An intelligent identification and prevention system for diseases and insect pests of ancient trees based on reinforcement learning

Abstract

The application discloses an ancient tree disease and pest intelligent identification and prevention system based on reinforcement learning, comprising: a data acquisition module for collecting image data and environmental monitoring data and preprocessing to generate a multi-source monitoring sequence; a feature extraction module for feature extraction and fusion of the multi-source monitoring sequence to generate a fusion feature sequence; a mapping identification module for performing frequency domain mapping and graph structure updating operations to generate a disease and pest identification sequence and a health state sequence; a spatial calculation module for constructing a spatial adjacency graph to generate a spatial feature sequence; a strategy generation module for aligning the health state sequence and the spatial feature sequence and generating a target prevention sequence using an A3C algorithm; and a feedback update module for executing the target prevention sequence and updating A3C algorithm parameters according to feedback data. The application has the advantages of high identification accuracy, strong adaptability and sustainable optimization by using the A3C algorithm.

Description

Technical Field

[0001] This invention relates to the field of intelligent identification technology, and in particular to an intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning. Background Technology

[0002] With the increasing awareness of urban ecological protection, the health monitoring and pest control of ancient and famous trees, as important natural heritage and ecological resources, are receiving more and more attention. Traditional protection of ancient and famous trees mainly relies on manual inspections. Staff regularly observe the tree crown shape, leaf color, bark damage, and signs of pests to determine the type and severity of pests and diseases, and then develop control plans based on experience. This method depends on human experience, the identification process is highly subjective, and it is difficult to accurately judge subtle lesions or early-stage pests. Furthermore, when ancient trees are widely distributed and numerous, manual inspections are inefficient and cannot achieve continuous, dynamic monitoring over time.

[0003] To improve identification efficiency, existing technologies have gradually introduced pest and disease identification methods based on image processing and convolutional neural networks. These methods achieve automatic identification by extracting features and classifying collected images of leaves or tree trunks. However, these methods often focus on processing single image data and lack comprehensive modeling of environmental factors such as temperature, humidity, light intensity, and soil moisture content. They also fail to construct a unified multi-source monitoring sequence, leading to discrepancies between the identification results and the actual ecological state. Furthermore, existing identification models typically analyze individual trees independently, neglecting the spatial adjacency relationships between ancient trees and the transmission paths of pests and diseases. This fails to characterize the spatial spread characteristics of pests and diseases, resulting in identification results lacking spatial correlation.

[0004] Regarding prevention and control strategies, existing technologies mostly employ preset rules or expert experience bases, outputting fixed prevention and control measures based on the identification results, such as spraying pesticides or pruning diseased branches. This approach lacks a joint modeling structure of health state sequences and spatial feature sequences, making it impossible to make dynamic decisions based on the health change trends and spatial location relationships of different ancient trees at different time steps. Some studies have attempted to introduce reinforcement learning methods to generate prevention and control strategies, but these often focus on simple state spaces and single-objective optimization, lacking the fusion processing of graph structure propagation features and spatial embedding results, making it difficult to achieve parallel decision-making and value assessment for multiple ancient trees. Furthermore, existing reinforcement learning methods often lack a complete feedback update mechanism in practical applications, failing to continuously optimize the strategy network parameters based on pest and disease change data after prevention and control implementation, resulting in slow strategy convergence speed and insufficient stability.

[0005] Therefore, how to provide an intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning. This invention fully utilizes the DeepLabV3+ model, Fourier neural operators, GATv2 network, Node2Vec algorithm, DiffPool model, and A3C algorithm to construct a complete technical path of multi-source monitoring sequence fusion, spatial adjacency modeling, graph structure feature propagation, and closed-loop optimization of strategy generation. It realizes dynamic identification of diseases and pests of ancient and famous trees and adaptive generation of control strategies, and has the advantages of high identification accuracy, strong spatial propagation characterization ability, strong adaptiveness of strategy generation, and stable continuous optimization ability.

[0007] According to an embodiment of the present invention, an intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning includes:

[0008] The data acquisition module is used to collect and preprocess image data and environmental monitoring data of the target ancient and famous trees at continuous time steps, generate multi-source monitoring sequences, and collect the spatial coordinates and spatial adjacency relationships of the target ancient and famous trees.

[0009] The feature extraction module is used to perform semantic segmentation and fully connected mapping operations on multi-source monitoring sequences using the DeepLabV3+ model, extract image features and environmental features, and perform fusion processing to generate fused feature sequences.

[0010] The mapping and recognition module is used to perform frequency domain mapping operation on the fused feature sequence using Fourier neural operators, construct a graph structure based on the mapping result and spatial adjacency relationship, update the node features of the graph structure through the GATv2 network, generate a pest and disease recognition sequence and connect it with the fused feature sequence to obtain a health status sequence.

[0011] The spatial computing module is used to construct a spatial adjacency graph based on spatial coordinates and spatial adjacency relationships. It uses the Node2Vec algorithm to perform node embedding calculations on the spatial adjacency graph and performs hierarchical pooling operations on the embedding results through the DiffPool model to generate spatial feature sequences.

[0012] The strategy generation module is used to align the health status sequence and the spatial feature sequence by time step, use the A3C algorithm to perform prevention and control strategy generation and value assessment on the alignment result, and select target prevention and control vectors according to the assessment results and arrange them by time step to form the target prevention and control sequence.

[0013] The feedback update module is used to execute the target prevention sequence, collect feedback data, and update the parameters of the A3C algorithm based on the feedback data.

[0014] Optionally, the image data represents a set of digital images containing visual information of the tree crown, trunk, and leaves collected at each time step; the environmental monitoring data represents a set of numerical data consisting of temperature, humidity, light intensity, and soil moisture content; the preprocessing includes performing resolution unification, noise reduction, and time alignment operations on the image data; and performing missing value imputation, dimension unification, and time alignment operations on the environmental monitoring data; the spatial coordinates represent the three-dimensional position vector of the target ancient and famous trees in a unified spatial reference system; the spatial adjacency relationship represents a set of connection relationships between ancient trees constructed according to a preset spatial distance threshold; and the pest and disease identification sequence includes pest and disease category identifiers and pest and disease severity identifiers.

[0015] Optionally, the feature extraction module includes:

[0016] Image data from each time step in the multi-source monitoring sequence are cropped to a preset size and input into the DeepLabV3+ model. Semantic feature maps under different receptive fields are extracted from the dilated convolutional structure of the DeepLabV3+ model. Dilated spatial pyramid pooling is performed on the semantic feature maps. Feature concatenation and convolution mapping are performed on the pooling results at each scale to obtain the initial feature map.

[0017] The proportion of lesion pixels in the initial feature map is calculated based on the preset lesion recognition rules. The initial feature map is then weighted according to the proportion of lesion pixels to generate a semantic feature map.

[0018] The semantic feature map is input into the fully connected mapping structure of the DeepLabV3+ model, and vector expansion and linear mapping operations are performed on the semantic feature map to generate image feature vectors.

[0019] The environmental monitoring data in the multi-source monitoring sequence is read, and the four types of environmental components at each time step are arranged in a preset order. The weight mapping operation is performed based on the change amplitude of each environmental component between adjacent time steps to generate an environmental feature vector with the same dimension as the image feature vector. The environmental components represent temperature, humidity, light intensity and soil moisture content.

[0020] In terms of feature dimension, the image feature vector and the environment feature vector are concatenated. The image feature vector is used as the backbone feature, and a linear mapping operation is performed on the environment feature vector. The environment feature vector is then added element-wise with the backbone feature to form the residual feature vector.

