Tomato disease classification method and system based on plant assimilate optimization

Abstract

The application discloses a tomato disease classification method and system based on plant assimilate optimization, belongs to the field of computing intelligence and optimization technology, and solves the technical problems that the existing technology has information island among sub-regions, imbalance between global exploration and local development, and low high-dimensional search efficiency when dealing with high-dimensional or multi-objective problems, thereby reducing the classification efficiency and accuracy of tomato diseases. A D-dimensional search space of a classification model is divided into N equal sub-regions; the fitness of different particles in each sub-region is evaluated to determine a local growth center and a global growth center; based on the local growth center and the global growth center, the particles are used to search the optimal position of each sub-region to form a candidate solution set; the candidate solution set is subjected to clustering analysis to generate high-quality solutions; and after the high-quality solutions are processed, a Pareto optimal solution is obtained. The application is used for realizing efficient and accurate classification of tomato diseases.

Description

Technical Field

[0001] This invention relates to the field of computational intelligence and optimization technology, specifically to a method and system for classifying tomato diseases based on plant assimilate optimization. Background Technology

[0002] Tomatoes are one of the most consumed vegetables globally and are widely cultivated. However, because they are grown in open-field environments, various uncertainties during their growth process make tomatoes susceptible to a variety of pests and diseases, which are often difficult to detect, leading to significant yield losses.

[0003] With the continuous advancement of precision agriculture, crop pest and disease identification based on deep learning has become a key area of ​​research in modern agriculture. Convolutional neural networks (CNNs) are a classic deep learning model. While deep learning-based methods for identifying tomato pests have shown significant advantages, the performance of deep learning models is highly dependent on hyperparameter configurations, including learning rate, batch size, and optimizer selection. Traditional hyperparameter tuning mainly relies on manual trial and error or grid search, which suffers from low efficiency, high computational cost, and difficulty in obtaining the global optimum. Furthermore, existing optimization algorithms face numerous challenges when dealing with high-dimensional or multi-objective problems. Multi-objective evolutionary algorithms based on decomposition (MOEA / D) are among the closest existing technologies. They decompose multi-objective problems into a series of single-objective sub-problems using weight vectors and perform collaborative optimization. However, MOEA / D has the following main drawbacks:

[0004] I. Insufficient interaction between sub-problems, leading to local optima. MOEA / D decomposes a multi-objective optimization problem into a series of single-objective sub-problems and then optimizes these sub-problems simultaneously. However, MOEA / D's neighbor update mechanism has a limited scope, which may cause the algorithm to converge to a local region of the Pareto front, failing to fully explore the entire front. Specifically, if the initial population distribution is poor, or if some sub-problems in certain regions are trapped in local optima, these local optima may be difficult to escape due to a lack of information exchange with distant sub-problems. In this case, the local neighborhood update mechanism may prevent some sub-problems from obtaining effective solutions, resulting in incomplete Pareto front coverage. When facing high-dimensional problems, as the number of objectives increases, the angle between weight vectors decreases, reducing the discriminative power between sub-problems and leading to reduced efficiency in collaborative optimization.

[0005] Second, it has poor adaptability to complex Pareto front shapes. When a real Pareto front has complex geometric characteristics such as non-convexity, discontinuity, and extreme non-uniformity, standard MOEA / D often struggles to generate high-quality, fully covering approximate solution sets. MOEA / D defines these subproblems using a set of weight vectors. The most commonly used methods are the weighted sum method and the Chebyshev method. The weighted sum method can only mathematically find points on the convex hull. If the entire Pareto front is non-convex, the weighted sum method will completely fail to find solutions for the concave parts. It will skip the concave parts and directly find the endpoints of the convex hull. The Chebyshev method performs better in this regard and can theoretically handle non-convex fronts. Because of its property of maximizing the minimum satisfaction, it can approximate the boundary of any shape. Therefore, standard MOEA / D, if using the Chebyshev method, has some capability for purely "non-convex" problems. However, problems often arise in combination. For discontinuous or discrete Pareto fronts, such as when the Pareto front consists of several unconnected segments and isolated regions, the Chebyshev algorithm may get trapped in local optima. In summary, MOEA / D uses a weight vector that is uniformly distributed in the target space. However, if the actual Pareto front is irregularly shaped, this uniform weight projection onto the Pareto front will produce a highly non-uniform solution distribution.

