Image recognition-based leaf nutrition state diagnosis and chemical fertilizer regulation method

Abstract

The invention discloses a leaf nutrition state diagnosis and chemical fertilizer adjustment method based on image recognition. The method comprises the following steps: S1, collecting a leaf image; s2, performing denoising, enhancement and background removal on the image, extracting color, shape and texture features, and generating a leaf feature vector; s3, modeling by using a nonlinear dynamic chaotic system, detecting abnormal changes, and performing multi-scale chaotic analysis; s4, constructing a topological network for high-dimensional topological data analysis, and performing multi-modal data fusion in combination with a Bayesian network; s5, dynamically adjusting algorithm parameters according to a fusion result; s6, generating a diagnosis result through a generative adversarial network, image fusion and a deep artifact removal algorithm; and S7, constructing a cognitive intelligent agent and group intelligence optimization algorithm, and carrying out dynamic environment adaptation and closed-loop feedback regulation and control. According to the invention, the efficiency and accuracy of nutrition state diagnosis are improved, the fertilizer application is optimized, and the environmental pollution is reduced.

Description

Technical Field

[0001] The invention relates to the technical field of agriculture, and in particular to a leaf nutritional status diagnosis and fertilizer adjustment method based on image recognition. Background Art

[0002] With the advancement of science and technology and the intelligent development of agricultural production, leaf nutrition status diagnosis and fertilizer adjustment methods based on image recognition have gradually become a hot topic in agricultural research. Traditional leaf nutrition status diagnosis methods mainly rely on manual observation and laboratory analysis. These methods have problems such as low efficiency, insufficient accuracy and poor real-time performance, and are difficult to meet the needs of modern agricultural production for precise management. Specifically, the existing technology has defects in the following aspects:

[0003] First, the limitations of traditional methods: Traditional leaf nutritional status diagnosis usually relies on the experience of agronomic experts, and the nutritional status of leaves is judged by observing the color, shape and other characteristics of leaves with the naked eye. This method has the disadvantages of strong subjectivity, large errors, and low efficiency. The judgment results between different observers may vary greatly, affecting the accuracy of diagnosis. Another common diagnostic method is to analyze the nutritional components of leaves in the laboratory. Although this method is more accurate, it requires multiple steps such as sampling, transportation, and testing. The process is cumbersome and time-consuming, and it is difficult to achieve real-time monitoring and rapid response. In addition, the cost of laboratory analysis is high and it is not suitable for large-scale crop management. Lack of real-time: Whether it is manual observation or laboratory analysis, it is difficult to reflect the dynamic changes of leaf nutritional status in a timely manner, and it cannot meet the real-time monitoring needs of nutritional status during crop growth. This lack of real-time diagnostic method leads to a lag in crop nutrition management, affecting the healthy growth and yield of crops.

[0004] Secondly, the application of fertilizers is not precise: Traditional fertilizer application methods are usually based on a unified fertilization plan and lack accurate data support for specific crop nutritional needs. This can easily lead to excessive or insufficient use of fertilizers, which not only wastes resources but also may pollute the environment. In addition, due to the lack of accurate nutritional status data support, traditional fertilizer application methods often lead to excessive or insufficient use of fertilizers, resulting in a waste of resources. For example, when fertilizers are used excessively, it not only increases agricultural production costs, but may also lead to crop nutritional excess and affect crop quality. Moreover, excessive fertilizer application not only wastes resources, but also causes soil and water pollution and damages the ecological environment. For example, excessive nitrogen fertilizers may enter water bodies through rainwater runoff, leading to eutrophication of water bodies and destroying aquatic ecosystems.

[0005] Third: Lack of intelligent control mechanism: Traditional nutrition status monitoring and fertilization management methods lack intelligent and dynamic control mechanisms, and cannot adjust fertilization strategies in time according to the actual situation of crop growth and environmental changes, affecting the optimal growth state and final yield of crops. Traditional methods lack intelligent control mechanisms, and it is difficult to make dynamic adjustments according to real-time changes in crop nutrition status, which affects the healthy growth of crops. For example, when the nutritional status of crops changes, the fertilization plan cannot be adjusted in time, which may lead to insufficient or excessive nutrition. The existing methods have a low degree of intelligence and cannot make full use of modern scientific and technological means for precision agricultural management. For example, the lack of intelligent diagnosis and control systems based on advanced technologies such as big data analysis and machine learning affects the scientificity and efficiency of crop management.

[0006] In summary, the traditional leaf nutrition status diagnosis and fertilizer adjustment methods have many shortcomings and cannot meet the needs of modern agriculture for efficient and precise management. Therefore, there is an urgent need for a leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition and intelligent algorithms. By introducing advanced technical means, the efficiency and accuracy of nutrition status diagnosis can be improved, the precise application of fertilizers can be achieved, resource waste and environmental pollution can be reduced, and the sustainable development of modern agriculture can be promoted. Summary of the invention

[0007] The present invention realizes efficient, accurate and intelligent management of leaf nutritional status diagnosis and fertilizer adjustment by innovatively combining multidisciplinary technologies, significantly improves agricultural production efficiency and resource utilization efficiency, reduces environmental pollution, and has broad application prospects and important practical significance.

[0008] The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition according to an embodiment of the present invention comprises the following steps:

[0009] The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition is characterized by comprising the following steps:

[0010] S1. Collect images of crop leaves using a camera or drone;

[0011] S2, performing denoising, image enhancement and background removal processing on the collected leaf image, extracting the color, shape and texture features of the leaf image, and generating a leaf feature vector;

[0012] S3, using nonlinear dynamic chaotic system to model blade eigenvectors, calculating Lyapunov exponents to detect abnormal changes, and performing multi-scale chaos analysis;

[0013] S4. Use high-dimensional topological data analysis technology to construct a topological network of leaf nutritional status data, combine it with Bayesian network to capture the causal relationship and probability dependence between variables, and perform multimodal data fusion;

[0014] S5. Dynamically adjust algorithm parameters according to the multimodal data fusion results;

[0015] S6, generating images through multi-view image fusion and deep artifact removal algorithm of generative adversarial network to generate leaf nutrition status diagnosis results;

[0016] S7. Build intelligent agents with cognitive capabilities and swarm intelligence optimization algorithms to work together to perform dynamic environmental adaptation and closed-loop feedback control.

[0017] Optionally, the S1 includes the following steps:

[0018] S11, selecting a camera or a drone as an image acquisition device to obtain a leaf image;

[0019] S12, adjusting the shooting parameters of the camera or the drone so that leaf images can be captured under different lighting conditions;

[0020] S13, setting the shooting frequency and time interval of the camera or drone, and regularly collecting leaf images according to the crop growth cycle and monitoring requirements;

[0021] S14. During the shooting process, the location information of the camera or drone is recorded, and the latitude, longitude and altitude data of the shooting point are obtained through GPS or other positioning systems;

[0022] S15, performing a preliminary inspection on the collected leaf images;

[0023] S16. Store the preliminarily inspected blade image and its position information in a data storage device.