[0021] Batch normalization and ReLU activation are performed on the residual feature vectors, and the vectors are arranged according to time steps to form a fused feature sequence.

[0022] Optionally, the semantic feature map generation process specifically includes:

[0023] Image data at each time step is input into the dilated convolutional structure of the DeepLabV3+ model. Convolutional calculations are performed on the image data in multiple convolutional branches at different scales. In each convolutional branch, the convolutional kernel slides along the spatial dimension of the image data with a preset stride. The pixel values ​​within the coverage area of ​​the convolutional kernel are multiplied element-wise with the convolutional kernel weight values, and the results are accumulated to obtain the convolutional feature map at the corresponding scale. The convolutional kernel weight values ​​are taken from the convolutional layer network parameters in the DeepLabV3+ model that match the corresponding convolutional branch.

[0024] A hollow spatial pyramid pooling operation is performed on each convolutional feature map. The pooling window size and pooling stride are set in multiple pooling branches of different scales. The pooling window slides along the spatial dimension of the convolutional feature map according to the pooling stride. The summation operation is performed on all feature values ​​within the pooling window coverage area and divided by the number of feature values ​​in the pooling window to obtain the spatial average value.

[0025] All spatial average values ​​are arranged in the sliding order of the pooling window to form pooling feature maps of the corresponding scales. Then, the pooling feature maps of each scale are spliced ​​together in the channel dimension to obtain the fused feature map.

[0026] In the convolutional mapping structure of the DeepLabV3+ model, a convolutional mapping operation is performed on the fused feature map. The preset convolutional kernel slides in the spatial dimension of the fused feature map with a preset stride. The feature value of each channel within the coverage area of ​​the convolutional kernel is multiplied element by element with the weight value of the corresponding channel, and then accumulated in the channel dimension to obtain the initial feature map. The weight value is taken from the convolutional layer network parameters corresponding to the convolutional mapping structure in the DeepLabV3+ model.

[0027] Based on the preset lesion recognition rules, the feature channel corresponding to the lesion category is determined in the channel dimension of the initial feature map. Each feature value of the feature channel in the spatial dimension is compared with the feature values ​​of all channels at the same spatial location. When the feature value of the feature channel is the largest feature value among all channels and is not less than the preset threshold, the spatial location is marked as the lesion pixel location.

[0028] Traverse all spatial locations, summarize all lesion pixel locations to form a lesion pixel set, count the number of pixels in the lesion pixel set and divide it by the total number of pixels in the initial feature map to obtain the lesion pixel ratio;

[0029] The initial feature map is scaled according to the proportion of lesion pixels. The product of the preset gain coefficient and the proportion of lesion pixels is incremented by one to obtain the scaling coefficient. Each feature value of each channel in the initial feature map is multiplied by the scaling coefficient to obtain the semantic feature map.

[0030] Optionally, the mapping recognition module includes:

[0031] The fused feature vectors at each time step in the fused feature sequence are reconstructed into a two-dimensional feature matrix in the feature dimension, and a discrete Fourier transform operation is performed on the two-dimensional feature matrix in the spatial dimension to map the two-dimensional feature matrix from the spatial domain to the frequency domain, thus obtaining the initial frequency domain matrix.

[0032] In the initial frequency domain matrix, each frequency component is divided into a high-frequency region and a low-frequency region according to a preset frequency division rule. The frequency components in the high-frequency region are subjected to amplitude scaling processing, while the amplitude of the frequency components in the low-frequency region remains unchanged, forming a frequency domain feature matrix. The frequency domain components represent the complex values ​​corresponding to each frequency position in the initial frequency domain matrix.

[0033] Performing an inverse discrete Fourier transform on the frequency domain feature matrix maps the frequency domain feature matrix back to the spatial domain, resulting in a mapped feature matrix.

[0034] A graph structure is constructed based on spatial adjacency relationships. The mapping feature matrix of each target ancient and famous tree is used as the initial node feature of the node in the graph structure, and the spatial adjacency relationship is used as the edge relationship in the graph structure.

[0035] The graph structure is input into the GATv2 network. Attention weights between nodes are calculated on each edge of the graph structure. The features of adjacent nodes are weighted and summed according to the attention weights in the node dimension. The weighted sum is then subjected to linear mapping and LeakyReLU activation to obtain the updated node features.

[0036] A fully connected mapping operation is performed on the updated node features along the channel dimension. The mapping results are divided into pest and disease category components and pest and disease severity components according to the preset dimension segmentation position, and arranged according to the time step to form a pest and disease identification sequence.

[0037] The disease and pest identification sequence and the fused feature sequence are spliced ​​together along the feature dimension, and the splicing result is subjected to linear mapping and batch normalization to generate a health status sequence.

[0038] Optionally, the amplitude scaling process specifically includes:

[0039] In the initial frequency domain matrix, locate the set of frequency indices for the high-frequency region according to the preset frequency division rules, and traverse each frequency index position in the set of frequency indices to read the corresponding complex value.

[0040] For each complex value, the summation operation is performed on the square of the real part and the square of the imaginary part, and then the square root operation is performed to calculate the amplitude of the complex value. All amplitudes are arranged according to frequency position to form a high-frequency amplitude matrix.

[0041] The high-frequency amplitude matrix is ​​probabilistically processed by calculating the sum of all amplitudes in the high-frequency amplitude matrix and dividing the amplitude at each frequency position by the sum to obtain the probability value, thus obtaining the high-frequency probability matrix.

[0042] In the high-frequency probability matrix, each frequency position is traversed with a preset neighborhood window. The set of probability values ​​in the neighborhood window corresponding to the frequency position is extracted. Logarithmic and multiplication operations are performed on each item of the probability value set. The multiplication results are accumulated and then the accumulated result is negative to obtain the local spectral entropy value of the frequency position. All local spectral entropy values ​​are arranged according to frequency position to form a spectral entropy matrix.

[0043] A threshold segmentation mapping operation is performed on the spectral entropy matrix. Based on the preset spectral entropy threshold, the spectral entropy matrix is ​​divided into enhancement regions and preservation regions. A corresponding weight value is generated for each frequency position in the enhancement region, and a weight value of one is generated for each frequency position in the preservation region. The weights are arranged according to the frequency positions to form a weight matrix.

[0044] The weight matrix is ​​multiplied by the amplitude at each frequency position while keeping the phase of the complex value at that frequency position unchanged, to obtain the enhanced frequency component.

[0045] Optionally, the spatial computing module includes:

[0046] Read the spatial coordinates and spatial adjacency relationships of the target ancient and famous trees, determine the sequence of adjacent node pairs based on the spatial adjacency relationships, calculate the spatial coordinate difference for each adjacent node pair as the spatial distance value, and arrange all spatial distance values ​​in the order of the adjacent node pair sequence to form a spatial distance sequence.

[0047] A spatial adjacency graph is constructed based on spatial adjacency relationships. Each target ancient and famous tree is defined as a node in the spatial adjacency graph, and spatial adjacency relationships are defined as edges in the spatial adjacency graph. The spatial distance sequence is written into the edge attribute set in edge order to generate a spatial adjacency graph with edge attribute set.

[0048] The Node2Vec algorithm is used to perform random walk operations in the spatial adjacency graph. Based on the spatial adjacency relationship, multiple fixed-length walk sequences are generated for each node, and all walk sequences are arranged by node index to form a walk sequence set.

[0049] Skip-Gram embedding training is performed on the wander sequence set. The center node and context node are extracted from the wander sequence set according to a fixed window length to form a node pair sample set. Vector mapping and similarity constraint calculation are performed on the node pair sample set to obtain the node embedding vector corresponding to each node. All node embedding vectors are arranged according to the node index to form a node embedding sequence.