[0006] Third, efficiency declines in high-dimensional spaces. As the dimensionality increases, the number of required subproblems increases dramatically, and the computational complexity rises significantly, leading to the "curse of dimensionality." As the number of objective functions M (i.e., the dimension of the objective) increases, the resources (number of subproblems N) needed to maintain the effectiveness of the algorithm need to grow exponentially.

[0007] In summary, existing technologies suffer from problems such as information silos between sub-regions, an imbalance between global exploration and local development, and low efficiency in high-dimensional search when dealing with high-dimensional or multi-objective problems. These issues lead to reduced efficiency and accuracy in tomato disease classification. Summary of the Invention

[0008] This invention solves the technical problems of reduced efficiency and accuracy in tomato disease classification caused by existing technologies when dealing with high-dimensional or multi-objective problems, such as information silos between sub-regions, imbalance between global exploration and local development, and low efficiency of high-dimensional search.

[0009] The tomato disease classification method based on plant assimilate optimization described in this invention includes the following steps:

[0010] Step 1: Obtain tomato disease image data as sample input;

[0011] Step 2: Divide the D-dimensional search space of the classification model into N equal sub-regions;

[0012] Step 3: Evaluate the different features within each sub-region. The fitness of the region is used to determine the local growth center of each sub-region and the global growth center of the D-dimensional search space;

[0013] Step 4: Based on the local growth center of each sub-region and the global growth center of the D-dimensional search space, use particle search to find the optimal position of each sub-region and form a candidate solution set;

[0014] Step 5: Perform cluster analysis on the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, obtain the Pareto optimal solution. Based on the Pareto optimal solution, construct a tomato disease classification model, input samples, and realize tomato disease classification.

[0015] Furthermore, in one embodiment of the present invention, the local growth center of each sub-region in step 3 is the particle with the highest fitness in each sub-region, and the global growth center of the D-dimensional search space is the particle with the highest fitness among the local growth centers of each sub-region.

[0016] Furthermore, in one embodiment of the present invention, step 4, which utilizes particle search to find the optimal location of each sub-region, includes an exploration phase and a development phase.

[0017] When R < 1, particle search is in the exploration stage; when R ≥ 1, particle search is in the development stage.

[0018] R = Ac / Cc;

[0019] Where R is the concentration ratio, Ac is the auxin concentration, simulating the decay characteristics with distance, and Cc is the cytokinin concentration, simulating the random pulse characteristics.

[0020] Furthermore, in one embodiment of the present invention, the exploration phase involves randomly searching for particles within each sub-region in the region, and the development phase involves searching for particles within each sub-region around the local growth center of the corresponding sub-region.

[0021] when ≥ At that time, the particle search is completed, among which, This represents the current iteration number. This represents the maximum number of iterations.

[0022] Furthermore, in one embodiment of the present invention, poor particles are annihilated in each iteration, and the same number of particles are randomly generated. The randomly generated particles are used for searching, and the optimal position of each sub-region is determined based on the fitness value of the positions searched by the particles.

[0023] Furthermore, in one embodiment of the present invention, when performing particle search during the exploration phase, the particle search location is:

[0024] ;

[0025] in, To explore the step size coefficient, This represents the (t+1)th generation position of the i-th particle. Let i be the position of the i-th particle in the t-th generation. A random number in the range of 0 to 1. This represents the boundary span of the current sub-region.

[0026] Furthermore, in one embodiment of the present invention, when performing particle search during the development phase, the particle search location is:

[0027] ;

[0028] in, A random number in the range of 0 to 1. It is the arctangent function. A random number in the range of 0 to 1. This represents the local growth center within the current subregion.

[0029] Furthermore, in one embodiment of the present invention, a global tabu table is established to record the positions searched by the particle. If the position searched by the particle has been recorded in the tabu table, a position that has not been searched is randomly generated in the corresponding sub-region.

[0030] Furthermore, in one embodiment of the present invention, in step 5, cluster analysis is performed on the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, Pareto optimal solutions are obtained, specifically as follows:

[0031] After performing cluster analysis on the candidate solution set, the candidate solution set is divided into multiple clusters. A high-quality solution is selected from each cluster, and a local particle search is performed based on the high-quality solution to obtain the Pareto optimal solution.

