A cherry leaf spot disease recognition method and system based on a deep convolutional network

By constructing a cherry blossom leaf spot disease identification system based on deep convolutional networks, the shortcomings of existing cherry blossom leaf spot disease identification systems in terms of growth cycle perception and environmental adaptability are solved. This achieves efficient and stable disease monitoring and identification, reduces system maintenance costs, and has good scalability.

CN120783187BActive Publication Date: 2025-11-18江西省经济作物研究所
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511236860.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-09-01
Publication Date
2025-11-18
Estimated Expiration
2045-09-01

AI Technical Summary

Technical Problem

Existing cherry blossom leaf spot disease identification systems lack the ability to perceive the growth cycle and cannot effectively cope with the dynamic changes in leaf characteristics at different growth stages, resulting in difficulty in guaranteeing identification accuracy and stability. Furthermore, the system suffers from model performance degradation and poor environmental adaptability during long-term operation.

Method used

A cherry blossom leaf spot disease identification system based on deep convolutional networks is constructed. By establishing a dual dynamic system mathematical model of plant growth and model evolution, a feature extractor for growth stage perception, a progressive knowledge base, a dual memory system, a concept drift detector, and a meta-learning optimizer, the system achieves long-term collaboration with the cherry blossom growth cycle. The feature extractor adapts to the feature differences of different growth stages, the progressive knowledge base and dual memory system improve the identification accuracy and stability, the concept drift detector monitors the long-term trend of disease manifestation, and the meta-learning optimizer learns the optimal adaptation strategy.

Benefits of technology

It significantly improves the identification accuracy and stability in the whole life cycle monitoring, reduces the dependence on new samples, enhances the system's response to environmental changes and disease evolution, reduces system maintenance costs, and has good scalability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120783187B_ABST
    Figure CN120783187B_ABST
Patent Text Reader

Abstract

The application relates to the technical field of plant disease identification, and discloses a cherry leaf spot disease identification method and system based on a deep convolutional network, wherein the cherry leaf spot disease identification method based on the deep convolutional network comprises the following steps: establishing a double dynamic system mathematical model of plant growth and model evolution, constructing a growth stage perception feature extractor, a progressive knowledge base, a double memory system, a concept drift detector and a meta-learning optimizer, integrating a biological technology co-adaptation framework, and realizing long-term cooperation between a cherry leaf spot disease identification system and a cherry growth cycle; the application realizes long-term cooperation between the identification system and the cherry growth cycle, solves the contradiction between stability and adaptability of a traditional deep learning method in long-term plant monitoring, and enables the cherry leaf spot disease identification system to maintain high efficiency and stability in a multi-year application process without frequent manual intervention and model updating.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of plant disease identification technology, and more specifically, to a method and system for identifying cherry blossom leaf spot disease based on deep convolutional networks. Background Technology

[0002] With the continuous expansion of cherry blossom planting and the increasing ornamental value, the prevention and control of cherry blossom leaf spot disease has become an important issue in garden maintenance. Leaf spot disease not only affects the ornamental effect of cherry blossoms but also leads to weakened tree vigor, stunted growth, and in severe cases, the death of the entire tree. Currently, the identification of cherry blossom leaf spot disease mainly relies on manual inspection. This method is not only time-consuming and labor-intensive but also easily affected by personnel experience and subjective judgment, making it difficult to achieve large-scale, high-efficiency disease monitoring.

[0003] The development of deep learning technology has provided new technical means for plant disease identification. However, as a perennial woody plant, cherry blossoms exhibit significant seasonal and periodic growth characteristics, with leaf morphology, color, and texture changing markedly with each growth stage. This biological characteristic presents unique challenges to deep learning-based disease identification systems, particularly in terms of model long-term stability, environmental adaptability, and growth cycle perception capabilities.

[0004] Furthermore, the symptoms of cherry blossom leaf spot disease dynamically evolve with changes in environmental conditions, pathogen evolution, and host status. This conceptual drift makes it difficult for traditional static identification models to maintain long-term accuracy. At the same time, the complex variations in environmental factors such as light conditions, temperature, and humidity in cherry blossom planting areas also place higher demands on the robustness of disease identification systems. Summary of the Invention

[0005] This invention provides a method and system for identifying cherry blossom leaf spot disease based on deep convolutional networks. It solves the technical problems in related technologies, such as the lack of perception of the cherry blossom growth cycle, the inability to effectively cope with the dynamic changes in leaf characteristics at different growth stages, the difficulty in guaranteeing the accuracy and stability of identification, and the existence of model performance degradation and poor environmental adaptability during long-term system operation.

[0006] This invention provides a method for identifying cherry blossom leaf spot disease based on deep convolutional networks, comprising the following steps:

[0007] Establish a dual dynamic system mathematical model of plant growth and model evolution to uniformly represent the plant growth process and the model evolution process;

[0008] A growth stage-aware feature extractor is constructed. Through a multi-scale feature extraction network and growth stage conditional coding, the feature extractor can adapt to the differences in cherry blossom leaf features at different growth stages.