[0050] The node embedding sequence is input into the DiffPool model, and clustering allocation calculation is performed to generate an allocation matrix. The allocation matrix is ​​normalized to construct hierarchical aggregation relationships. Based on the hierarchical aggregation relationships, hierarchical pooling operations are performed on the node embedding sequence to obtain a multi-level pooled embedding set.

[0051] The pooled embedding sets are concatenated in hierarchical order, and linear mapping and batch normalization operations are performed on the concatenation results to generate spatial feature sequences.

[0052] Optionally, the random walk operation specifically includes:

[0053] In the spatial adjacency graph, each node is selected sequentially according to the node index order. For each node, the corresponding walk sequence is initialized. The node is written into the corresponding walk sequence as the starting node and the current node. At the same time, the node before the starting node is marked as null.

[0054] Read all adjacent nodes of the current node in the spatial adjacency graph, find the spatial distance value between the current node and each adjacent node in the edge attribute set, perform the reciprocal calculation on the spatial distance value to obtain the initial transition value, perform a summation operation on all initial transition values, and divide each initial transition value by the summation result to obtain the transition probability set;

[0055] When the current node is null, select the target node from the adjacent nodes according to the transition probability set;

[0056] When the previous node is not null, for each adjacent node, determine whether the adjacent node has an edge connection with the previous node in the spatial adjacency graph. If there is an edge connection, keep the transition probability value unchanged. If there is no edge connection, multiply the corresponding transition probability value by a preset attenuation coefficient, and perform summation and normalization operations on the product result to obtain a new set of transition probabilities. Then, continue to select target nodes according to the new set of transition probabilities.

[0057] Write the target node into the current traversal sequence, update the previous node to the current node, and update the current node to the target node;

[0058] Determine whether the current walking sequence length has reached the preset walking length. If it has not reached the preset walking length, repeat the transition probability calculation, node selection and node writing operations. If it has reached the preset walking length, end the generation of the current walking sequence.

[0059] For each node in the spatial adjacency graph, the walk sequence generation operation is repeated to obtain multiple walk sequences. All walk sequences are then arranged in the order of node indices to form a walk sequence set.

[0060] Optionally, the calculation process of the DiffPool model specifically includes:

[0061] The node embedding sequence is reconstructed into a node feature matrix according to the node index order. At the same time, the adjacency relationship of the spatial adjacency graph is represented as an adjacency matrix, and symmetry and normalization operations are performed on the adjacency matrix to obtain a normalized adjacency matrix.

[0062] Perform a linear mapping operation on the node feature matrix and perform matrix multiplication with the normalized adjacency matrix to obtain the clustering score matrix;

[0063] Perform exponential and normalization operations on the clustering score matrix along the clustering dimension to obtain the assignment matrix;

[0064] Perform matrix multiplication between the transpose of the allocation matrix and the node feature matrix to obtain the pooling feature matrix;

[0065] After performing matrix multiplication on the transpose of the allocation matrix and the normalized adjacency matrix, perform matrix multiplication again on the allocation matrix to obtain the pooled adjacency matrix.

[0066] The pooling feature matrix and pooling adjacency matrix are used as inputs. The allocation matrix calculation and pooling operation are performed iteratively until the preset number of iterations is reached, generating a multi-level pooling embedding set.

[0067] Optionally, the strategy generation module includes:

[0068] Perform a time alignment operation on the health state sequence and the spatial feature sequence, and concatenate the aligned health state vector and the spatial feature vector along the feature dimension to generate an aligned feature sequence;

[0069] The aligned feature sequence is input into the policy network of the A3C algorithm step by step. Based on the policy weight matrix and bias vector in the policy network, matrix multiplication and bias addition are performed layer by layer on each aligned feature vector, and ReLU function activation is performed to obtain the action mapping vector.

[0070] In the action dimension, exponential calculation and normalization operations are performed on the action mapping vector to generate the prevention and control probability sequence at each time step;

[0071] Based on the prevention and control probability sequence, at each time step, the corresponding prevention and control action vector is selected from the preset prevention and control action set as a candidate prevention and control set;

[0072] The aligned feature sequence is input into the value network of the A3C algorithm, and matrix multiplication and bias addition are performed on the aligned feature vectors at each time step according to the value weight matrix and bias vector in the value network to generate the state value sequence.

[0073] In the time dimension, the difference between adjacent time steps is calculated on the state value sequence to obtain the time difference sequence. The time difference sequence is matched with the candidate prevention and control set in the time step dimension. The time difference is used as the screening criterion. In the same time step, the prevention and control action vector with the positive time difference and the largest value is taken as the target prevention and control vector.

[0074] All target prevention and control vectors are arranged in chronological order to form a target prevention and control sequence.

[0075] The beneficial effects of this invention are:

[0076] First, this invention constructs a multi-source monitoring sequence, fusing image data and environmental monitoring data at a unified time step. By combining the semantic segmentation structure of the DeepLabV3+ model and the frequency domain mapping mechanism of the Fourier neural operator, it enhances the expression of lesion areas and high-frequency abnormal features, achieving refined extraction and structured modeling of pest and disease features. This significantly improves the accuracy and robustness of pest and disease identification sequences, effectively solving the problems of large identification errors and difficulty in early disease detection caused by existing technologies relying solely on single visual information.

[0077] Secondly, this invention constructs a spatial adjacency graph containing spatial distance attributes and introduces the Node2Vec algorithm and DiffPool model to perform hierarchical embedding modeling of the spatial structure, transforming the spatial propagation relationship between ancient trees into a computable spatial feature sequence. At the same time, it combines the GATv2 network to perform attention-weighted updates on node features, enabling the spatial propagation influence to participate in the construction of the health status sequence, thereby improving the ability to characterize the spatial spread trend of pests and diseases and overcoming the problem of lack of global consistency in judgment results caused by the failure to consider spatial correlation in existing technologies.

[0078] Finally, this invention constructs a reinforcement learning decision structure based on health status sequences and spatial feature sequences, realizes joint computation of policy networks and value networks through the A3C algorithm, and continuously updates network parameters based on feedback data after prevention and control implementation, forming a closed-loop mechanism that couples identification, spatial modeling, policy generation and feedback optimization. This enables prevention and control strategies to be dynamically adjusted and gradually converged over time, improving the adaptability, stability and resource utilization efficiency of prevention and control decisions. Attached Figure Description

[0079] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0080] Figure 1 This is a module structure diagram of an intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning, as proposed in this invention.

[0081] Figure 2 This is a flowchart of the mapping and recognition process for an intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning, as proposed in this invention.

[0082] Figure 3 This is a flowchart of the spatial computation and strategy generation process for an intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning, as proposed in this invention. Detailed Implementation

[0083] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0084] refer to Figures 1-3 A reinforcement learning-based intelligent identification and control system for diseases and pests of ancient and famous trees, comprising:

[0085] The data acquisition module is used to collect and preprocess image data and environmental monitoring data of the target ancient and famous trees at continuous time steps, generate multi-source monitoring sequences, and collect the spatial coordinates and spatial adjacency relationships of the target ancient and famous trees.

[0086] The feature extraction module is used to perform semantic segmentation and fully connected mapping operations on multi-source monitoring sequences using the DeepLabV3+ model, extract image features and environmental features, and perform fusion processing to generate fused feature sequences.

[0087] The mapping and recognition module is used to perform frequency domain mapping operation on the fused feature sequence using Fourier neural operators, construct a graph structure based on the mapping result and spatial adjacency relationship, update the node features of the graph structure through the GATv2 network, generate a pest and disease recognition sequence and connect it with the fused feature sequence to obtain a health status sequence.