[0024] Optionally, S2 includes the following steps:

[0025] S21, performing denoising on the collected leaf image, using Gaussian filtering or median filtering algorithm to remove noise in the image, retaining the main feature information of the leaf, and obtaining a denoised image;

[0026] S22, performing image processing on the denoised leaf image by using a histogram equalization or a contrast-limited adaptive histogram equalization algorithm;

[0027] S23, performing background removal processing, using threshold segmentation or edge detection algorithm to perform background removal processing;

[0028] S24, extracting color features from the leaf image after background removal, calculating the average color value, color histogram and color distribution of the leaf, and generating a color feature vector;

[0029] S25, extracting shape features from the leaf image after background removal, calculating feature parameters such as the leaf contour, area, perimeter and shape moment, and generating a shape feature vector;

[0030] S26, extracting texture features from the leaf image after background removal, calculating the texture features of the leaf using a gray-level co-occurrence matrix or a local binary pattern algorithm, and generating a texture feature vector;

[0031] S27, merging the color feature vector, the shape feature vector and the texture feature vector to generate a comprehensive leaf feature vector.

[0032] Optionally, S3 includes the following steps:

[0033] S31. Construct a nonlinear dynamic chaotic system model integrating multi-source data assimilation and define the state equation:

[0034] x(t+1)=F(x(t),p)+N(t);

[0035] Among them, x(t) is the blade characteristic vector at the current moment, F is the nonlinear mapping function, p is the system parameter vector, and N(t) is the random noise vector;

[0036] S32. Use multi-modal collaborative chaos optimization algorithm to optimize the chaotic system parameter p:

[0037] S321. Generate an initial population P(0) in the parameter space, where each individual represents a possible parameter vector:

[0038] P(0)={p1(0),p2(0),…,p N (0)};

[0039] S322, different subpopulations P i (t) co-evolution, using mutation and crossover operations to simulate the natural selection process:

[0040] p i (t+1)=p i (t)+α·(p j (t)-p k (t))+β·G i (t);

[0041] Among them, p i (t), p j (t) and p k (t) is the parameter vector of the i-th, j-th and k-th individuals, α is the variation factor, β is the crossover factor, G i (t) is the guide vector, defined as:

[0042]

[0043] Among them, γ is the learning rate, w m is the weight vector, x m (t) is the mth modal eigenvector, M is the number of modes;

[0044] S323, introduce improved Logistic mapping to optimize the population, increase the diversity of the search space and the global exploration ability, and avoid local optimal solutions:

[0045] p i (t+1)=μ·p i (t)·(1-p i (t));

[0046]

[0047] Among them, μ is the chaos parameter, f(p i (t)) is the fitness function, f max is the maximum fitness;

[0048] S324, through the online learning mechanism, dynamically adjust the evolution strategy and chaos mapping parameters according to environmental changes:

[0049] p i (t+1)=p i (t)+η·(p best (t)-p i (t));

[0050] Among them, η is the adaptive learning rate, p best (t) is the current optimal parameter vector;

[0051] S325, judging whether the optimization process is finished according to the convergence of the objective function, and outputting the optimized system parameter vector p MMCCOA ;

[0052] S33, the optimized system parameter p MMCCOA Input into the chaotic system model, calculate the Lyapunov exponent λ, and detect the chaotic characteristics and abnormal changes of the system:

[0053]

[0054] δx(i)=x(i)-x′(i);

[0055] Among them, δx(i) is the disturbance vector at the i-th moment, x'(i) is the blade characteristic vector after introducing a small disturbance,

[0056] S34. By calculating the growth rate of the disturbance vector at each moment, the chaotic characteristics and abnormal changes of the system are judged. When λ>0, the system is in a chaotic state and there are abnormal changes.

[0057] S35. Perform multi-scale chaos analysis, use adaptive multi-scale transformation to decompose the multi-modal feature vector, and extract features at different time scales:

[0058]

[0059] Among them, ψ j,k (t) is the adaptive multi-scale transformation basis function, c j (k) is the coefficient of the jth layer;

[0060] S36, extract the characteristic vector x of each time scale through adaptive multi-scale analysis method j (t);

[0061] S37. Use the Bayesian dynamic fusion-time series prediction algorithm to combine the multi-scale analysis results with the optimized chaotic system model:

[0062]

[0063] in, Optimize the LSTM network weights for the optimized Bayesian inference. P(D|w) is the likelihood function under Bayesian inference. is the predicted leaf feature vector, F LSTM Optimizing the prediction function of long short-term memory networks for Bayesian reasoning that integrates multi-scale features and chaotic system models;

[0064] S38. Comprehensively evaluate abnormal changes in leaf nutritional status.

[0065] Optionally, S4 includes the following steps:

[0066] S41, construct a multimodal data matrix X, integrating multimodal data such as leaf images, soil composition, and climate conditions, where X = [x1, x2, …, x N ], each x i is a data vector containing different modal features;

[0067] S42. Calculate the similarity matrix S between data samples using the improved Gaussian kernel:

[0068]

[0069] Among them, S ij is the similarity between the i-th and j-th samples, σ is the bandwidth parameter of the kernel function;

[0070] S43. Use high-dimensional topological data analysis technology to analyze the similarity matrix S, construct a topological network K, and obtain topological features by calculating persistent homology:

[0071]

[0072] Among them, β k is the kth order Betti number, is a k-order boundary operator;

[0073] S44. Construct a spatiotemporal dynamic Bayesian network to capture the complex spatiotemporal dependencies between variables, and use graph convolutional networks and temporal recurrent neural networks to extract spatiotemporal features:

[0074] H (l+1) =σ(AH (l) W (l) );

[0075] h t =σ(W h x t +U h h t-1 +b h )

[0076] Among them, H (l) is the node feature matrix of the lth layer, A is the adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function, h t is the hidden state at time t, W h , U h and b h are the weight matrix and bias vector respectively;

[0077] S45, perform multimodal data fusion and use adaptive variational autoencoder to generate high-dimensional fusion features;

[0078]

[0079] where z is the latent variable, μ(x) and σ(x) are the mean and standard deviation of the encoder parameterization, and μ θ (z) and is the mean and standard deviation of the decoder parameterization;

[0080] S46. Design an adaptive quantum immune memory module, combining quantum computing and immune algorithms to record and update historical data:

[0081]

[0082] Among them, α i is the complex amplitude, |i> is the ground state;

[0083] S47. Use the quantum clone selection algorithm to update the memory bank, select the quantum state with the highest fitness for cloning, perform mutation operations on the quantum state, measure the quantum state, and obtain classical memory cells:

[0084]

[0085] M(|ψ>)={|ψ1>,|ψ2>,…,|ψ m >};

[0086] Among them, U mutation |ψ> is the quantum mutation operator, (|ψ i > is the classical state after measurement;

[0087] S48, the newly collected data sample x new Match with quantum immune memory bank M, generate new diagnostic results and update the memory bank to form a comprehensive analysis report: x new →Match(M)→Update(M).

