. To better illustrate the workflow of DOPSO, the pseudocode is presented in Algorithm 1.

Tomato disease classification with DOPSO-LightGBM

Tomato Disease Classification with DOPSO-LightGBM represents a comprehensive application of advanced machine learning optimization techniques to address the critical agricultural challenge of automated plant disease diagnosis. This approach leverages a dataset of tomato leaf images encompassing ten distinct categories, including healthy specimens and nine disease conditions such as bacterial spot, early blight, late blight, leaf mold, septoria leaf spot, spider mites, target spot, tomato mosaic virus, and yellow leaf curl virus. The methodology extracts ten carefully selected image features spanning color, texture, shape, and statistical characteristics, including red and green channel means, red-green ratio, Gray-level co-occurrence matrix (GCLM) contrast and energy, Hue, Saturation and Value (HSV) means, intensity standard deviation, edge density, and color variance. These features are specifically designed to capture the visual manifestations of disease symptoms in tomato leaves, providing discriminative information for accurate classification. The DOPSO-LightGBM framework for tomato disease classification integrates the robust gradient boosting capabilities of LightGBM with the enhanced optimization power of DOPSO to achieve superior diagnostic performance. The system employs DOPSO’s dual subgroup oscillatory mechanism to optimize seven critical LightGBM hyperparameters simultaneously, ensuring optimal model configuration for the multi-class classification task. Through the anti-phase perturbation strategy and periodic information exchange between subgroups, the optimization process effectively explores the hyperparameter space while maintaining population diversity, leading to improved convergence to high-quality parameter combinations. This optimized DOPSO-LightGBM model demonstrates enhanced accuracy, precision, recall, and F1-score metrics compared to conventional optimization approaches, providing farmers and agricultural specialists with a reliable, automated tool for early disease detection and management, ultimately contributing to improved crop yield and food security.

The tomato disease classification pipeline begins with the acquisition of tomato leaf images representing ten distinct categories: one healthy class and nine disease conditions including bacterial spot, early blight, late blight, leaf mold, septoria leaf spot, spider mites, target spot, tomato mosaic virus, and yellow leaf curl virus. Each category contains 1,000 preprocessed images that have been standardized in terms of size and quality to ensure consistency across the dataset. The images are systematically organized in separate directories corresponding to their respective disease labels, facilitating automated processing and maintaining clear categorical boundaries for supervised learning. The feature extraction phase constitutes a critical component of the pipeline, where ten discriminative features are computed from each image to capture the essential visual characteristics of disease manifestations. The extraction process encompasses multiple feature domains: color features including red and green channel means, red-green ratio, and HSV hue and saturation means to capture disease-induced color variations; texture features such as GLCM contrast and energy that quantify surface irregularities and uniformity patterns; statistical features including intensity standard deviation and color variance that measure pixel distribution characteristics; and shape features represented by edge density that captures boundary information of diseased regions. This comprehensive feature set is specifically designed to encode the visual signatures of various tomato diseases while maintaining computational efficiency. The flowchart of the classification process is shown in Fig. 1.

Experimental data

The experimental framework utilizes a balanced tomato leaf pathology datasets containing 10,000 standardized images organized into ten classification categories, each represented by 1,000 samples to maintain statistical equilibrium. The datasets incorporates healthy control specimens and nine disease manifestations spanning bacterial, fungal, viral, and pest-induced pathologies. All image samples underwent preprocessing normalization protocols to achieve uniform dimensional and quality standards, enabling consistent feature extraction and robust performance assessment across pathological classifications.

The experimental comparison encompasses a comprehensive set of baseline methods representing different paradigms in machine learning for agricultural image classification. Traditional machine learning group includes Support Vector Machine (SVM), Logistic Regression K-Nearest Neighbors (KNN), and Random Forest. To evaluate the effectiveness of DOPSO, traditional PSO, Differential Evolution (DE), and Genetic Algorithm (GA) are also used to optimize the hyperparameters of LightGBM. LightGBM based group includes traditional LightGBM, PSO-LightGBM, DE-LightGBM and GA-LightGBM. To ensure fairness in the experiments, all population-based algorithms used the same parameter settings: a population size of 20 and a maximum of 30 iterations. For DOPSO, parameters are set: population size 20, subgroup size 10, inertia weight 0.7, $c_1$ = $c_2$ = 2.0.