[0088] The spatial computing module is used to construct a spatial adjacency graph based on spatial coordinates and spatial adjacency relationships. It uses the Node2Vec algorithm to perform node embedding calculations on the spatial adjacency graph and performs hierarchical pooling operations on the embedding results through the DiffPool model to generate spatial feature sequences.

[0089] The strategy generation module is used to align the health status sequence and the spatial feature sequence by time step, use the A3C algorithm to perform prevention and control strategy generation and value assessment on the alignment result, and select target prevention and control vectors according to the assessment results and arrange them by time step to form the target prevention and control sequence.

[0090] The feedback update module is used to execute the target prevention sequence, collect feedback data, and update the parameters of the A3C algorithm based on the feedback data.

[0091] In this embodiment, the image data represents a set of digital images containing visual information of the tree crown, trunk, and leaves collected at each time step; the environmental monitoring data represents a set of numerical data consisting of temperature, humidity, light intensity, and soil moisture content; the preprocessing includes performing resolution unification, noise reduction, and time alignment operations on the image data, and performing missing value imputation, dimension unification, and time alignment operations on the environmental monitoring data; the spatial coordinates represent the three-dimensional position vector of the target ancient and famous trees in a unified spatial reference system; the spatial adjacency relationship represents a set of connection relationships between ancient trees constructed according to a preset spatial distance threshold; and the pest and disease identification sequence includes pest and disease category identifiers and pest and disease severity identifiers.

[0092] In this embodiment, the feature extraction module includes:

[0093] Image data from each time step in the multi-source monitoring sequence are cropped to a preset size and input into the DeepLabV3+ model. Semantic feature maps under different receptive fields are extracted from the dilated convolutional structure of the DeepLabV3+ model. Dilated spatial pyramid pooling is performed on the semantic feature maps. Feature concatenation and convolution mapping are performed on the pooling results at each scale to obtain the initial feature map.

[0094] The proportion of lesion pixels in the initial feature map is calculated based on the preset lesion recognition rules. The initial feature map is then weighted according to the proportion of lesion pixels to generate a semantic feature map.

[0095] The semantic feature map is input into the fully connected mapping structure of the DeepLabV3+ model, and vector expansion and linear mapping operations are performed on the semantic feature map to generate image feature vectors.

[0096] The environmental monitoring data in the multi-source monitoring sequence is read, and the four types of environmental components at each time step are arranged in a preset order. The weight mapping operation is performed based on the change amplitude of each environmental component between adjacent time steps to generate an environmental feature vector with the same dimension as the image feature vector. The environmental components represent temperature, humidity, light intensity and soil moisture content.

[0097] In terms of feature dimension, the image feature vector and the environment feature vector are concatenated. The image feature vector is used as the backbone feature, and a linear mapping operation is performed on the environment feature vector. The environment feature vector is then added element-wise with the backbone feature to form the residual feature vector.

[0098] Batch normalization and ReLU activation are performed on the residual feature vectors, and the vectors are arranged according to time steps to form a fused feature sequence.

[0099] In this embodiment, the semantic feature map generation process specifically includes:

[0100] Image data at each time step is input into the dilated convolutional structure of the DeepLabV3+ model. Convolutional calculations are performed on the image data in multiple convolutional branches at different scales. In each convolutional branch, the convolutional kernel slides along the spatial dimension of the image data with a preset stride. The pixel values ​​within the coverage area of ​​the convolutional kernel are multiplied element-wise with the convolutional kernel weight values, and the results are accumulated to obtain the convolutional feature map at the corresponding scale. The convolutional kernel weight values ​​are taken from the convolutional layer network parameters in the DeepLabV3+ model that match the corresponding convolutional branch.

[0101] A hollow spatial pyramid pooling operation is performed on each convolutional feature map. The pooling window size and pooling stride are set in multiple pooling branches of different scales. The pooling window slides along the spatial dimension of the convolutional feature map according to the pooling stride. The summation operation is performed on all feature values ​​within the pooling window coverage area and divided by the number of feature values ​​in the pooling window to obtain the spatial average value.

[0102] All spatial average values ​​are arranged in the sliding order of the pooling window to form pooling feature maps of the corresponding scales. Then, the pooling feature maps of each scale are spliced ​​together in the channel dimension to obtain the fused feature map.

[0103] In the convolutional mapping structure of the DeepLabV3+ model, a convolutional mapping operation is performed on the fused feature map. The preset convolutional kernel slides in the spatial dimension of the fused feature map with a preset stride. The feature value of each channel within the coverage area of ​​the convolutional kernel is multiplied element by element with the weight value of the corresponding channel, and then accumulated in the channel dimension to obtain the initial feature map. The weight value is taken from the convolutional layer network parameters corresponding to the convolutional mapping structure in the DeepLabV3+ model.

[0104] Based on the preset lesion recognition rules, the feature channel corresponding to the lesion category is determined in the channel dimension of the initial feature map. Each feature value of the feature channel in the spatial dimension is compared with the feature values ​​of all channels at the same spatial location. When the feature value of the feature channel is the largest feature value among all channels and is not less than the preset threshold, the spatial location is marked as the lesion pixel location.

[0105] Traverse all spatial locations, summarize all lesion pixel locations to form a lesion pixel set, count the number of pixels in the lesion pixel set and divide it by the total number of pixels in the initial feature map to obtain the lesion pixel ratio;

[0106] The initial feature map is scaled according to the proportion of lesion pixels. The product of the preset gain coefficient and the proportion of lesion pixels is incremented by one to obtain the scaling coefficient. Each feature value of each channel in the initial feature map is multiplied by the scaling coefficient to obtain the semantic feature map.

[0107] In this embodiment, the mapping recognition module includes:

[0108] The fused feature vectors at each time step in the fused feature sequence are reconstructed into a two-dimensional feature matrix in the feature dimension, and a discrete Fourier transform operation is performed on the two-dimensional feature matrix in the spatial dimension to map the two-dimensional feature matrix from the spatial domain to the frequency domain, thus obtaining the initial frequency domain matrix.

[0109] In the initial frequency domain matrix, each frequency component is divided into a high-frequency region and a low-frequency region according to a preset frequency division rule. The frequency components in the high-frequency region are subjected to amplitude scaling processing, while the amplitude of the frequency components in the low-frequency region remains unchanged, forming a frequency domain feature matrix. The frequency domain components represent the complex values ​​corresponding to each frequency position in the initial frequency domain matrix.

[0110] Performing an inverse discrete Fourier transform on the frequency domain feature matrix maps the frequency domain feature matrix back to the spatial domain, resulting in a mapped feature matrix.

[0111] A graph structure is constructed based on spatial adjacency relationships. The mapping feature matrix of each target ancient and famous tree is used as the initial node feature of the node in the graph structure, and the spatial adjacency relationship is used as the edge relationship in the graph structure.

[0112] The graph structure is input into the GATv2 network. Attention weights between nodes are calculated on each edge of the graph structure. The features of adjacent nodes are weighted and summed according to the attention weights in the node dimension. The weighted sum is then subjected to linear mapping and LeakyReLU activation to obtain the updated node features.

[0113] A fully connected mapping operation is performed on the updated node features along the channel dimension. The mapping results are divided into pest and disease category components and pest and disease severity components according to the preset dimension segmentation position, and arranged according to the time step to form a pest and disease identification sequence.

[0114] The disease and pest identification sequence and the fused feature sequence are spliced ​​together along the feature dimension, and the splicing result is subjected to linear mapping and batch normalization to generate a health status sequence.