[0088] Optionally, S5 includes the following steps:

[0089] S51, the multimodal data fusion result Z is input into the adaptive multimodal hybrid generative adversarial network, and the adaptive multimodal hybrid generative adversarial network combines multimodal data, generative adversarial network and variational autoencoder. The objective function is:

[0090] Generator objective function:

[0091]

[0092] Discriminator objective function:

[0093]

[0094] Among them, p is the parameter vector to be optimized for the generator, q is the parameter vector to be optimized for the discriminator, G is the generator function, D is the discriminator function, E is the encoder function, λ is the regularization parameter, and z is the latent variable;

[0095] S52. Construct a proxy model based on a multimodal graph neural network and a variational generative network. The multimodal graph neural network is used to extract graph structure and multimodal features, and the variational generative network is used to generate high-dimensional latent features:

[0096] H (l+1) =σ(AH (l) W (l) )+βT (l) ;

[0097] q φ (z|x)=N(z;μ(x),σ2 (x));

[0098]

[0099] Among them, H (l) is the node feature matrix of the lth layer, A is the adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function, T (l) is the multimodal feature matrix, β is the fusion weight; z is the latent variable, μ(x) and σ(x) are the mean and standard deviation of the encoder, μ θ (z) and is the mean and standard deviation of the decoder;

[0100] S53, using the expected improvement strategy of adaptive multi-modal generation adversarial optimization to perform parameter optimization and calculate the expected improvement value:

[0101]

[0102] Among them, p + is the current optimal parameter, G(z|p) is the generator function value, and D is the discriminator function value;

[0103] S54, the optimized parameter p * Input into the leaf nutrition status assessment model, build an adaptive multimodal generative adversarial network, and combine the leaf nutrition status assessment with graph neural network and generative adversarial network:

[0104] G(z|p)=Decoder(z;p);

[0105] D(x|p)=σ(W d xb d );

[0106] Among them, z is the latent variable, p is the optimization parameter, and W d and b d are the weight and bias of the discriminator, σ is the activation function;

[0107] S55. Optimizing the Adaptive Multimodal Generative Adversarial Network Model Using the Multimodal Graph Neural Network Attention Mechanism:

[0108] a t =softmax(W a [h t ;c t ])+ηGNN(H t ,A);

[0109] Among them, a t is the attention weight at the tth moment, W a is the attention weight matrix, ct is the context vector, η is the graph neural network weight, GNN is the graph neural network function, H t is the node feature matrix at the tth moment, and A is the adjacency matrix;

[0110] S56. Use the output of the multimodal graph neural network model to perform real-time evaluation and generate an evaluation report R, including leaf nutrition indicators:

[0111] R={N(h T ),P(h T ),K(h T )};

[0112] Among them, N(h T )、P(h T ) and K(h T ) are the estimated values ​​of nitrogen, phosphorus and potassium content respectively;

[0113] S57. According to the evaluation result R, dynamically adjust the fertilizer application plan to generate a fertilization plan F:

[0114] F={f1,f2,…,f n};

[0115] Among them, f i is the application amount of the i-th fertilizer.

[0116] Optionally, the S6 comprises the following steps:

[0117] S61. Input the multimodal data into the adaptive recursive hash network. The adaptive recursive hash network extracts the high-dimensional features of the leaf image through a multi-layer recursive hash function and maps it to a low-dimensional hash space:

[0118]

[0119] Among them, p is the parameter to be optimized of the recursive hash network, x i and x i ' are the input image and its enhanced image, H is the recursive hash function, z j is a latent variable, α is a regularization parameter;

[0120] S62. Design a deep artifact removal algorithm based on a sparse encoder-decoder network. The sparse encoder extracts sparse features of the leaf image, and the sparse decoder reconstructs the image:

[0121]

[0122] Among them, x i is the i-th input image, E is the sparse encoder function, D is the sparse decoder function, h jis the jth hidden layer feature, D SEDN is the artifact removal decoder function, β is the regularization parameter;

[0123] S63, input the processed leaf image into a multi-scale wavelet transform module, and use wavelet transform to decompose the image into different scales:

[0124]

[0125] Among them, W m,n is the wavelet coefficient of the mth scale and the nth time, x(t) is the input image signal, ψ is the mother wavelet function, δ is the time interval, s m is the scale parameter;

[0126] S64, reconstructing the image using multi-scale wavelet coefficients and inputting it into a multimodal spatiotemporal convolutional network. The multimodal spatiotemporal convolutional network detects and analyzes the nutritional status of leaves in combination with spatiotemporal features. The optimization objective function is:

[0127]

[0128] Among them, W k is the multi-scale wavelet coefficient, F MSTCN is the spatiotemporal convolution function, y k is the leaf nutritional status label, H t is the spatiotemporal feature, G MSTCN is the spatiotemporal convolutional network function, γ is the regularization coefficient;

[0129] S65. Generate leaf nutrition status diagnosis results, use the spatiotemporal features output by the multimodal spatiotemporal convolutional network, and perform classification through the fully connected layer. The classification results include the nitrogen, phosphorus, and potassium contents of the leaves:

[0130] y=softmax(W fc h+b fc );

[0131] Among them, softmax is the activation function, W fc and b fc are the weight and bias of the fully connected layer respectively, and h is the spatiotemporal feature of the multimodal spatiotemporal convolutional network model.

[0132] Optionally, the S7 includes the following steps:

[0133] S71. Design a cognitive intelligent agent model. The cognitive intelligent agent model simulates the knowledge and experience of agricultural experts through a Bayesian reasoning system to build a cognitive model:

[0134]

[0135] Where E is the set of leaf nutritional state variables, H is the set of environmental data variables, P(E|H) is the posterior probability of leaf nutritional state under environmental data, P(H|E) is the likelihood of leaf nutritional state for environmental data, P(E) is the prior probability of leaf nutritional state, and P(H) is the marginal probability of environmental data;

[0136] S72. Build a multi-agent deep reinforcement learning system, where each agent represents a crop area, and use the deep deterministic policy gradient algorithm for collaborative optimization, policy update and Q value update:

[0137]

[0138] Among them, θ π is the parameter of the policy network, α is the learning rate, J(θ π ) is the policy objective function, θ Q are the parameters of the Q-value network, β is the learning rate, r is the reward value, γ is the discount factor, Q' and π' are the target networks, s and a are the state and action respectively;

[0139] S73. Using a multi-objective differential evolution algorithm, combined with leaf nutrition status diagnosis results and environmental data, an optimal fertilization strategy is formed, and the optimization objective function is:

[0140]

[0141] Among them, F is the fertilization strategy, N j , P j and K j are the nitrogen, phosphorus and potassium contents in the jth crop area, and are the target nitrogen, phosphorus and potassium contents, w1, w2, w3 and w4 are weight coefficients, and C(F) is the fertilization cost;

[0142] S74, realize real-time adjustment of fertilization scheme, and optimize fertilization strategy by combining recursive neural network:

[0143] F(t+1)=F(t)+α·(F best -F(t))+β·(F rand -F(t));

[0144] Among them, F(t) is the current fertilization strategy, F best is the optimal fertilization strategy, F rand is the randomly selected fertilization strategy, α and β are the adjustment coefficients;

[0145] S75. Through the closed-loop feedback control system, the crop growth and environmental changes are continuously monitored, and the fertilization strategy is adjusted according to the real-time data:

[0146] ΔF(t)=η(E(t)-E * (t))+ζ(H(t)-H * (t));

[0147] Among them, ΔF(t) is the fertilizer adjustment amount at the tth moment, E(t) and H(t) are the current leaf nutrition status and environmental data, respectively, and E * (t) and H * (t) is the target state, η and ζ are the adjustment coefficients.

[0148] The beneficial effects of the present invention are:

[0149] (1) The present invention proposes a method of using a nonlinear dynamic chaotic system to model leaf eigenvectors, calculate the Lyapunov exponent to detect abnormal changes, and perform multi-scale chaos analysis, which effectively improves the real-time and accuracy of nutritional status diagnosis and can promptly detect abnormal conditions in crop growth.