[0032] The present invention discloses a tomato disease classification system based on plant assimilate optimization. The system is constructed based on the aforementioned tomato disease classification method and includes the following modules:

[0033] The acquisition module acquires tomato disease image data as sample input;

[0034] The partitioning module divides the D-dimensional search space of the classification model into N equal sub-regions;

[0035] The evaluation module evaluates the different features within each sub-region. The fitness of the region is used to determine the local growth center of each sub-region and the global growth center of the D-dimensional search space;

[0036] The search module, based on the local growth center of each sub-region and the global growth center of the D-dimensional search space, uses particles to search for the optimal position of each sub-region, forming a candidate solution set;

[0037] The module constructs a cluster analysis of the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, the Pareto optimal solution is obtained. Based on the Pareto optimal solution, a tomato disease classification model is constructed. Input samples are then used to classify tomato diseases.

[0038] This invention solves the technical problems of existing technologies, such as information silos between sub-regions, imbalance between global exploration and local development, and low efficiency in high-dimensional search, which lead to reduced efficiency and accuracy in tomato disease classification when dealing with high-dimensional or multi-objective problems. Specific beneficial effects include:

[0039] 1. This invention proposes a tomato disease classification method based on plant assimilate optimization. For the first time, the complete biological mechanism of "plant assimilate allocation and transport" is introduced into the optimization method design, and a collaborative search framework of "partition-center-transport-redistribution" is constructed, which effectively solves the problems of information silos between sub-regions, imbalance between global exploration and local development, and low efficiency of high-dimensional search.

[0040] 2. This invention proposes a tomato disease classification method based on plant assimilate optimization. Based on the dynamic exploration-development conversion mechanism of the virtual auxin (Ac) to cytokinin (Cc) concentration ratio, it achieves smoother and more adaptive search behavior control.

[0041] 3. This invention proposes a tomato disease classification method based on plant assimilate optimization, which has excellent high-dimensional problem handling capabilities. It decomposes high-dimensional problems into multiple subspaces, significantly alleviating the "curse of dimensionality". By exploring different regions of the target space through partitioning, and combined with subsequent clustering and reinforcement search, it more comprehensively approximates the Pareto front of complex shapes. Attached Figure Description

[0042] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0043] Figure 1 This is a flowchart of the tomato disease classification method based on plant assimilate optimization as described in Implementation Method 1. Detailed Implementation

[0044] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0045] Implementation Method 1: A tomato disease classification method based on plant assimilate optimization, comprising the following steps:

[0046] Step 1: Obtain tomato disease image data as sample input;

[0047] Step 2: Divide the D-dimensional search space of the classification model into N equal sub-regions;

[0048] Step 3: Evaluate the different features within each sub-region. The fitness of the region is used to determine the local growth center of each sub-region and the global growth center of the D-dimensional search space;

[0049] Step 4: Based on the local growth center of each sub-region and the global growth center of the D-dimensional search space, use particle search to find the optimal position of each sub-region and form a candidate solution set;

[0050] Step 5: Perform cluster analysis on the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, obtain the Pareto optimal solution. Based on the Pareto optimal solution, construct a tomato disease classification model, input samples, and realize tomato disease classification.

[0051] In high-dimensional search spaces, traditional methods often divide the solution space into multiple sub-regions for parallel exploration. Due to the lack of effective information exchange mechanisms between sub-regions, local optima are difficult to share across regions, forming "information silos." This phenomenon hinders the discovery of the global optimum, as different sub-regions may become trapped in their own local optima and fail to improve collaboratively. Furthermore, most optimization methods need to maintain a balance between global exploration and local exploitation. In high-dimensional spaces, this balance is easily disrupted; overexploration leads to a decrease in convergence speed, while overexploitation may result in premature convergence. Simultaneously, as dimensionality increases, the search space expands exponentially, causing the algorithm to consume significant computational resources.

[0052] To address the aforementioned issues, this implementation provides a tomato disease classification method based on plant assimilate optimization. It simulates the mechanism by which photosynthetic products (assimilates) within plants are "regionally produced, supplied locally, and globally allocated" according to the needs of growth centers. A novel optimization method is proposed, which effectively solves the problems of information silos between sub-regions, imbalance between global exploration and local development, and low efficiency of high-dimensional search by dynamically managing the search space and coordinating the guidance of global and local growth centers.

[0053] The core of this implementation method lies in simulating four key biological principles in plant growth:

[0054] 1) Priority supply to the growth center; 2) Supply to nearby locations and same-side transportation; 3) No assimilation supply relationship between functional leaves; 4) Redistribution of photosynthetic products.