[0009] Construct a progressive knowledge base, which stores and manages feature prototypes and recognition rules from different periods through a hierarchical knowledge storage structure and knowledge update mechanism;

[0010] Construct a dual memory system that distinguishes between long-term stable knowledge and short-term adaptive knowledge through long-term memory and short-term memory subsystems;

[0011] A concept drift detector was constructed to monitor the long-term trends of disease manifestations and plant characteristics through multi-granularity feature distribution monitoring and multi-timescale change detection.

[0012] A meta-learning optimizer is constructed, which learns the optimal adaptation strategy from the historical adaptation process through meta-parameter space optimization and historical experience learning;

[0013] By integrating a biotechnology co-adaptation framework and through system architecture and biological clock synchronization mechanisms, a long-term synergy between the cherry blossom leaf spot disease identification system and the cherry blossom growth cycle can be achieved.

[0014] In a preferred embodiment, the mathematical model of the dual dynamic system of plant growth and model evolution includes:

[0015] Plant state dynamic evolution model: used to describe the changing pattern of plant state over time, and calculate the rate of change of plant state based on the current state of the plant and environmental conditions;

[0016] Identify the dynamic evolution model of model parameters: This model is used to describe the update pattern of model parameters over time and calculates the rate of change of model parameters based on the current model parameters, plant status, and collected data.

[0017] System coupling model: used to describe the acquisition process of plant status observation data and the dynamic update mechanism of environmental conditions. The update of environmental conditions takes into account the influence of model parameters and plant status.

[0018] In a preferred embodiment, the growth stage-aware feature extractor includes:

[0019] Multi-scale feature extraction network: includes a backbone network, a multi-scale feature extraction module, and a feature pyramid network, which extracts and fuses features from different scales;

[0020] Growth stage conditional encoding: This includes growth stage encoding, conditional batch normalization layer and feature map modulation, which encodes cherry blossom growth stage information and injects it into the feature extraction process;

[0021] Feature adaptation module: includes dynamic feature transformation matrix, dynamic convolutional kernel generation network and dual attention modulation module, which enables the feature extractor to adaptively adjust according to the growth stage;

[0022] Cross-stage feature mapping: This includes feature alignment loss function, encoder-decoder network for different growth stages, periodic consistency constraint and identity preservation mechanism, to establish the mapping relationship between feature spaces at different growth stages.

[0023] In a preferred embodiment, the progressive knowledge base includes:

[0024] Hierarchical knowledge storage structure: knowledge is organized into an instance layer, a prototype layer, a rule layer, and a meta-knowledge layer;

[0025] Knowledge caching and persistence management includes short-term caching, long-term storage, indexing systems, and compressed representation;

[0026] Knowledge update and verification mechanisms include incremental updates, conflict resolution, knowledge verification, and version control.

[0027] Knowledge management strategies related to biological cycles include periodic organization, seasonal activation, integration of historical experience, and predictive caching.

[0028] In a preferred embodiment, the dual-memory system includes:

[0029] Long-term memory subsystem: stores core features, invariant knowledge, long-term statistical models, and category prototypes;

[0030] Short-term memory subsystem: storage environment adaptation parameters, seasonality characteristics, recent sample caching, and temporary hypothesis model;

[0031] Memory interaction and fusion mechanisms include bidirectional information flow, knowledge distillation, memory reset gates, and long-term guidance;

[0032] Memory update strategies include stability assessment, importance weighting, forgetting mechanisms, and consolidation mechanisms.

[0033] In a preferred embodiment, the concept drift detector includes:

[0034] Multi-granularity feature distribution monitoring system: monitors the original feature distribution, semantic feature distribution, discrimination boundary, and prediction confidence;

[0035] Multi-timescale change detection: including short-term fluctuation detection, medium-term trend detection, long-term cycle detection, and non-periodic mutation detection;

[0036] Distribution offset quantification model: including statistical distance calculation, rate of change assessment, threshold adaptation, and confidence interval;

[0037] Drift response and early warning mechanism: including drift classification, severity assessment, early warning triggering, and update recommendations.

[0038] In a preferred embodiment, the meta-learning optimizer includes:

[0039] Meta-parameter space and optimization objectives: including meta-parameter definitions, multi-objective functions, constraints, and reward functions;

[0040] Data collection and representation in the meta-learning process: including empirical trajectories, contextual representations, performance metrics, and metadata annotation;

[0041] Meta-policy learning based on historical experience includes pattern extraction, conditional mapping, policy networks, and reinforcement learning.

[0042] Adaptive execution and evaluation mechanisms include exploratory balancing, metaparameter scheduling, A / B testing, and feedback loops.

[0043] In a preferred embodiment, the biotechnology co-adaptation framework includes:

[0044] System overall architecture: including multi-level design, modular structure, communication interfaces, and event-driven architecture;

[0045] Biological clock synchronization mechanisms include growth cycle modeling, model cycle synchronization, predictive adjustment, and periodic reset.