[0150] (2) The present invention proposes a method of multimodal data fusion by using high-dimensional topological data analysis technology to construct a topological network of leaf nutritional status data, combining it with a Bayesian network to capture the causal relationship and probabilistic dependency between variables, thereby enhancing the comprehensiveness and reliability of diagnosis and being able to more accurately reflect the actual nutritional status of crops.

[0151] (3) The present invention proposes a method of generating high-quality leaf images by using a generative adversarial network (GAN) and a deep artifact removal algorithm, and combining it with multi-view image fusion technology, which significantly improves the accuracy of nutritional status diagnosis and reduces the impact of image artifacts on the diagnosis results.

[0152] (4) The present invention proposes a method for constructing an intelligent agent with cognitive capabilities and a swarm intelligence optimization algorithm to perform dynamic environmental adaptation and closed-loop feedback control, thereby achieving precise fertilization, adjusting the fertilization plan in real time according to the actual needs of crop growth, optimizing crop growth conditions, improving agricultural production efficiency, and reducing fertilizer waste and environmental pollution. BRIEF DESCRIPTION OF THE DRAWINGS

[0153] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0154] Figure 1 A flow chart of the leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition proposed by the present invention;

[0155] Figure 2 A schematic diagram of a nonlinear dynamic chaotic system for the leaf nutrition status diagnosis and fertilizer regulation method based on image recognition proposed by the present invention;

[0156] Figure 3 This is a multimodal data processing flow chart of the leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition proposed by the present invention. DETAILED DESCRIPTION

[0157] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0158] refer to Figure 1-3 , the leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition includes the following steps:

[0159] S1. Collect images of crop leaves using a camera or drone;

[0160] S2, performing denoising, image enhancement and background removal processing on the collected leaf image, extracting the color, shape and texture features of the leaf image, and generating a leaf feature vector;

[0161] S3, using nonlinear dynamic chaotic system to model blade eigenvectors, calculating Lyapunov exponents to detect abnormal changes, and performing multi-scale chaos analysis;

[0162] S4. Use high-dimensional topological data analysis technology to construct a topological network of leaf nutritional status data, combine it with Bayesian network to capture the causal relationship and probability dependence between variables, and perform multimodal data fusion;

[0163] S5. Dynamically adjust algorithm parameters according to the multimodal data fusion results;

[0164] S6, generating images through multi-view image fusion and deep artifact removal algorithm of generative adversarial network to generate leaf nutrition status diagnosis results;

[0165] S7. Build intelligent agents with cognitive capabilities and swarm intelligence optimization algorithms to work together to perform dynamic environmental adaptation and closed-loop feedback control.

[0166] In this implementation, S1 includes the following steps:

[0167] S11, selecting a camera or a drone as an image acquisition device to obtain a leaf image;

[0168] S12, adjusting the shooting parameters of the camera or the drone so that leaf images can be captured under different lighting conditions;

[0169] S13, setting the shooting frequency and time interval of the camera or drone, and regularly collecting leaf images according to the crop growth cycle and monitoring requirements;

[0170] S14. During the shooting process, the location information of the camera or drone is recorded, and the latitude, longitude and altitude data of the shooting point are obtained through GPS or other positioning systems;

[0171] S15, performing a preliminary inspection on the collected leaf images;

[0172] S16. Store the preliminarily inspected blade image and its position information in a data storage device.

[0173] In this implementation, S2 includes the following steps:

[0174] S21, performing denoising on the collected leaf image, using Gaussian filtering or median filtering algorithm to remove noise in the image, retaining the main feature information of the leaf, and obtaining a denoised image;

[0175] S22, performing image processing on the denoised leaf image by using a histogram equalization or a contrast-limited adaptive histogram equalization algorithm;

[0176] S23, performing background removal processing, using threshold segmentation or edge detection algorithm to perform background removal processing;

[0177] S24, extracting color features from the leaf image after background removal, calculating the average color value, color histogram and color distribution of the leaf, and generating a color feature vector;

[0178] S25, extracting shape features from the leaf image after background removal, calculating feature parameters such as the leaf contour, area, perimeter and shape moment, and generating a shape feature vector;

[0179] S26, extracting texture features from the leaf image after background removal, calculating the texture features of the leaf using a gray-level co-occurrence matrix or a local binary pattern algorithm, and generating a texture feature vector;

[0180] S27, merging the color feature vector, the shape feature vector and the texture feature vector to generate a comprehensive leaf feature vector.

[0181] In this implementation, S3 includes the following steps:

[0182] S31. Construct a nonlinear dynamic chaotic system model integrating multi-source data assimilation and define the state equation:

[0183] x(t+1)=F(x(t),p)+N(t);

[0184] Among them, x(t) is the blade characteristic vector at the current moment, F is the nonlinear mapping function, p is the system parameter vector, and N(t) is the random noise vector;

[0185] S32. Use multi-modal collaborative chaos optimization algorithm to optimize the chaotic system parameter p:

[0186] S321. Generate an initial population P(0) in the parameter space, where each individual represents a possible parameter vector:

[0187] P(0)={p1(0),p2(0),…,p N (0)};

[0188] S322, different subpopulations P i (t) co-evolution, using mutation and crossover operations to simulate the natural selection process:

[0189] p i (t+1)=p i (t)+α·(p j (t)-p k (t))+β·G i (t);

[0190] Among them, p i (t), p j (t) and p k (t) is the parameter vector of the i-th, j-th and k-th individuals, α is the variation factor, β is the crossover factor, G i (t) is the guide vector, defined as:

[0191]

[0192] Among them, γ is the learning rate, w m is the weight vector, x m (t) is the mth modal eigenvector, M is the number of modes;

[0193] S323, introduce improved Logistic mapping to optimize the population, increase the diversity of the search space and the global exploration ability, and avoid local optimal solutions:

[0194] p i (t+1)=μ·p i (t)·(1-p i (t));

[0195]

[0196] Among them, μ is the chaos parameter, f(p i (t)) is the fitness function, f max is the maximum fitness;

[0197] S324, through the online learning mechanism, dynamically adjust the evolution strategy and chaos mapping parameters according to environmental changes:

[0198] p i (t+1)=p i (t)+η·(p best (t)-p i (t));

[0199] Among them, η is the adaptive learning rate, p best (t) is the current optimal parameter vector;

[0200] S325, judging whether the optimization process is finished according to the convergence of the objective function, and outputting the optimized system parameter vector p MMCCOA ;

[0201] S33, the optimized system parameter p MMCCOA Input into the chaotic system model, calculate the Lyapunov exponent λ, and detect the chaotic characteristics and abnormal changes of the system:

[0202]

[0203] δx(i)=x(i)-x′(i);

[0204] Among them, δx(i) is the disturbance vector at the i-th moment, x'(i) is the blade characteristic vector after introducing a small disturbance,

[0205] S34. By calculating the growth rate of the disturbance vector at each moment, the chaotic characteristics and abnormal changes of the system are judged. When λ>0, the system is in a chaotic state and there are abnormal changes.