[0055] This implementation method initializes parameters and divides regions by simulating functional leaf partitioning:

[0056] 1. Initialize the population. Randomly generate an initial population P of size M in a D-dimensional search space.

[0057] 2. Divide the search space into sub-regions. Divide the entire D-dimensional search space into N equal sub-regions {R1, R2, ..., RN}. Each sub-region is initially allocated approximately M / N search particles (individuals).

[0058] Furthermore, in one embodiment of this implementation, the parameters are initialized as follows: maximum number of iterations MaxFEs = 300000; population particle count PopSize = 100; dimension is divided according to the specific problem, for example, 30 in the ZDT test set (a set of benchmark functions used to evaluate the performance of algorithms in the field of multi-objective optimization); number of sub-regions N = 5; decay coefficient σ = 0.1; annihilation period T = 20.

[0059] Furthermore, in one embodiment of this implementation, the EfficientNetV2 deep learning model is used as the basic classification model. The EfficientNetV2 network structure is optimized using a plant assimilate optimization algorithm, and the hyperparameters—network learning rate, batch size, and optimizer—are optimized. The search space consists of two parts: 1. the EfficientNetV2 network structure, and 2. training hyperparameters. The EfficientNetV2 structure consists of multiple stages (layers), each stage containing several layers. Optimizable structural parameters include the number of repeated layers in each stage, the number of output channels in each stage, and the kernel size. Training hyperparameters include the learning rate and batch size. Therefore, the entire search space has a dimension D of 6. With N=3, the entire search region is divided into 3 sub-regions, each a subset of the entire search space, used to search for different structures and parameters within the search range. Particles refer to surrogates used by the plant assimilation algorithm to search for optimal parameters in space. They can be understood as arrays with the following structure: [number of repetition layers per stage, number of output channels per stage, kernel size, learning rate, batch size, optimizer type]. Local growth centers refer to the particles with the best fitness values ​​within a sub-region, global growth centers refer to the particles with the best fitness values ​​across the entire search region, and optimal position refers to the solution formed by the values ​​of all parameters when the model's fitness is at its best.

[0060] Implementation Method 2: This implementation method further defines the tomato disease classification method based on plant assimilate optimization described in Implementation Method 1. In step 3, the local growth center of each sub-region is the particle with the highest fitness in each sub-region, and the global growth center of the D-dimensional search space is the particle with the highest fitness among the local growth centers of each sub-region.

[0061] This implementation method evaluates and determines growth centers, and prioritizes the supply of growth centers:

[0062] 1. Evaluate fitness. Calculate the fitness of each particle in each region at the current iteration number. The fitness value below.

[0063] 2. Determine the growth center. In each region, the particle with the best fitness is the growth center of that region, also known as the local growth center. Other particles search around this growth center. The particle with the best fitness in all regions is the global growth center. Local growth centers search around the global growth center within their respective regions.

[0064] In this implementation, the particle with the best fitness in each region is designated as the local growth center, and other particles perform directional searches around this center, avoiding the waste of resources from random exploration. Furthermore, the establishment of a global growth center ensures that all local centers evolve collaboratively towards the global optimum, guaranteeing that the growth centers in each sub-region are fully developed, forming a multi-level guidance system, maintaining diversity between regions, and avoiding local optima. In the subsequent development and exploration phases, a decision mechanism based on the Ac / Cc concentration ratio ensures that particles adaptively switch between growth center guidance and autonomous exploration. During the development phase, the local growth center becomes a strong attractor, and particles perform fine-grained searches around it using Gaussian perturbations with adaptive step sizes. The development intensity is dynamically adjusted as the search progresses, effectively improving convergence accuracy and concentrating computational resources on the most promising search regions. When particles are far from the growth center, they automatically switch to exploration mode to prevent premature convergence. Each sub-region's growth center corresponds to different candidate solutions on the Pareto front. Partial exploration comprehensively covers the target space. Later, cluster analysis and enhanced search are performed on the candidate solution set of the growth centers to better approximate the Pareto front with complex shapes.

[0065] Implementation Method 3: This implementation method further defines the tomato disease classification method based on plant assimilate optimization described in Implementation Method 1. In step 4, the optimal location of each sub-region is searched using particles, which includes an exploration phase and a development phase.

[0066] When R < 1, particle search is in the exploration stage; when R ≥ 1, particle search is in the development stage.