All experiments were conducted on a workstation equipped with an Intel Core i7-10700 CPU (8 cores, 16 logical processors) running at 2.90 GHz, 32 GB of DDR4 RAM, and an NVIDIA GeForce RTX 3080 GPU with 4 GB of dedicated memory, Microsoft Windows 11 operation system.

The classification stage employs the DOPSO-LightGBM framework, where the extracted features serve as input to the optimized gradient boosting model. The DOPSO algorithm systematically explores the eight-dimensional hyperparameter space of LightGBM through its dual subgroup architecture and oscillatory perturbation mechanism. Details of hyperparameters are shown in Table 1.

The optimized model undergoes training on 80% of the dataset while utilizing 20% for testing, with additional validation splits employed during the optimization process. The final pipeline outputs classification predictions with associated confidence scores, enabling accurate disease identification and providing agricultural practitioners with reliable diagnostic capabilities for early intervention and crop management strategies. Details of datasets are shown in Table 2.

Prior to feature extraction, each input image undergoes a standardized preprocessing procedure. All images are resized to a uniform dimension of 256 * 256 pixels to ensure consistent feature computation across the datasets.

The feature extraction process constitutes a critical component of the tomato disease classification pipeline, where ten carefully selected features are computed from each image to capture the essential visual characteristics of disease manifestations. These features span multiple domains including color information, texture properties, statistical characteristics, and shape descriptors, providing comprehensive representation of disease-induced visual changes in tomato leaves.

The following ten features were systematically extracted from each tomato leaf image:

• Red channel mean (red_mean): Computes the average intensity of the red color channel across all pixels in the image. This feature is particularly valuable for detecting diseases that cause reddish-brown discoloration, such as bacterial spot and early blight, where pathological changes manifest as distinct red pigmentation patterns.

• Green channel mean (green_mean): Calculates the mean intensity of the green color channel, representing the overall greenness of the leaf tissue. Healthy leaves typically exhibit high green channel values, while diseased areas often show reduced green intensity due to chlorophyll degradation and tissue necrosis.

• Red-green ratio (rg_ratio): Derived as the ratio between red and green channel means, this feature effectively captures the color balance shift that occurs during disease progression. Diseased regions typically exhibit higher red-green ratios compared to healthy tissue, making this feature particularly discriminative for pathological condition identification.

• HSV hue mean (hsv_hue_mean): Represents the average hue value in the HSV color space, providing perceptually uniform color information that aligns with human visual perception. This feature captures subtle color variations that may not be apparent in RGB space, particularly useful for distinguishing between similar disease symptoms.

• HSV saturation mean (hsv_saturation_mean): Measures the average color saturation, indicating the purity and vividness of colors in the leaf image. Disease-affected areas often exhibit altered saturation levels due to tissue degradation and pigment changes.

• GLCM contrast (glcm_contrast): Computed using Gray-Level Co-occurrence Matrix analysis, this feature quantifies the local intensity variations and texture roughness. Diseased regions typically exhibit higher contrast values due to irregular surface patterns, lesion boundaries, and tissue deterioration.

• GLCM energy (glcm_energy): Measures texture uniformity and homogeneity through GLCM analysis. Healthy leaf surfaces demonstrate higher energy values indicating uniform texture, while diseased areas show reduced energy due to irregular patterns and surface disruptions.

• Intensity standard deviation (intensity_std): Calculates the standard deviation of grayscale pixel intensities, capturing the variability in brightness across the image. Diseased leaves typically exhibit higher intensity variation due to the presence of spots, lesions, and irregular pigmentation patterns.

• Color variance (color_variance): Computes the average variance across all three RGB color channels, providing a comprehensive measure of color distribution complexity. Disease-affected regions generally demonstrate increased color variance due to mixed healthy and pathological tissue areas.