[0206] S35. Perform multi-scale chaos analysis, use adaptive multi-scale transformation to decompose the multi-modal feature vector, and extract features at different time scales:

[0207]

[0208] Among them, ψ j,k (t) is the adaptive multi-scale transformation basis function, c j (k) is the coefficient of the jth layer;

[0209] S36, extract the characteristic vector x of each time scale through adaptive multi-scale analysis method j (t);

[0210] S37. Use the Bayesian dynamic fusion-time series prediction algorithm to combine the multi-scale analysis results with the optimized chaotic system model:

[0211]

[0212] in, Optimize the LSTM network weights for the optimized Bayesian inference. P(D|w) is the likelihood function under Bayesian inference. is the predicted leaf feature vector, F LSTM Optimizing the prediction function of long short-term memory networks for Bayesian reasoning that integrates multi-scale features and chaotic system models;

[0213] S38. Comprehensively evaluate abnormal changes in leaf nutritional status.

[0214] In this implementation, S4 includes the following steps:

[0215] S41, construct a multimodal data matrix X, integrating multimodal data such as leaf images, soil composition, and climate conditions, where X = [x1, x2, …, x N ], each x i is a data vector containing different modal features;

[0216] S42. Calculate the similarity matrix S between data samples using the improved Gaussian kernel:

[0217]

[0218] Among them, S ij is the similarity between the i-th and j-th samples, σ is the bandwidth parameter of the kernel function;

[0219] S43. Use high-dimensional topological data analysis technology to analyze the similarity matrix S, construct a topological network K, and obtain topological features by calculating persistent homology:

[0220]

[0221] Among them, β k is the kth order Betti number, is a k-order boundary operator;

[0222] S44. Construct a spatiotemporal dynamic Bayesian network to capture the complex spatiotemporal dependencies between variables, and use graph convolutional networks and temporal recurrent neural networks to extract spatiotemporal features:

[0223] H (l+1) =σ(AH (l) W (l) );

[0224] h t =σ(W h x t +U h h t-1 +b h )

[0225] Among them, H (l)is the node feature matrix of the lth layer, A is the adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function, h t is the hidden state at time t, W h , U h and b h are the weight matrix and bias vector respectively;

[0226] S45, perform multimodal data fusion and use adaptive variational autoencoder to generate high-dimensional fusion features;

[0227]

[0228] where z is the latent variable, μ(x) and σ(x) are the mean and standard deviation of the encoder parameterization, and μ θ (z) and is the mean and standard deviation of the decoder parameterization;

[0229] S46. Design an adaptive quantum immune memory module, combining quantum computing and immune algorithms to record and update historical data:

[0230]

[0231] Among them, α i is the complex amplitude, |i> is the ground state;

[0232] S47. Use the quantum clone selection algorithm to update the memory bank, select the quantum state with the highest fitness for cloning, perform mutation operations on the quantum state, measure the quantum state, and obtain classical memory cells:

[0233]

[0234] M(|ψ>)={|ψ1>,|ψ2>,…,|ψ m >};

[0235] Among them, U mutation |ψ> is the quantum mutation operator, (|ψ i > is the classical state after measurement;

[0236] S48, the newly collected data sample x new Match with quantum immune memory bank M, generate new diagnostic results and update the memory bank to form a comprehensive analysis report: x new →Match(M)→Update(M).

[0237] In this implementation, S5 includes the following steps:

[0238] S51, the multimodal data fusion result Z is input into the adaptive multimodal hybrid generative adversarial network, and the adaptive multimodal hybrid generative adversarial network combines multimodal data, generative adversarial network and variational autoencoder to optimize the objective function:

[0239] Generator objective function:

[0240]

[0241] Discriminator objective function:

[0242]

[0243] Among them, p is the parameter vector to be optimized for the generator, q is the parameter vector to be optimized for the discriminator, G is the generator function, D is the discriminator function, E is the encoder function, λ is the regularization parameter, and z is the latent variable;

[0244] S52. Construct a proxy model based on a multimodal graph neural network and a variational generative network. The multimodal graph neural network is used to extract graph structure and multimodal features, and the variational generative network is used to generate high-dimensional latent features:

[0245] H (l+1) =σ(AH (l) W (l) )+βT (l) ;

[0246] q φ (z|x)=N(z;μ(x),σ 2 (x));

[0247]

[0248] Among them, H (l) is the node feature matrix of the lth layer, A is the adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function, T (l) is the multimodal feature matrix, β is the fusion weight; z is the latent variable, μ(x) and σ(x) are the mean and standard deviation of the encoder, μ θ (z) and is the mean and standard deviation of the decoder;

[0249] S53, using the expected improvement strategy of adaptive multi-modal generation adversarial optimization to perform parameter optimization and calculate the expected improvement value:

[0250]

[0251] Among them, p + is the current optimal parameter, G(z|p) is the generator function value, and D is the discriminator function value;

[0252] S54, the optimized parameter p * Input into the leaf nutrition status assessment model, build an adaptive multimodal generative adversarial network, and combine the leaf nutrition status assessment with graph neural network and generative adversarial network:

[0253] G(z|p)=Decoder(z;p);

[0254] D(x|p)=σ(W d xb d );

[0255] Among them, z is the latent variable, p is the optimization parameter, and W d and b d are the weight and bias of the discriminator, σ is the activation function;

[0256] S55. Optimizing the Adaptive Multimodal Generative Adversarial Network Model Using the Multimodal Graph Neural Network Attention Mechanism:

[0257] a t =softmax(W a [h t ;c t ])+ηGNN(H t ,A);

[0258] Among them, a t is the attention weight at the tth moment, W a is the attention weight matrix, c t is the context vector, η is the graph neural network weight, GNN is the graph neural network function, H t is the node feature matrix at the tth moment, and A is the adjacency matrix;

[0259] S56. Use the output of the multimodal graph neural network model to perform real-time evaluation and generate an evaluation report R, including leaf nutrition indicators:

[0260] R={N(h T ),P(h T ),K(h T )};

[0261] Among them, N(h T )、P(h T ) and K(h T ) are the estimated values ​​of nitrogen, phosphorus and potassium content respectively;

[0262] S57. According to the evaluation result R, dynamically adjust the fertilizer application plan to generate a fertilization plan F:

[0263] F={f1,f2,…,f n};

[0264] Among them, f i is the application amount of the i-th fertilizer.

[0265] In this implementation, S6 includes the following steps:

[0266] S61. Input the multimodal data into the adaptive recursive hash network. The adaptive recursive hash network extracts the high-dimensional features of the leaf image through a multi-layer recursive hash function and maps it to a low-dimensional hash space:

[0267]

[0268] Among them, p is the parameter to be optimized of the recursive hash network, x i and x i ' are the input image and its enhanced image, H is the recursive hash function, z j is a latent variable, α is a regularization parameter;

[0269] S62. Design a deep artifact removal algorithm based on a sparse encoder-decoder network. The sparse encoder extracts sparse features of the leaf image, and the sparse decoder reconstructs the image:

[0270]

[0271] Among them, x i is the i-th input image, E is the sparse encoder function, D is the sparse decoder function, h j is the jth hidden layer feature, D SEDN is the artifact removal decoder function, β is the regularization parameter;

[0272] S63, input the processed leaf image into a multi-scale wavelet transform module, and use wavelet transform to decompose the image into different scales:

[0273]

[0274] Among them, W m,n is the wavelet coefficient of the mth scale and the nth time, x(t) is the input image signal, ψ is the mother wavelet function, δ is the time interval, s m is the scale parameter;

[0275] S64, reconstructing the image using multi-scale wavelet coefficients and inputting it into a multimodal spatiotemporal convolutional network. The multimodal spatiotemporal convolutional network detects and analyzes the nutritional status of leaves in combination with spatiotemporal features. The optimization objective function is:

[0276]

[0277] Among them, W k is the multi-scale wavelet coefficient, FMSTCN is the spatiotemporal convolution function, y k is the leaf nutritional status label, H t is the spatiotemporal feature, G MSTCN is the spatiotemporal convolutional network function, γ is the regularization coefficient;

[0278] S65. Generate leaf nutrition status diagnosis results, use the spatiotemporal features output by the multimodal spatiotemporal convolutional network, and perform classification through the fully connected layer. The classification results include the nitrogen, phosphorus, and potassium contents of the leaves:

[0279] y=softmax(W fc h+b fc );

[0280] Among them, softmax is the activation function, W fc and b fc are the weight and bias of the fully connected layer respectively, and h is the spatiotemporal feature of the multimodal spatiotemporal convolutional network model.