[0067] R = Ac / Cc;

[0068] Where R is the concentration ratio, Ac is the auxin concentration, simulating the decay characteristics with distance, and Cc is the cytokinin concentration, simulating the random pulse characteristics.

[0069] The exploration phase involves randomly searching for particles within each sub-region in that region, while the development phase involves searching for particles within each sub-region around the local growth center of the corresponding sub-region.

[0070] when ≥ At that time, the particle search is completed, among which, This represents the current iteration number. This represents the maximum number of iterations.

[0071] In this embodiment, particle decision-making behavior is designed by simulating exploration and development based on auxin / cytokinin concentrations, and virtual auxin concentrations (Ac) and cytokinin concentrations (Cc) are introduced to control particle behavior. For each particle i, its behavior is determined by the concentration ratio R = Ac / Cc.

[0072] The Ac concentration simulates the decay characteristic with distance, defined as Ac = exp(-d / σ), where d is the Euclidean distance from the particle to the local growth center of its region, and σ is the decay coefficient. The Cc concentration simulates its random impulse characteristics, defined as random sampling from a Poisson or gamma distribution.

[0073] When R < 1 (i.e., the Ac concentration is relatively low), the particles enter the exploration phase. During the exploration phase, the particle search location is:

[0074] ;

[0075] in, The step size coefficient is a random number or constant between (0,1) that controls the range of exploration. This represents the (t+1)th generation position of the i-th particle. Let i be the position of the i-th particle in the t-th generation. A random number in the range of 0 to 1. The boundary span of the current sub-region (such as the length of the region in each dimension) ensures that the exploration is confined to this region.

[0076] For a D-dimensional optimization problem, where the lower bound of the search space is LowerBound and the upper bound is UpperBound, it can be represented as:

[0077] ;

[0078] ;

[0079] in, The minimum value for the D-th dimension. The maximum value in the D-th dimension is the total span of the entire search space. The span of each sub-region in the D-th dimension is fixed:

[0080] ;

[0081] in, Let be the local span of the subregion in the D-th dimension. Let be the global span of the entire search space in the D-th dimension.

[0082] ;

[0083] If new location If a position is already marked as existing in the taboo list, discard that position and regenerate:

[0084] ;

[0085] in, A function to generate random points in a specified area. A defined region for generating random points. This is a list of contraindications.

[0086] When the ratio of auxin concentration to cytokinin concentration (Ac / Cc) within a region is greater than 1, the particles enter the development phase. During the development phase, the local growth center of the current sub-region becomes a strong attractor, and the particles perform a fine-tuning search around it to find a better solution. The particle search location is:

[0087] ;

[0088] in, A random number in the range of 0 to 1. It is the current iteration number. It represents the maximum number of iterations, and tanh() is the arctangent function. A random number in the range of 0 to 1. This represents the local growth center within the current subregion.

[0089] This implementation boasts powerful global exploration capabilities. Through partitioned search and branching strategies in the exploration phase, the algorithm can broadly cover the search space and effectively discover potential global optima. Furthermore, by employing Gaussian perturbations guided by growth centers during the development phase, this implementation can perform fine-grained searches near the optimal solution, improving convergence accuracy. Simultaneously, the Ac / Cc-based dynamic decision-making mechanism can automatically adjust the strategy according to the search progress, achieving an adaptive balance between the two.

[0090] This implementation performs information exchange and population updates, and establishes a global tabu list to record all the positions searched by particles. If the position searched by a particle exists in the tabu list, a position that has not been searched is randomly generated in that sub-region.

[0091] Each search particle within a region is only responsible for searching and developing that region. In each iteration, particles with lower fitness are annihilated. Based on the prior knowledge of the region's particles and the current optimal solution within that region, the same number of particles are randomly generated. The search is performed using these randomly generated particles, and the optimal position for each sub-region is determined based on the fitness values ​​of the positions searched by the particles.

[0092] Implementation Method Four: This implementation method further defines the tomato disease classification method based on plant assimilate optimization described in Implementation Method One. In step 5, cluster analysis is performed on the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, the Pareto optimal solution is obtained, specifically as follows:

[0093] After performing cluster analysis on the candidate solution set, the candidate solution set is divided into multiple clusters. A high-quality solution is selected from each cluster, and a local particle search is performed based on the high-quality solution to obtain the Pareto optimal solution.

[0094] This implementation integrates and outputs global information. During the algorithm's operation, it continuously records the current local growth center found in each sub-region, forming a candidate solution set A.