[0115] In this embodiment, the amplitude scaling process specifically includes:

[0116] In the initial frequency domain matrix, locate the set of frequency indices for the high-frequency region according to the preset frequency division rules, and traverse each frequency index position in the set of frequency indices to read the corresponding complex value.

[0117] For each complex value, the summation operation is performed on the square of the real part and the square of the imaginary part, and then the square root operation is performed to calculate the amplitude of the complex value. All amplitudes are arranged according to frequency position to form a high-frequency amplitude matrix.

[0118] The high-frequency amplitude matrix is ​​probabilistically processed by calculating the sum of all amplitudes in the high-frequency amplitude matrix and dividing the amplitude at each frequency position by the sum to obtain the probability value, thus obtaining the high-frequency probability matrix.

[0119] In the high-frequency probability matrix, each frequency position is traversed with a preset neighborhood window. The set of probability values ​​in the neighborhood window corresponding to the frequency position is extracted. Logarithmic and multiplication operations are performed on each item of the probability value set. The multiplication results are accumulated and then the accumulated result is negative to obtain the local spectral entropy value of the frequency position. All local spectral entropy values ​​are arranged according to frequency position to form a spectral entropy matrix.

[0120] A threshold segmentation mapping operation is performed on the spectral entropy matrix. Based on the preset spectral entropy threshold, the spectral entropy matrix is ​​divided into enhancement regions and preservation regions. A corresponding weight value is generated for each frequency position in the enhancement region, and a weight value of one is generated for each frequency position in the preservation region. The weights are arranged according to the frequency positions to form a weight matrix.

[0121] The weight matrix is ​​multiplied by the amplitude at each frequency position while keeping the phase of the complex value at that frequency position unchanged, to obtain the enhanced frequency component.

[0122] In this embodiment, the process of generating the weight values ​​for each frequency position in the enhancement region specifically includes:

[0123] The maximum and minimum spectral entropy values ​​of the spectral entropy matrix within the enhanced region are statistically analyzed and used as the upper and lower bounds of the spectral entropy, respectively.

[0124] Preset upper and lower bound values ​​for the weights;

[0125] For each frequency location within the enhancement region, the local spectral entropy value at that frequency location is read. The lower bound of the spectral entropy is subtracted from the local spectral entropy value, and then divided by the difference between the upper bound and the lower bound of the spectral entropy to obtain the normalized spectral entropy value.

[0126] Multiply the normalized spectral entropy value by the difference between the upper and lower bounds of the weight, and then add the lower bound of the weight to obtain the weight value for that frequency position.

[0127] In this embodiment, the spatial calculation module includes:

[0128] Read the spatial coordinates and spatial adjacency relationships of the target ancient and famous trees, determine the sequence of adjacent node pairs based on the spatial adjacency relationships, calculate the spatial coordinate difference for each adjacent node pair as the spatial distance value, and arrange all spatial distance values ​​in the order of the adjacent node pair sequence to form a spatial distance sequence.

[0129] A spatial adjacency graph is constructed based on spatial adjacency relationships. Each target ancient and famous tree is defined as a node in the spatial adjacency graph, and spatial adjacency relationships are defined as edges in the spatial adjacency graph. The spatial distance sequence is written into the edge attribute set in edge order to generate a spatial adjacency graph with edge attribute set.

[0130] The Node2Vec algorithm is used to perform random walk operations in the spatial adjacency graph. Based on the spatial adjacency relationship, multiple fixed-length walk sequences are generated for each node, and all walk sequences are arranged by node index to form a walk sequence set.

[0131] Skip-Gram embedding training is performed on the wander sequence set. The center node and context node are extracted from the wander sequence set according to a fixed window length to form a node pair sample set. Vector mapping and similarity constraint calculation are performed on the node pair sample set to obtain the node embedding vector corresponding to each node. All node embedding vectors are arranged according to the node index to form a node embedding sequence.

[0132] The node embedding sequence is input into the DiffPool model, and clustering allocation calculation is performed to generate an allocation matrix. The allocation matrix is ​​normalized to construct hierarchical aggregation relationships. Based on the hierarchical aggregation relationships, hierarchical pooling operations are performed on the node embedding sequence to obtain a multi-level pooled embedding set.

[0133] The pooled embedding sets are concatenated in hierarchical order, and linear mapping and batch normalization operations are performed on the concatenation results to generate spatial feature sequences.

[0134] In this embodiment, the random walk operation specifically includes:

[0135] In the spatial adjacency graph, each node is selected sequentially according to the node index order. For each node, the corresponding walk sequence is initialized. The node is written into the corresponding walk sequence as the starting node and the current node. At the same time, the node before the starting node is marked as null.

[0136] Read all adjacent nodes of the current node in the spatial adjacency graph, find the spatial distance value between the current node and each adjacent node in the edge attribute set, perform the reciprocal calculation on the spatial distance value to obtain the initial transition value, perform a summation operation on all initial transition values, and divide each initial transition value by the summation result to obtain the transition probability set;

[0137] When the current node is null, select the target node from the adjacent nodes according to the transition probability set;

[0138] When the previous node is not null, for each adjacent node, determine whether the adjacent node has an edge connection with the previous node in the spatial adjacency graph. If there is an edge connection, keep the transition probability value unchanged. If there is no edge connection, multiply the corresponding transition probability value by a preset attenuation coefficient, and perform summation and normalization operations on the product result to obtain a new set of transition probabilities. Then, continue to select target nodes according to the new set of transition probabilities.

[0139] Write the target node into the current traversal sequence, update the previous node to the current node, and update the current node to the target node;

[0140] Determine whether the current walking sequence length has reached the preset walking length. If it has not reached the preset walking length, repeat the transition probability calculation, node selection and node writing operations. If it has reached the preset walking length, end the generation of the current walking sequence.

[0141] For each node in the spatial adjacency graph, the walk sequence generation operation is repeated to obtain multiple walk sequences. All walk sequences are then arranged in the order of node indices to form a walk sequence set.

[0142] In this embodiment, Skip-Gram embedding training specifically includes:

[0143] Traverse the set of walk sequences according to the node index order, extract the center node and context node in each walk sequence according to the preset window length to form a node pair sample set, and arrange all node pair sample sets in the order of appearance.

[0144] Initialize the corresponding node embedding vector for each node, and form an embedding matrix by combining all node embedding vectors in the order of node index.

[0145] For each node pair in the node pair sample set, read the node embedding vector corresponding to the center node and the node embedding vector corresponding to the context node, and perform a dot product calculation on the two node embedding vectors to obtain the similarity value.

[0146] Perform an exponential operation on the similarity value, and sum the exponential results of all context nodes corresponding to the same central node. Divide each exponential result by the sum to obtain the conditional probability value.

[0147] The objective function is constructed based on the conditional probability value, the gradient of the objective function is calculated, and the node embedding vectors of the center node and the context node are numerically updated according to the gradient result.

[0148] Repeatedly perform node traversal of the sample set and node embedding vector update operations until the preset training rounds are reached, generating a node embedding sequence.

[0149] In this embodiment, the calculation process of the DiffPool model specifically includes:

[0150] The node embedding sequence is reconstructed into a node feature matrix according to the node index order. At the same time, the adjacency relationship of the spatial adjacency graph is represented as an adjacency matrix, and symmetry and normalization operations are performed on the adjacency matrix to obtain a normalized adjacency matrix.