[0281] In this implementation, S7 includes the following steps:

[0282] S71. Design a cognitive intelligent agent model. The cognitive intelligent agent model simulates the knowledge and experience of agricultural experts through a Bayesian reasoning system to build a cognitive model:

[0283]

[0284] Where E is the set of leaf nutritional state variables, H is the set of environmental data variables, P(E|H) is the posterior probability of leaf nutritional state under environmental data, P(H|E) is the likelihood of leaf nutritional state for environmental data, P(E) is the prior probability of leaf nutritional state, and P(H) is the marginal probability of environmental data;

[0285] S72. Build a multi-agent deep reinforcement learning system, where each agent represents a crop area, and use the deep deterministic policy gradient algorithm for collaborative optimization, policy update and Q value update:

[0286]

[0287] Among them, θ π is the parameter of the policy network, α is the learning rate, J(θ π ) is the policy objective function, θ Q are the parameters of the Q-value network, β is the learning rate, r is the reward value, γ is the discount factor, Q' and π' are the target networks, s and a are the state and action respectively;

[0288] S73. Using a multi-objective differential evolution algorithm, combined with leaf nutrition status diagnosis results and environmental data, an optimal fertilization strategy is formed, and the optimization objective function is:

[0289]

[0290] Among them, F is the fertilization strategy, N j , P j and K j are the nitrogen, phosphorus and potassium contents in the jth crop area, and are the target nitrogen, phosphorus and potassium contents, w1, w2, w3 and w4 are weight coefficients, and C(F) is the fertilization cost;

[0291] S74, realize real-time adjustment of fertilization scheme, and optimize fertilization strategy by combining recursive neural network:

[0292] F(t+1)=F(t)+α·(F best -F(t))+β·(F rand -F(t));

[0293] Among them, F(t) is the current fertilization strategy, F best is the optimal fertilization strategy, F rand is the randomly selected fertilization strategy, α and β are the adjustment coefficients;

[0294] S75. Through the closed-loop feedback control system, the crop growth and environmental changes are continuously monitored, and the fertilization strategy is adjusted according to the real-time data:

[0295] ΔF(t)=η(E(t)-E * (t))+ζ(H(t)-H * (t));

[0296] Among them, ΔF(t) is the fertilizer adjustment amount at the tth moment, E(t) and H(t) are the current leaf nutrition status and environmental data, respectively, and E * (t) and H * (t) is the target state, η and ζ are the adjustment coefficients.

[0297] Embodiment 1:

[0298] In a certain agricultural demonstration area, a large area of ​​rice is planted. For a long time, the area has relied on traditional manual observation and laboratory analysis methods to diagnose the nutritional status of leaves and adjust fertilizers. However, this method is inefficient and difficult to achieve real-time monitoring and rapid response, resulting in uneven crop growth, low fertilizer utilization efficiency, and even environmental pollution. Therefore, the leaf nutritional status diagnosis and fertilizer regulation method based on image recognition of the present invention is introduced into the demonstration area in order to improve the accuracy and efficiency of crop management and optimize fertilizer use.

[0299] High-definition cameras and drones are installed in the demonstration area to collect images of rice leaves. The cameras are installed on ground equipment, and the drones fly regularly to obtain images of leaves in different areas. By setting the shooting frequency and time interval, it is ensured that high-quality leaf image data can be collected regularly.

[0300] The collected images are first processed by denoising, image enhancement and background removal to extract the color, shape and texture features of the leaves and generate leaf feature vectors. These feature vectors are modeled by nonlinear dynamic chaotic systems, the Lyapunov exponents are calculated to detect abnormal changes, and multi-scale chaos analysis is performed.

[0301] Next, we use high-dimensional topological data analysis technology to construct a topological network of leaf nutritional status data, and combine it with the Bayesian network to capture the causal relationship and probabilistic dependency between variables, and perform multimodal data fusion. Based on the fusion results, the system dynamically adjusts the algorithm parameters to ensure the accuracy and real-time nature of the diagnosis results.

[0302] By generating adversarial networks and deep artifact removal algorithms, high-quality leaf images are generated to further improve diagnostic accuracy. Combined with multi-view image fusion technology, accurate leaf nutritional status diagnosis results are ultimately generated.

[0303] Build cognitive intelligent agents and group intelligence optimization algorithms to perform dynamic environmental adaptation and closed-loop feedback control. Intelligent agents simulate the knowledge and experience of agricultural experts and provide accurate fertilizer application recommendations based on real-time nutritional status data and environmental data. The system adjusts fertilization plans in real time based on these recommendations to achieve intelligent and precise crop nutrition management.

[0304] From April to October 2023, the method of the present invention was applied in the demonstration area, and detailed data recording and analysis were carried out in comparison with traditional methods.

[0305] Table 1 Comparison of diagnostic efficiency and accuracy

[0306] project Traditional methods Method of the present invention Diagnosis time (hours) 48 2 Diagnostic accuracy (%) 70 95 Abnormal detection rate (%) 50 90 Real-time Difference excellent

[0307] Table 2 Fertilizer use efficiency and environmental impact

[0308]

[0309] Table 3 Comparison of crop growth and yield

[0310] project Traditional methods Method of the present invention Crop growth uniformity (%) 70 95 Production increase (%) 0 20 Leaf health index (%) 65 90 Nutritional deficiency rate (%) 40 5

[0311] Through the above comparison, it can be seen that the method of the present invention is superior to the traditional method in terms of diagnostic efficiency, accuracy, real-time performance, fertilizer utilization efficiency, environmental impact, crop growth uniformity and yield. Specifically, the diagnosis time is reduced from 48 hours to 2 hours, the accuracy rate is increased from 70% to 95%, and the abnormal detection rate is increased from 50% to 90%. The total amount of fertilizer used is reduced by 30%, the fertilizer utilization rate is increased to 90%, the environmental pollution index is reduced by 66%, and the resource saving rate reaches 30%. In addition, the uniformity of crop growth is increased to 95%, the yield is increased by 20%, the leaf health index is increased to 90%, and the nutrient deficiency rate is significantly reduced.