[0095] Before the algorithm terminates, cluster analysis is performed on the candidate solution set A. A high-quality solution is selected from each cluster, and a further enhanced local search is conducted in the vicinity of these high-quality solutions to improve the quality of the solutions. The final non-dominated solution set is either taken as the Pareto optimal solution or the global optimal solution is directly output.

[0096] This implementation focuses on finding high-quality solutions, using a smaller step size and a more refined search method than usual. It employs a Gaussian small-step perturbation, as shown in the formula:

[0097] X new =x current +N(0,δ);

[0098] Among them, X new Let x be the value of the optimized variable after perturbation. current The value of the current optimization variable is N(0,δ), which is a Gaussian distributed random perturbation term, and δ is a perturbation strength parameter, which is a random number between 0.1 and 0.3.

[0099] This implementation improves search efficiency by performing cluster analysis on the candidate solution set, concentrating computational resources on targeted local searches near representative solutions in each cluster. It groups similar solutions into the same cluster, effectively avoiding excessive clustering of solutions on the Pareto front. High-quality solutions are selected from each cluster for local searches. When the true Pareto front has complex geometric characteristics such as non-convexity, discontinuity, and extreme non-uniformity, cluster analysis can identify different characteristic regions of the front. Selecting high-quality solutions within each characteristic region for enhanced local searches better approximates complex-shaped Pareto fronts. Furthermore, the clustering process naturally divides candidate solutions into multiple representative categories, avoiding repeated ineffective searches in similar regions. Selecting high-quality solutions from each cluster for local searches ensures both the quality and diversity of solutions.

[0100] This implementation first broadly covers the search space through partitioned search, then systematically organizes the obtained candidate solutions through cluster analysis, and finally further improves the quality of the solutions through local reinforcement search. The interaction between this and the partitioned search mechanism achieves an organic unity of global exploration and local development.

[0101] Implementation Method 5: This implementation method is a specific embodiment of the tomato disease classification method based on plant assimilate optimization described in Implementation Methods 1 to 4. In the classic 30-dimensional single-objective function and bi-objective ZDT function, the parameters are set as follows: population size M = 100, number of sub-regions N = 5, decay coefficient σ = 0.1 (to control the decay rate of Ac), and annihilation period T = 20.

[0102] Optimal Range and Performance: Number of Subregions N: When N is in the range of 4 to 8, the algorithm achieves the best balance between maintaining diversity and reducing computational overhead. If N is too small (e.g., N=2), the partitioning effect is not obvious; if N is too large (e.g., N=15), resources for a single region are insufficient, and search efficiency decreases.

[0103] The attenuation coefficient σ works best in the range of [0.05, 0.2]. When σ=0.1, particles can smoothly transition from exploration to development. If σ is too small, particles will enter development too early and get trapped in local optima; if σ is too large, the exploration will be too strong and the convergence speed will be slow.

[0104] Annihilation rate: The effect is stable within the range of 5% to 15%. An annihilation rate of 10% can effectively eliminate inferior particles without causing too much impact on the stability of the population.

[0105] In actual tests, the plant assimilate optimization algorithm with the above-mentioned optimal parameter range outperformed PSO (Particle Swarm Optimization), WOA (Whale Optimization), GWO (Grey Wolf Optimization), and RIME (Rime Optimization) in convergence accuracy on 30-dimensional single-objective functions. The Pareto front distribution uniformity obtained on test functions such as ZDT1 and ZDT2 was superior to the MOEA / D algorithm.

[0106] In summary, the key points of the tomato disease classification method based on plant assimilate optimization described in this embodiment are as follows:

[0107] 1. Dynamic partitioning management of the search space;

[0108] 2. Coordinated guidance of global and local growth centers;

[0109] 3. Parameter settings and behavioral decisions for the Ac / Cc model;

[0110] 4. Combining taboo lists and random regeneration to prevent duplicate searches.

[0111] Implementation Method Six: This implementation method is a specific embodiment based on the tomato disease classification method based on plant assimilate optimization described in Implementation Methods One to Four.

[0112] In complex field environments, varying light conditions and differences in insect morphology require models to possess dynamic adaptability, while fixed hyperparameters are insufficient to meet the diverse needs of practical applications. Secondly, the structural design of convolutional neural networks largely relies on expert experience and repeated experiments, lacking theoretical guidance. Manually designed network architectures often suffer from redundant computations or insufficient feature extraction, limiting the model's deployment capability on resource-constrained devices.