[0151] Perform a linear mapping operation on the node feature matrix and perform matrix multiplication with the normalized adjacency matrix to obtain the clustering score matrix;

[0152] Perform exponential and normalization operations on the clustering score matrix along the clustering dimension to obtain the assignment matrix;

[0153] Perform matrix multiplication between the transpose of the allocation matrix and the node feature matrix to obtain the pooling feature matrix;

[0154] After performing matrix multiplication on the transpose of the allocation matrix and the normalized adjacency matrix, perform matrix multiplication again on the allocation matrix to obtain the pooled adjacency matrix.

[0155] The pooling feature matrix and pooling adjacency matrix are used as inputs. The allocation matrix calculation and pooling operation are performed iteratively until the preset number of iterations is reached, generating a multi-level pooling embedding set.

[0156] In this embodiment, the strategy generation module includes:

[0157] Perform a time alignment operation on the health state sequence and the spatial feature sequence, and concatenate the aligned health state vector and the spatial feature vector along the feature dimension to generate an aligned feature sequence;

[0158] The aligned feature sequence is input into the policy network of the A3C algorithm step by step. Based on the policy weight matrix and bias vector in the policy network, matrix multiplication and bias addition are performed layer by layer on each aligned feature vector, and ReLU function activation is performed to obtain the action mapping vector.

[0159] In the action dimension, exponential calculation and normalization operations are performed on the action mapping vector to generate the prevention and control probability sequence at each time step;

[0160] Based on the prevention and control probability sequence, at each time step, the corresponding prevention and control action vector is selected from the preset prevention and control action set as a candidate prevention and control set;

[0161] The aligned feature sequence is input into the value network of the A3C algorithm, and matrix multiplication and bias addition are performed on the aligned feature vectors at each time step according to the value weight matrix and bias vector in the value network to generate the state value sequence.

[0162] In the time dimension, the difference between adjacent time steps is calculated on the state value sequence to obtain the time difference sequence. The time difference sequence is matched with the candidate prevention and control set in the time step dimension. The time difference is used as the screening criterion. In the same time step, the prevention and control action vector with the positive time difference and the largest value is taken as the target prevention and control vector.

[0163] All target prevention and control vectors are arranged in chronological order to form a target prevention and control sequence.

[0164] In this embodiment, the feedback update module includes:

[0165] Read the feedback data after the target prevention and control sequence is executed, align the feedback data with the health status sequence according to the time step, and calculate the status change value at each time step to form a status change sequence;

[0166] By mapping the state change sequence to the state value sequence in the time dimension, a reward sequence is generated based on the corresponding difference.

[0167] The parameters of the A3C algorithm are updated based on the reward sequence.

[0168] Example 1: To verify the feasibility of this invention in practice, it was applied to the daily maintenance of a concentrated area of ​​ancient and famous trees. This area contains various ancient and famous trees over 100 years old, with complex tree species structures and a network-like spatial distribution. These trees are significantly affected by climate fluctuations and soil moisture changes, and have long relied on manual inspections for pest and disease identification and control decisions. Traditional methods suffer from problems such as long inspection cycles, strong subjectivity in identification, difficulty in judging spatial spread trends, and difficulty in dynamically adjusting control measures. This results in some diseases not being detected in their early stages, leading to a year-on-year increase in control costs, and frequent occurrences of repeated spraying and over-treatment.

[0169] In this scenario, image acquisition and environmental monitoring terminals are first deployed around each ancient tree to continuously collect image data of the canopy, trunk, and leaves, as well as temperature, humidity, light intensity, and soil moisture content. All data are aligned at a unified time step to construct a multi-source monitoring sequence. Image data is input into a DeepLabV3+ model for semantic segmentation. Multi-scale features are extracted using dilated convolution and dilated spatial pyramid pooling structures, and weighted according to the proportion of lesion pixels to obtain a semantic feature map. This map is then mapped using a fully connected layer to form an image feature vector. Environmental monitoring data is used to construct an environmental feature vector through linear mapping, and this vector is concatenated with the image feature vector to generate a fused feature sequence.

[0170] Subsequently, the fused feature sequence is input into a Fourier neural operator for frequency domain mapping. Spectral entropy weighting enhancement is performed on the high-frequency region, and then an inverse transform is used to restore it to the spatial domain, yielding the mapped feature matrix. A graph structure is constructed based on spatial adjacency relationships, and node features are updated using a GATv2 network, outputting pest and disease identification sequences and health status sequences. Simultaneously, a spatial adjacency graph is constructed based on spatial coordinates, and random walk embedding is performed using the Node2Vec algorithm, followed by hierarchical pooling using the DiffPool model to form spatial feature sequences. The health status sequence and spatial feature sequences are time-aligned and input into the A3C algorithm for policy generation, outputting a target control sequence, including parameters such as whether to apply pesticides, pesticide dosage, and control frequency. Feedback data is collected after control implementation, and the parameters of the policy network and value network are updated based on changes in pests and diseases, forming a continuous optimization mechanism.

[0171] After a period of continuous operation, statistical analysis was conducted on indicators such as system identification accuracy, prevention and control success rate, resource utilization efficiency, and spatial propagation prediction accuracy. These results were compared with traditional manual inspection methods and single image recognition methods, yielding the following data:

[0172] Table 1. Comparative Statistics of the Effects of Intelligent Identification and Control of Diseases and Pests on Ancient and Famous Trees

[0173]

[0174] As shown in Table 1, the present invention achieves a pest and disease identification accuracy of 95.4%, which is 16.8 percentage points higher than traditional manual inspection and 9.1 percentage points higher than a single image recognition model. This indicates that integrating multi-source monitoring sequences and frequency domain enhancement processing can effectively improve identification capabilities. The early-stage disease identification rate reaches 91.7%, demonstrating that the enhanced lesion pixel ratio and high spectral entropy adjustment mechanism have a significant enhancement effect on early, subtle lesions. The spatial propagation prediction accuracy reaches 92.3%, nearly 37 percentage points higher than traditional methods, proving that spatial adjacency graph and hierarchical embedding modeling can effectively characterize the spread trend of pests and diseases.

[0175] In terms of control effectiveness, the average control response time was shortened to 8 hours, the control success rate reached 93.8%, the incidence of repeated application of pesticides decreased to 3.6%, and the average annual control cost decreased to 860 yuan. This indicates that the reinforcement learning decision structure constructed by the A3C algorithm can dynamically adjust the control strategy according to the health status sequence and spatial feature sequence, and converge quickly under the feedback update mechanism. The recurrence rate of pests and diseases decreased to 6.9%, and the comprehensive resource utilization efficiency increased by 34.8%, indicating that this invention achieves closed-loop collaboration of identification, spatial modeling, and strategy optimization.

[0176] As can be seen from the above embodiments, the present invention can effectively solve the problems of lagging identification, unpredictable spatial propagation, and inability to dynamically optimize prevention and control strategies in actual ancient and famous tree maintenance scenarios, and achieves a comprehensive effect of high-precision identification, high-efficiency decision-making, and continuous optimization of prevention and control.