[0312] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. Leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition, characterized in that: The following steps are involved: S1. Collect images of crop leaves using a camera or drone; S2, performing denoising, image enhancement and background removal processing on the collected leaf image, extracting the color, shape and texture features of the leaf image, and generating a leaf feature vector; S3, using nonlinear dynamic chaotic system to model blade eigenvectors, calculating Lyapunov exponents to detect abnormal changes, and performing multi-scale chaos analysis; S4. Use high-dimensional topological data analysis technology to construct a topological network of leaf nutritional status data, combine it with Bayesian network to capture the causal relationship and probability dependence between variables, and perform multimodal data fusion; S5. Dynamically adjust algorithm parameters according to the multimodal data fusion results; S6, generating images through multi-view image fusion and deep artifact removal algorithm of generative adversarial network to generate leaf nutrition status diagnosis results; S7. Build intelligent agents with cognitive capabilities and swarm intelligence optimization algorithms to work together to perform dynamic environmental adaptation and closed-loop feedback control.

2. The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition according to claim 1 is characterized in that: The step S1 comprises: S11, selecting a camera or a drone as an image acquisition device to obtain a leaf image; S12, adjusting the shooting parameters of the camera or the drone so that leaf images can be captured under different lighting conditions; S13, setting the shooting frequency and time interval of the camera or drone, and regularly collecting leaf images according to the crop growth cycle and monitoring requirements; S14. During the shooting process, the location information of the camera or drone is recorded, and the latitude, longitude and altitude data of the shooting point are obtained through GPS or other positioning systems; S15, performing a preliminary inspection on the collected leaf images; S16. Store the preliminarily inspected blade image and its position information in a data storage device.

3. The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition according to claim 1, characterized in that: The step S2 comprises: S21, performing denoising on the collected leaf image, using Gaussian filtering or median filtering algorithm to remove noise in the image, retaining the main feature information of the leaf, and obtaining a denoised image; S22, performing image processing on the denoised leaf image by using a histogram equalization or a contrast-limited adaptive histogram equalization algorithm; S23, performing background removal processing, using threshold segmentation or edge detection algorithm to perform background removal processing; S24, extracting color features from the leaf image after background removal, calculating the average color value, color histogram and color distribution of the leaf, and generating a color feature vector; S25, extracting shape features from the leaf image after background removal, calculating feature parameters such as the leaf contour, area, perimeter and shape moment, and generating a shape feature vector; S26, extracting texture features from the leaf image after background removal, calculating the texture features of the leaf using a gray-level co-occurrence matrix or a local binary pattern algorithm, and generating a texture feature vector; S27, merging the color feature vector, the shape feature vector and the texture feature vector to generate a comprehensive leaf feature vector.

4. The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition according to claim 1, characterized in that: The step S3 comprises: S31. Construct a nonlinear dynamic chaotic system model integrating multi-source data assimilation and define the state equation: x(t+1)=F(x(t),p)+N(t); Among them, x(t) is the blade characteristic vector at the current moment, F is the nonlinear mapping function, p is the system parameter vector, and N(t) is the random noise vector; S32. Use multi-modal collaborative chaos optimization algorithm to optimize the chaotic system parameter p: S321. Generate an initial population P(0) in the parameter space, where each individual represents a possible parameter vector: P(0)={p1(0),p2(0),…,p N (0)}; S322, different subpopulations P i (t) co-evolution, using mutation and crossover operations to simulate the natural selection process: p i (t+1)=p i (t)+α·(p j (t)-p k (t))+β·G i (t); Among them, p i (t), p j (t) and p k (t) is the parameter vector of the i-th, j-th and k-th individuals, α is the variation factor, β is the crossover factor, G i (t) is the guide vector, defined as: Among them, γ is the learning rate, w m is the weight vector, x m (t) is the mth modal eigenvector, M is the number of modes; S323, introduce improved Logistic mapping to optimize the population, increase the diversity of the search space and the global exploration ability, and avoid local optimal solutions: p i (t+1)=μ·p i (t)·(1-p i (t)); Among them, μ is the chaos parameter, f(p i (t)) is the fitness function, f max is the maximum fitness; S324, through the online learning mechanism, dynamically adjust the evolution strategy and chaos mapping parameters according to environmental changes: p i (t+1)=p i (t)+η·(p best (t)-p i (t)); Among them, η is the adaptive learning rate, p best (t) is the current optimal parameter vector; S325, judging whether the optimization process is finished according to the convergence of the objective function, and outputting the optimized system parameter vector p MMCCOA ; S33, the optimized system parameter p MMCCOA Input into the chaotic system model, calculate the Lyapunov exponent λ, and detect the chaotic characteristics and abnormal changes of the system: δx(i)=x(i)-x′(i); Among them, δx(i) is the disturbance vector at the i-th moment, x'(i) is the blade characteristic vector after introducing a small disturbance, S34. By calculating the growth rate of the disturbance vector at each moment, the chaotic characteristics and abnormal changes of the system are judged. When λ>0, the system is in a chaotic state and there are abnormal changes. S35. Perform multi-scale chaos analysis, use adaptive multi-scale transformation to decompose the multi-modal feature vector, and extract features at different time scales: Among them, ψ j,k (t) is the adaptive multi-scale transformation basis function, c j (k) is the coefficient of the jth layer; S36, extract the characteristic vector x of each time scale through adaptive multi-scale analysis method j (t); S37. Use the Bayesian dynamic fusion-time series prediction algorithm to combine the multi-scale analysis results with the optimized chaotic system model: in, Optimize the LSTM network weights for the optimized Bayesian inference. P(D|w) is the likelihood function under Bayesian inference. is the predicted leaf feature vector, F LSTM Optimizing the prediction function of long short-term memory networks for Bayesian reasoning that integrates multi-scale features and chaotic system models; S38. Comprehensively evaluate abnormal changes in leaf nutritional status.

5. The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition according to claim 1, characterized in that: The step S4 comprises: S41, construct a multimodal data matrix X, integrating multimodal data such as leaf images, soil composition, and climate conditions, where X = [x1, x2, …, x N ], each x i is a data vector containing different modal features; S42. Calculate the similarity matrix S between data samples using the improved Gaussian kernel: Among them, S ij is the similarity between the i-th and j-th samples, σ is the bandwidth parameter of the kernel function; S43. Use high-dimensional topological data analysis technology to analyze the similarity matrix S, construct a topological network K, and obtain topological features by calculating persistent homology: Among them, β k is the kth order Betti number, is a k-order boundary operator; S44. Construct a spatiotemporal dynamic Bayesian network to capture the complex spatiotemporal dependencies between variables, and use graph convolutional networks and temporal recurrent neural networks to extract spatiotemporal features: H (l+1) =σ(AH (l) W (l) ); h t =σ(W h x t +U h h t-1 +b h ) Among them, H (l) is the node feature matrix of the lth layer, A is the adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function, h t is the hidden state at time t, W h , U h and b h are the weight matrix and bias vector respectively; S45, perform multimodal data fusion and use adaptive variational autoencoder to generate high-dimensional fusion features; where z is the latent variable, μ(x) and σ(x) are the mean and standard deviation of the encoder parameterization, and μ θ (z) and is the mean and standard deviation of the decoder parameterization; S46. Design an adaptive quantum immune memory module, combining quantum computing and immune algorithms to record and update historical data: Among them, α i is the complex amplitude, |i> is the ground state; S47. Use the quantum clone selection algorithm to update the memory bank, select the quantum state with the highest fitness for cloning, perform mutation operations on the quantum state, measure the quantum state, and obtain classical memory cells: M(|ψ>)={|ψ1>,|ψ2>,…,|ψ m >}; Among them, U mutation |ψ> is the quantum mutation operator, (|ψ i > is the classical state after measurement; S48, the newly collected data sample x new Match with quantum immune memory bank M, generate new diagnostic results and update the memory bank to form a comprehensive analysis report: x new →Match(M)→Update(M).