[0113] This implementation method, based on a tomato disease classification method optimized by plant assimilates, selected seven common pests that severely damage to tomato crops as research subjects: whitefly, cotton bollworm, green peach aphid, beet armyworm, taro caterpillar, two-spotted spider mite, and melon fly. The original pest image dataset, containing 585 images, was obtained from publicly available online resources. To enhance the model's generalization ability, six image augmentation methods were used to augment the original images, expanding the training samples and improving the model's robustness to changes in lighting and angle. The six image augmentation methods were brightness enhancement, color enhancement, contrast enhancement, flipping, rotation, and random color adjustment. After data augmentation, a dataset containing 5153 pest images was finally constructed. All enhanced images were uniformly adjusted (cropped / scaled) to a fixed size of 224 × 224 × 3 pixels and converted to JPG format to meet the input requirements of deep convolutional neural networks.

[0114] This implementation selects the EffcientNetV2-S network for structural optimization and hyperparameter tuning. The EffcientNetV2-S structure consists of multiple stages, each containing several layers. Optimizable structural parameters include the number of repeated layers in each stage, the number of output channels in each stage, and the kernel size. Training hyperparameters include the learning rate and batch size, as detailed in Table 1.

[0115] Table 1

[0116]

[0117] This implementation concatenates the aforementioned search dimensions into a real-number vector or an integer vector, which serves as a particle in the plant assimilation algorithm. The particle encoding can be represented as:

[0118] Individual = [Stage_l, Stage_ch, Kernel_Size, Lr, Batch_S, Optim_S];

[0119] Where Stage_l is the number of repetition layers for each Stage, data type int, range [1, 5], Stage_ch is the number of output channels for each Stage, data type int, range [32, 48, 64, 80, 96, 112, 128, 160, 192, 224, 256, 320, 384, 512], Kernel_Size is the kernel size, data type int, range {3, 5}, Lr is the learning rate, data type float, range [1e-4, 1e-2], and Batch_S is the batch size, data type int, range {4, 8, 16, 32, 64, 128, 256, 512}. 512}, Optim_S is the optimizer type, data type int, range {1, 2, 3, 4}, where 1 represents Adam (adaptive moment estimator optimizer), 2 represents AdamW (a variant of Adam optimizer), 3 represents SGD (stochastic gradient descent optimizer), and 4 represents AdaGrad (adaptive gradient optimizer).

[0120] The algorithm parameters are set as follows: population size 30, dimension 6, number of iterations 100. Model training times: 200.

[0121] Fitness function design:

[0122] Fitness=α·Accuracy-β·(NumParams);

[0123] Where α=0.8, β=0.2, Accuracy is the model's accuracy, NumParams are the model parameters, and a larger Fitness value indicates better performance.

[0124] First, the algorithm initializes the model structure and hyperparameters within the search range. This involves decoding the particles found by the algorithm and converting the decoded vectors into an EfficientNetV2 configuration dictionary. Then, based on the dictionary and configuration, a PyTorch (an open-source deep learning framework for machine learning and deep learning) model is dynamically generated. At this point, a network model and hyperparameter configuration automatically found by the algorithm are obtained. Subsequently, a tomato pest and disease dataset is used, with the training, validation, and test sets divided in a 7:2:1 ratio. Images are input into the model for training, and an evaluation score is obtained using a fitness function. After optimization by the algorithm, the optimal model structure and hyperparameter values ​​are output.

[0125] The final expected results are as follows:

[0126] Table 2

[0127]

[0128] Compared with other models:

[0129] Table 3

[0130]

[0131] Optimal individual = [ [2, 3, 4, 3, 3, 2], # Stage_l: Number of repetition layers in each Stage

[0132] [48, 80, 128, 192, 256, 320], # Stage_ch: Output channels for each Stage

[0133] 3, # Kernel_Size

[0134] 0.0023, # Lr

[0135] 64, # Batch_S

[0136] 2 # Optim_S → AdamW ]

[0137] Implementation Method Seven: This implementation method describes a tomato disease classification system based on plant assimilate optimization. The system is constructed based on the tomato disease classification method described in Implementation Method One and includes the following modules:

[0138] The acquisition module acquires tomato disease image data as sample input;

[0139] The partitioning module divides the D-dimensional search space of the classification model into N equal sub-regions;

[0140] The evaluation module evaluates the different features within each sub-region. The fitness of the region is used to determine the local growth center of each sub-region and the global growth center of the D-dimensional search space;

[0141] The search module, based on the local growth center of each sub-region and the global growth center of the D-dimensional search space, uses particles to search for the optimal position of each sub-region, forming a candidate solution set;

[0142] The module constructs a cluster analysis of the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, the Pareto optimal solution is obtained. Based on the Pareto optimal solution, a tomato disease classification model is constructed. Input samples are then used to classify tomato diseases.