[0177] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A system for intelligent identification and control of diseases and pests in ancient and famous trees based on reinforcement learning, characterized in that, include: The data acquisition module is used to collect and preprocess image data and environmental monitoring data of the target ancient and famous trees at continuous time steps, generate multi-source monitoring sequences, and collect the spatial coordinates and spatial adjacency relationships of the target ancient and famous trees. The feature extraction module is used to perform semantic segmentation and fully connected mapping operations on multi-source monitoring sequences using the DeepLabV3+ model, extract image features and environmental features, and perform fusion processing to generate fused feature sequences. The mapping and recognition module is used to perform frequency domain mapping operation on the fused feature sequence using Fourier neural operators, construct a graph structure based on the mapping result and spatial adjacency relationship, update the node features of the graph structure through the GATv2 network, generate a pest and disease recognition sequence and connect it with the fused feature sequence to obtain a health status sequence. The spatial computing module is used to construct a spatial adjacency graph based on spatial coordinates and spatial adjacency relationships. It uses the Node2Vec algorithm to perform node embedding calculations on the spatial adjacency graph and performs hierarchical pooling operations on the embedding results through the DiffPool model to generate spatial feature sequences. The strategy generation module is used to align the health status sequence and the spatial feature sequence by time step, use the A3C algorithm to perform prevention and control strategy generation and value assessment on the alignment result, and select target prevention and control vectors according to the assessment results and arrange them by time step to form the target prevention and control sequence. The feedback update module is used to execute the target prevention sequence, collect feedback data, and update the parameters of the A3C algorithm based on the feedback data.

2. The intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning according to claim 1, characterized in that, The image data represents a set of digital images containing visual information of the tree crown, trunk, and leaves collected at each time step. The environmental monitoring data represents a set of numerical data consisting of temperature, humidity, light intensity, and soil moisture content. The preprocessing includes performing resolution unification, noise reduction, and time alignment operations on the image data, and performing missing value imputation, dimension unification, and time alignment operations on the environmental monitoring data. The spatial coordinates represent the three-dimensional position vectors of the target ancient and famous trees in a unified spatial reference system. The spatial adjacency relationship represents a set of connection relationships between ancient trees constructed based on a preset spatial distance threshold. The pest and disease identification sequence includes pest and disease category identifiers and pest and disease severity identifiers.

3. The intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning according to claim 1, characterized in that, The feature extraction module includes: Image data from each time step in the multi-source monitoring sequence are cropped to a preset size and input into the DeepLabV3+ model. Semantic feature maps under different receptive fields are extracted from the dilated convolutional structure of the DeepLabV3+ model. Dilated spatial pyramid pooling is performed on the semantic feature maps. Feature concatenation and convolution mapping are performed on the pooling results at each scale to obtain the initial feature map. The proportion of lesion pixels in the initial feature map is calculated based on the preset lesion recognition rules. The initial feature map is then weighted according to the proportion of lesion pixels to generate a semantic feature map. The semantic feature map is input into the fully connected mapping structure of the DeepLabV3+ model, and vector expansion and linear mapping operations are performed on the semantic feature map to generate image feature vectors. The environmental monitoring data in the multi-source monitoring sequence is read, and the four types of environmental components at each time step are arranged in a preset order. The weight mapping operation is performed based on the change amplitude of each environmental component between adjacent time steps to generate an environmental feature vector with the same dimension as the image feature vector. The environmental components represent temperature, humidity, light intensity and soil moisture content. In terms of feature dimension, the image feature vector and the environment feature vector are concatenated. The image feature vector is used as the backbone feature, and a linear mapping operation is performed on the environment feature vector. The environment feature vector is then added element-wise with the backbone feature to form the residual feature vector. Batch normalization and ReLU activation are performed on the residual feature vectors, and the vectors are arranged according to time steps to form a fused feature sequence.

4. The intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning according to claim 3, characterized in that, The process of generating the semantic feature map specifically includes: Image data at each time step is input into the dilated convolutional structure of the DeepLabV3+ model. Convolutional calculations are performed on the image data in multiple convolutional branches at different scales. In each convolutional branch, the convolutional kernel slides along the spatial dimension of the image data with a preset stride. The pixel values ​​within the coverage area of ​​the convolutional kernel are multiplied element-wise with the convolutional kernel weight values, and the results are accumulated to obtain the convolutional feature map at the corresponding scale. The convolutional kernel weight values ​​are taken from the convolutional layer network parameters in the DeepLabV3+ model that match the corresponding convolutional branch. A hollow spatial pyramid pooling operation is performed on each convolutional feature map. The pooling window size and pooling stride are set in multiple pooling branches of different scales. The pooling window slides along the spatial dimension of the convolutional feature map according to the pooling stride. The summation operation is performed on all feature values ​​within the pooling window coverage area and divided by the number of feature values ​​in the pooling window to obtain the spatial average value. All spatial average values ​​are arranged in the sliding order of the pooling window to form pooling feature maps of the corresponding scales. Then, the pooling feature maps of each scale are spliced ​​together in the channel dimension to obtain the fused feature map. In the convolutional mapping structure of the DeepLabV3+ model, a convolutional mapping operation is performed on the fused feature map. The preset convolutional kernel slides in the spatial dimension of the fused feature map with a preset stride. The feature value of each channel within the coverage area of ​​the convolutional kernel is multiplied element by element with the weight value of the corresponding channel, and then accumulated in the channel dimension to obtain the initial feature map. The weight value is taken from the convolutional layer network parameters corresponding to the convolutional mapping structure in the DeepLabV3+ model. Based on the preset lesion recognition rules, the feature channel corresponding to the lesion category is determined in the channel dimension of the initial feature map. Each feature value of the feature channel in the spatial dimension is compared with the feature values ​​of all channels at the same spatial location. When the feature value of the feature channel is the largest feature value among all channels and is not less than the preset threshold, the spatial location is marked as the lesion pixel location. Traverse all spatial locations, summarize all lesion pixel locations to form a lesion pixel set, count the number of pixels in the lesion pixel set and divide it by the total number of pixels in the initial feature map to obtain the lesion pixel ratio; The initial feature map is scaled according to the proportion of lesion pixels. The product of the preset gain coefficient and the proportion of lesion pixels is incremented by one to obtain the scaling coefficient. Each feature value of each channel in the initial feature map is multiplied by the scaling coefficient to obtain the semantic feature map.

5. The intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning according to claim 1, characterized in that, The mapping recognition module includes: The fused feature vectors at each time step in the fused feature sequence are reconstructed into a two-dimensional feature matrix in the feature dimension, and a discrete Fourier transform operation is performed on the two-dimensional feature matrix in the spatial dimension to map the two-dimensional feature matrix from the spatial domain to the frequency domain, thus obtaining the initial frequency domain matrix. In the initial frequency domain matrix, each frequency component is divided into a high-frequency region and a low-frequency region according to a preset frequency division rule. The frequency components in the high-frequency region are subjected to amplitude scaling processing, while the amplitude of the frequency components in the low-frequency region remains unchanged, forming a frequency domain feature matrix. The frequency domain components represent the complex values ​​corresponding to each frequency position in the initial frequency domain matrix. Performing an inverse discrete Fourier transform on the frequency domain feature matrix maps the frequency domain feature matrix back to the spatial domain, resulting in a mapped feature matrix. A graph structure is constructed based on spatial adjacency relationships. The mapping feature matrix of each target ancient and famous tree is used as the initial node feature of the node in the graph structure, and the spatial adjacency relationship is used as the edge relationship in the graph structure. The graph structure is input into the GATv2 network. Attention weights between nodes are calculated on each edge of the graph structure. The features of adjacent nodes are weighted and summed according to the attention weights in the node dimension. The weighted sum is then subjected to linear mapping and LeakyReLU activation to obtain the updated node features. A fully connected mapping operation is performed on the updated node features along the channel dimension. The mapping results are divided into pest and disease category components and pest and disease severity components according to the preset dimension segmentation position, and arranged according to the time step to form a pest and disease identification sequence. The disease and pest identification sequence and the fused feature sequence are spliced ​​together along the feature dimension, and the splicing result is subjected to linear mapping and batch normalization to generate a health status sequence.