6. The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition according to claim 1, characterized in that: The step S5 comprises: S51, the multimodal data fusion result Z is input into the adaptive multimodal hybrid generative adversarial network, and the adaptive multimodal hybrid generative adversarial network combines multimodal data, generative adversarial network and variational autoencoder. The objective function is: Generator objective function: Discriminator objective function: Among them, p is the parameter vector to be optimized for the generator, q is the parameter vector to be optimized for the discriminator, G is the generator function, D is the discriminator function, E is the encoder function, λ is the regularization parameter, and z is the latent variable; S52. Construct a proxy model based on a multimodal graph neural network and a variational generative network. The multimodal graph neural network is used to extract graph structure and multimodal features, and the variational generative network is used to generate high-dimensional latent features: H (l+1) =σ(AH (l) W (l) )+βT (l) ; q φ (z|x)=N(z;μ(x),σ 2 (x)); Among them, H (l) is the node feature matrix of the lth layer, A is the adjacency matrix, W (l) is the weight matrix of the lth layer, σ is the activation function, T (l) is the multimodal feature matrix, β is the fusion weight; z is the latent variable, μ(x) and σ(x) are the mean and standard deviation of the encoder, μ θ (z) and is the mean and standard deviation of the decoder; S53, using the expected improvement strategy of adaptive multi-modal generation adversarial optimization to perform parameter optimization and calculate the expected improvement value: Among them, p + is the current optimal parameter, G(z|p) is the generator function value, and D is the discriminator function value; S54, the optimized parameter p * Input into the leaf nutrition status assessment model to build an adaptive multimodal generative adversarial network. The leaf nutrition status assessment model combines graph neural network and generative adversarial network: G(z|p)=Decoder(z;p); D(x|p)=σ(W d xb d ); Among them, z is the latent variable, p is the optimization parameter, and W d and b d are the weight and bias of the discriminator, σ is the activation function; S55. Optimizing the Adaptive Multimodal Generative Adversarial Network Model Using the Multimodal Graph Neural Network Attention Mechanism: a t =softmax(W a [h t ;c t ])+ηGNN(H t ,A); Among them, a t is the attention weight at the tth moment, W a is the attention weight matrix, c t is the context vector, η is the graph neural network weight, GNN is the graph neural network function, H t is the node feature matrix at the tth moment, and A is the adjacency matrix; S56. Use the output of the multimodal graph neural network model to perform real-time evaluation and generate an evaluation report R, including leaf nutrition indicators: R={N(h T ),P(h T ),K(h T )}; Among them, N(h T )、P(h T ) and K(h T ) are the estimated values ​​of nitrogen, phosphorus and potassium content respectively; S57. According to the evaluation result R, dynamically adjust the fertilizer application plan to generate a fertilization plan F: F={f1,f2,…,f n }; Among them, f i is the application amount of the i-th fertilizer.

7. The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition according to claim 1, characterized in that: The step S6 comprises: S61. Input the multimodal data into the adaptive recursive hash network. The adaptive recursive hash network extracts the high-dimensional features of the leaf image through a multi-layer recursive hash function and maps it to a low-dimensional hash space: Among them, p is the parameter to be optimized of the recursive hash network, x i and x i ' are the input image and its enhanced image, H is the recursive hash function, z j is a latent variable, α is a regularization parameter; S62. Design a deep artifact removal algorithm based on a sparse encoder-decoder network. The sparse encoder extracts sparse features of the leaf image, and the sparse decoder reconstructs the image: Among them, x i is the i-th input image, E is the sparse encoder function, D is the sparse decoder function, h j is the jth hidden layer feature, D SEDN is the artifact removal decoder function, β is the regularization parameter; S63, input the processed leaf image into a multi-scale wavelet transform module, and use wavelet transform to decompose the image into different scales: Among them, W m,n is the wavelet coefficient of the mth scale and the nth time, x(t) is the input image signal, ψ is the mother wavelet function, δ is the time interval, s m is the scale parameter; S64, reconstructing the image using multi-scale wavelet coefficients and inputting it into a multimodal spatiotemporal convolutional network. The multimodal spatiotemporal convolutional network detects and analyzes the nutritional status of leaves in combination with spatiotemporal features. The optimization objective function is: Among them, W k is the multi-scale wavelet coefficient, F MSTCN is the spatiotemporal convolution function, y k is the leaf nutritional status label, H t is the spatiotemporal feature, G MSTCN is the spatiotemporal convolutional network function, γ is the regularization coefficient; S65. Generate leaf nutrition status diagnosis results, use the spatiotemporal features output by the multimodal spatiotemporal convolutional network, and perform classification through the fully connected layer. The classification results include the nitrogen, phosphorus, and potassium contents of the leaves: y =softmax(W fc h+b fc ): Among them, softmax is the activation function, W fc and b fc are the weight and bias of the fully connected layer respectively, and h is the spatiotemporal feature of the multimodal spatiotemporal convolutional network model.

8. The leaf nutrition status diagnosis and fertilizer adjustment method based on image recognition according to claim 1, characterized in that: The step S7 comprises: S71. Design a cognitive intelligent agent model. The cognitive intelligent agent model simulates the knowledge and experience of agricultural experts through a Bayesian reasoning system to build a cognitive model: Where E is the set of leaf nutritional state variables, H is the set of environmental data variables, P(E|H) is the posterior probability of leaf nutritional state under environmental data, P(H|E) is the likelihood of leaf nutritional state for environmental data, P(E) is the prior probability of leaf nutritional state, and P(H) is the marginal probability of environmental data; S72. Build a multi-agent deep reinforcement learning system, where each agent represents a crop area, and use the deep deterministic policy gradient algorithm for collaborative optimization, policy update and Q value update: Among them, θ π is the parameter of the policy network, α is the learning rate, J(θ π ) is the policy objective function, θ Q are the parameters of the Q-value network, β is the learning rate, r is the reward value, γ is the discount factor, Q' and π' are the target networks, s and a are the state and action respectively; S73. Using a multi-objective differential evolution algorithm, combined with leaf nutrition status diagnosis results and environmental data, an optimal fertilization strategy is formed, and the optimization objective function is: Among them, F is the fertilization strategy, N j , P j and K j are the nitrogen, phosphorus and potassium contents in the jth crop area, and are the target nitrogen, phosphorus and potassium contents, w1, w2, w3 and w4 are weight coefficients, and C(F) is the fertilization cost; S74, realize real-time adjustment of fertilization scheme, and optimize fertilization strategy by combining recursive neural network: F(t+1)=F(t)+α·(F best -F(t))+β·(F rand -F(t)); Among them, F(t) is the current fertilization strategy, F best is the optimal fertilization strategy, F rand is the randomly selected fertilization strategy, α and β are the adjustment coefficients; S75. Through the closed-loop feedback control system, the crop growth and environmental changes are continuously monitored, and the fertilization strategy is adjusted according to the real-time data: ΔF(t)=η(E(t)-E * (t))+ζ(H(t)-H * (t)); Among them, ΔF(t) is the fertilizer adjustment amount at the tth moment, E(t) and H(t) are the current leaf nutrition status and environmental data, respectively, and E * (t) and H * (t) is the target state, η and ζ are the adjustment coefficients.

Images