[0143] The above provides a detailed description of the tomato disease classification method and system based on plant assimilate optimization proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A tomato disease classification method based on plant assimilate optimization, characterized in that, Includes the following steps: Step 1: Obtain tomato disease image data as sample input; Step 2: Divide the D-dimensional search space of the classification model into N equal sub-regions; Step 3: Evaluate the fitness of different particles in each sub-region to determine the local growth center and the global growth center of the D-dimensional search space for each sub-region; In step 3, the local growth center of each sub-region is the particle with the highest fitness in each sub-region, and the global growth center of the D-dimensional search space is the particle with the highest fitness among the local growth centers of each sub-region. Step 4: Based on the local growth center of each sub-region and the global growth center of the D-dimensional search space, use particle search to find the optimal position of each sub-region and form a candidate solution set; Step 4, which uses particle search to find the optimal location for each sub-region, includes an exploration phase and a development phase. When R < 1, particle search is in the exploration stage; when R ≥ 1, particle search is in the development stage. R = Ac / Cc; Where R is the concentration ratio, Ac is the auxin concentration, simulating the decay characteristics with distance, and Cc is the cytokinin concentration, simulating the random pulse characteristics. The exploration phase involves randomly searching for particles within each sub-region in that region, while the development phase involves searching for particles within each sub-region around the local growth center of the corresponding sub-region. when ≥ At that time, the particle search is completed, among which, This represents the current iteration number. This represents the maximum number of iterations. In each iteration, poor particles are annihilated and the same number of particles are randomly generated. The randomly generated particles are used for searching, and the optimal position of each sub-region is determined based on the fitness value of the positions searched by the particles. During the exploration phase, when searching for particles, the particle search location is: ； in, To explore the step size coefficient, This represents the (t+1)th generation position of the i-th particle. Let i be the position of the i-th particle in the t-th generation. A random number in the range of 0 to 1. This represents the boundary span of the current sub-region; During the development phase, when performing particle search, the particle search location is: ； in, A random number in the range of 0 to 1. It is the arctangent function. A random number in the range of 0 to 1. This serves as the local growth center within the current sub-region; Step 5: Perform cluster analysis on the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, obtain the Pareto optimal solution. Based on the Pareto optimal solution, construct a tomato disease classification model, input samples, and realize tomato disease classification.

2. The tomato disease classification method based on plant assimilate optimization according to claim 1, characterized in that, A global tabu list is created to record the locations searched by the particle. If the location searched by the particle is already recorded in the tabu list, a location that has not been searched is randomly generated in the corresponding sub-region.

3. The tomato disease classification method based on plant assimilate optimization according to claim 1, characterized in that, In step 5, cluster analysis is performed on the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, the Pareto optimal solution is obtained, specifically as follows: After performing cluster analysis on the candidate solution set, the candidate solution set is divided into multiple clusters. A high-quality solution is selected from each cluster, and a local particle search is performed based on the high-quality solution to obtain the Pareto optimal solution.

4. A tomato disease classification system based on plant assimilate optimization, said system being constructed based on the tomato disease classification method of claim 1, characterized in that, Includes the following modules: The acquisition module acquires tomato disease image data as sample input; The partitioning module divides the D-dimensional search space of the classification model into N equal sub-regions; The evaluation module assesses the fitness of different particles within each sub-region, and determines the local growth center and the global growth center of the D-dimensional search space for each sub-region. The search module, based on the local growth center of each sub-region and the global growth center of the D-dimensional search space, uses particles to search for the optimal position of each sub-region, forming a candidate solution set; The module constructs a cluster analysis of the candidate solution set to generate high-quality solutions. After processing the high-quality solutions, the Pareto optimal solution is obtained. Based on the Pareto optimal solution, a tomato disease classification model is constructed. Input samples are then used to classify tomato diseases.

Images