6. The intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning according to claim 5, characterized in that, The amplitude scaling process specifically includes: In the initial frequency domain matrix, locate the set of frequency indices for the high-frequency region according to the preset frequency division rules, and traverse each frequency index position in the set of frequency indices to read the corresponding complex values. For each complex value, the summation operation is performed on the square of the real part and the square of the imaginary part, and then the square root operation is performed to calculate the amplitude of the complex value. All amplitudes are arranged according to frequency position to form a high-frequency amplitude matrix. The high-frequency amplitude matrix is ​​probabilistically processed by calculating the sum of all amplitudes in the high-frequency amplitude matrix and dividing the amplitude at each frequency position by the sum to obtain the probability value, thus obtaining the high-frequency probability matrix. In the high-frequency probability matrix, each frequency position is traversed with a preset neighborhood window. The set of probability values ​​in the neighborhood window corresponding to the frequency position is extracted. Logarithmic and multiplication operations are performed on each item of the probability value set. The multiplication results are accumulated and then the accumulated result is negative to obtain the local spectral entropy value of the frequency position. All local spectral entropy values ​​are arranged according to frequency position to form a spectral entropy matrix. A threshold segmentation mapping operation is performed on the spectral entropy matrix. Based on the preset spectral entropy threshold, the spectral entropy matrix is ​​divided into enhancement regions and preservation regions. A corresponding weight value is generated for each frequency position in the enhancement region, and a weight value of one is generated for each frequency position in the preservation region. The weights are arranged according to the frequency positions to form a weight matrix. The weight matrix is ​​multiplied by the amplitude at each frequency position while keeping the phase of the complex value at that frequency position unchanged, to obtain the enhanced frequency component.

7. The intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning according to claim 1, characterized in that, The spatial computing module includes: Read the spatial coordinates and spatial adjacency relationships of the target ancient and famous trees, determine the sequence of adjacent node pairs based on the spatial adjacency relationships, calculate the spatial coordinate difference for each adjacent node pair as the spatial distance value, and arrange all spatial distance values ​​in the order of the adjacent node pair sequence to form a spatial distance sequence. A spatial adjacency graph is constructed based on spatial adjacency relationships. Each target ancient and famous tree is defined as a node in the spatial adjacency graph, and spatial adjacency relationships are defined as edges in the spatial adjacency graph. The spatial distance sequence is written into the edge attribute set in edge order to generate a spatial adjacency graph with edge attribute set. The Node2Vec algorithm is used to perform random walk operations in the spatial adjacency graph. Based on the spatial adjacency relationship, multiple fixed-length walk sequences are generated for each node, and all walk sequences are arranged by node index to form a walk sequence set. Skip-Gram embedding training is performed on the wander sequence set. The center node and context node are extracted from the wander sequence set according to a fixed window length to form a node pair sample set. Vector mapping and similarity constraint calculation are performed on the node pair sample set to obtain the node embedding vector corresponding to each node. All node embedding vectors are arranged according to the node index to form a node embedding sequence. The node embedding sequence is input into the DiffPool model, and clustering allocation calculation is performed to generate an allocation matrix. The allocation matrix is ​​normalized to construct hierarchical aggregation relationships. Based on the hierarchical aggregation relationships, hierarchical pooling operations are performed on the node embedding sequence to obtain a multi-level pooled embedding set. The pooled embedding sets are concatenated in hierarchical order, and linear mapping and batch normalization operations are performed on the concatenation results to generate spatial feature sequences.

8. The intelligent identification and control system for diseases and pests of ancient and famous trees based on reinforcement learning according to claim 7, characterized in that, The random walk operation specifically includes: In the spatial adjacency graph, each node is selected sequentially according to the node index order. For each node, the corresponding walk sequence is initialized. The node is written into the corresponding walk sequence as the starting node and the current node. At the same time, the node before the starting node is marked as null. Read all adjacent nodes of the current node in the spatial adjacency graph, find the spatial distance value between the current node and each adjacent node in the edge attribute set, perform the reciprocal calculation on the spatial distance value to obtain the initial transition value, perform a summation operation on all initial transition values, and divide each initial transition value by the summation result to obtain the transition probability set; When the current node is null, select the target node from the adjacent nodes according to the transition probability set; When the previous node is not null, for each adjacent node, determine whether the adjacent node has an edge connection with the previous node in the spatial adjacency graph. If there is an edge connection, keep the transition probability value unchanged. If there is no edge connection, multiply the corresponding transition probability value by a preset attenuation coefficient, and perform summation and normalization operations on the product result to obtain a new set of transition probabilities. Then, continue to select target nodes according to the new set of transition probabilities. Write the target node into the current traversal sequence, update the previous node to the current node, and update the current node to the target node; Determine whether the current walking sequence length has reached the preset walking length. If it has not reached the preset walking length, repeat the transition probability calculation, node selection and node writing operations. If it has reached the preset walking length, end the generation of the current walking sequence. For each node in the spatial adjacency graph, the walk sequence generation operation is repeated to obtain multiple walk sequences. All walk sequences are then arranged in the order of node indices to form a walk sequence set.

9. A reinforcement learning-based intelligent identification and control system for diseases and pests of ancient and famous trees, as described in claim 7, is characterized in that... The calculation process of the DiffPool model specifically includes: The node embedding sequence is reconstructed into a node feature matrix according to the node index order. At the same time, the adjacency relationship of the spatial adjacency graph is represented as an adjacency matrix, and symmetry and normalization operations are performed on the adjacency matrix to obtain a normalized adjacency matrix. Perform a linear mapping operation on the node feature matrix and perform matrix multiplication with the normalized adjacency matrix to obtain the clustering score matrix; Perform exponential and normalization operations on the clustering score matrix along the clustering dimension to obtain the assignment matrix; Perform matrix multiplication between the transpose of the allocation matrix and the node feature matrix to obtain the pooling feature matrix; After performing matrix multiplication on the transpose of the allocation matrix and the normalized adjacency matrix, perform matrix multiplication again on the allocation matrix to obtain the pooled adjacency matrix. The pooling feature matrix and pooling adjacency matrix are used as inputs. The allocation matrix calculation and pooling operation are performed iteratively until the preset number of iterations is reached, generating a multi-level pooling embedding set.

10. A reinforcement learning-based intelligent identification and control system for diseases and pests of ancient and famous trees according to claim 1, characterized in that, The strategy generation module includes: Perform a time alignment operation on the health state sequence and the spatial feature sequence, and concatenate the aligned health state vector and the spatial feature vector along the feature dimension to generate an aligned feature sequence; The aligned feature sequence is input into the policy network of the A3C algorithm step by step. Based on the policy weight matrix and bias vector in the policy network, matrix multiplication and bias addition are performed layer by layer on each aligned feature vector, and ReLU function activation is performed to obtain the action mapping vector. In the action dimension, exponential calculation and normalization operations are performed on the action mapping vector to generate the prevention and control probability sequence at each time step; Based on the prevention and control probability sequence, at each time step, the corresponding prevention and control action vector is selected from the preset prevention and control action set as a candidate prevention and control set; The aligned feature sequence is input into the value network of the A3C algorithm, and matrix multiplication and bias addition are performed on the aligned feature vectors at each time step according to the value weight matrix and bias vector in the value network to generate the state value sequence. In the time dimension, the difference between adjacent time steps is calculated on the state value sequence to obtain the time difference sequence. The time difference sequence is matched with the candidate prevention and control set in the time step dimension. The time difference is used as the screening criterion. In the same time step, the prevention and control action vector with the positive time difference and the largest value is taken as the target prevention and control vector. All target prevention and control vectors are arranged in chronological order to form a target prevention and control sequence.

Images