• Edge density (edge_density): Determined using Canny edge detection followed by edge pixel ratio calculation, this feature captures the boundary information and structural complexity. Diseased leaves typically exhibit higher edge density due to lesion boundaries, spot perimeters, and irregular tissue structures characteristic of pathological conditions.

The heat map of experimental data is shown in Fig. 2.

Experimental results

The experimental results presented in Table 3 demonstrate the superior performance of the proposed DOPSO-enhanced algorithms compared to traditional machine learning approaches and standard PSO optimization methods. DOPSO-LightGBM achieved the highest classification accuracy of 82.25%, representing a significant improvement over conventional approaches. This outstanding performance can be attributed to the synergistic combination of LightGBM’s efficient gradient boosting framework and DOPSO’s enhanced hyperparameter optimization capabilities.

The experimental results presented in Table 3 demonstrate the superior performance of the proposed DOPSO-LightGBM method across all evaluation metrics for tomato disease classification. DOPSO-LightGBM achieved the highest accuracy of 82.25%, with corresponding precision of 82.21%, recall of 82.25%, and F1-score of 82.16%, while maintaining the lowest error rate of 17.75%. Among traditional machine learning approaches, Logistic Regression exhibited the poorest performance with 61.05% accuracy, while SVM and KNN achieved comparable results around 76%. Random Forest, as an ensemble method, demonstrated improved performance at 80.35%, highlighting the advantage of ensemble learning over single classifiers. The baseline LightGBM algorithm further enhanced classification accuracy to 81.65%, confirming the effectiveness of gradient boosting methods for this image classification task.

Details are shown in Figs. 3, 4, 5, 6, 7, 8, 9, and 10.

Analysis and discussion

The integration of metaheuristic optimization algorithms with LightGBM yielded varying degrees of improvement, with DOPSO demonstrating the most significant enhancement. Specifically, DOPSO-LightGBM outperformed PSO-LightGBM (81.75%), DE-LightGBM (81.70%), and GA-LightGBM (81.05%) by 0.50, 0.55, and 1.20 percentage points, respectively. This performance hierarchy validates the efficacy of the proposed dual oscillatory mechanism in hyperparameter optimization, achieving a 0.60 percentage point improvement over the baseline LightGBM. The consistent values across precision, recall, and F1-score metrics for all methods indicate balanced classification performance across different tomato disease categories, suggesting no significant bias toward specific classes. These results empirically confirm that the DOPSO algorithm’s anti-phase oscillatory operators and dual-subgroup architecture effectively address premature convergence issues, leading to superior hyperparameter configurations for agricultural image classification tasks.

The oscillatory coefficients create a dynamic balance between exploration and exploitation. In early iterations, the large amplitude oscillations promote extensive search space exploration, while the gradual amplitude decay in later iterations facilitates precise convergence to optimal hyperparameter configurations. The periodic information exchange between subgroups prevents algorithmic stagnation and ensures comprehensive coverage of the hyperparameter landscape.

Limitations and future research directions

Several limitations of the current study warrant consideration. First, although DOPSO demonstrates effectiveness in optimizing hyperparameters for gradient boosting frameworks, its current application scope remains confined to parameter tuning rather than structural optimization. Second, the computational cost per fitness evaluation is relatively high due to the requirement of training and validating machine learning models, which represents a primary direction for future optimization.

Future research will address these limitations through three strategic directions. First, we plan to extend the functionality of DOPSO from hyperparameter optimization to neural architecture search (NAS), thereby enabling automatic discovery of optimal network structures for disease classification tasks. This extension would allow the algorithm to simultaneously optimize both architectural configurations and training parameters, potentially yielding more effective classification models. Second, we aim to develop a lightweight, real-time detection framework suitable for unmanned aerial vehicle (UAV) deployment. This framework will integrate model compression techniques, such as knowledge distillation and network pruning, with efficient inference strategies to achieve real-time responsiveness on resource-constrained mobile devices. Third, we intend to generalize the DOPSO-based recognition methodology to diverse plant species, enabling comprehensive crop disease identification across a broader range of agricultural applications. This generalization will involve the construction of multi-species disease datasets and the development of transfer learning approaches to facilitate cross-domain knowledge sharing.