[0046] Adaptive reasoning engine: including multi-model fusion, context awareness, confidence assessment, and decision adjustment;

[0047] System self-reflection and optimization: including performance monitoring, error analysis, improvement suggestions, and adaptive adjustments.

[0048] In a preferred embodiment, a cherry blossom leaf spot disease identification system based on deep convolutional networks is used to execute a cherry blossom leaf spot disease identification method based on deep convolutional networks, including:

[0049] Image acquisition module: used to acquire images of cherry blossom leaves;

[0050] Data preprocessing module: used to standardize, enhance, and perform quality control on the acquired images;

[0051] Biotechnology Co-adaptation Framework: A comprehensive process for identifying cherry blossom leaf spot disease;

[0052] Results display module: used to present the recognition results in a visual way;

[0053] Early warning and intervention module: Used to generate early warning information and provide intervention suggestions based on the identification results.

[0054] In a preferred embodiment, a cherry blossom leaf spot disease identification system based on deep convolutional networks further includes:

[0055] Environmental sensing module: used to collect environmental parameters and provide environmental condition vectors for the model;

[0056] Edge computing unit: used for preprocessing and initial identification at the data acquisition site;

[0057] Cloud servers: used to perform complex calculations, store long-term knowledge bases, and coordinate multi-location systems;

[0058] Human-computer interface: used to receive expert feedback and manually adjust system parameters;

[0059] API interface: Used for integration with other agricultural management systems.

[0060] The beneficial effects of this invention are as follows:

[0061] By unifying the modeling of plant growth processes with the evolutionary process through a biotechnology co-adaptation framework, long-term co-adaptation between identification technology and biological systems has been achieved. A feature extractor based on growth stage perception can adaptively adjust the identification strategy according to the differences in cherry blossom leaf characteristics at different growth stages, significantly improving the identification accuracy and stability in full life-cycle monitoring.

[0062] The progressive knowledge base and dual-memory system enable the system to efficiently store and utilize historical data, reducing reliance on new samples. Simultaneously, the concept drift detector can promptly detect long-term trends in disease manifestations, providing a basis for model updates and improving the system's responsiveness to environmental changes and disease evolution.

[0063] The meta-learning optimizer learns the optimal adaptation strategy from historical adaptation processes, improving the efficiency and stability of model updates. The overall system design achieves common characteristics such as "balance between long-term memory and short-term adaptation" and "predictive adaptation capability," enabling it to maintain stable identification of core disease characteristics while adapting to seasonal and growth stage changes during long-term operation.

[0064] By replacing the traditional periodic retraining scheme with a unified co-adaptation framework, the system maintenance cost is reduced. The system has good scalability and can be transferred to other plant disease identification tasks with low adaptation cost, showing broad application prospects. Attached Figure Description

[0065] Figure 1 This is a flowchart of a cherry blossom leaf spot disease identification method based on deep convolutional networks according to the present invention;

[0066] Figure 2 This is a radar chart comparing the multi-dimensional performance of the system of this invention with that of traditional methods;

[0067] Figure 3 This is a bar chart comparing the update frequency of the model at different growth stages of the present invention;

[0068] Figure 4This is an area graph comparing the early disease detection capabilities of this invention;

[0069] Figure 5 This is a line graph comparing the accuracy of cherry blossom leaf spot disease identification at different growth stages according to the present invention.

[0070] Figure 6 This is a bar chart comparing the performance improvement of the method of this invention with that of traditional methods;

[0071] Figure 7 This is a line graph comparing the long-term operational stability of the system of this invention with that of traditional methods. Detailed Implementation

[0072] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0073] At least one embodiment of the present invention discloses a method for identifying cherry blossom leaf spot disease based on deep convolutional networks, such as... Figure 1 As shown, it includes the following steps:

[0074] Step 1: Establish a dual dynamic system mathematical model of plant growth and model evolution to uniformly represent the plant growth process and the model evolution process;

[0075] Specifically, the following steps are included:

[0076] Step 1.1: Construct a dynamic evolution model of plant states, representing the plant growth process as a differential equation of the plant state vector changing over time:

[0077] ;

[0078] in, This represents the rate of change of the plant state vector over time, i.e., the instantaneous rate of change of the plant state. This represents a plant state vector, which includes features such as leaf morphology, color, and texture. The environmental condition vector is a multi-dimensional vector. The plant state transition function describes the mathematical laws governing how the plant state dynamically changes with environmental conditions.

[0079] Step 1.2: Construct a dynamic evolution model of the identification model parameters, representing the model update process as a differential equation of how the model parameters change with plant state and data:

[0080] ;

[0081] in, It represents the rate of change of the model parameter vector over time, that is, the instantaneous rate of change of the model parameters; This represents the model parameter vector, which contains all trainable parameters of the deep learning model. This represents a plant state vector, which includes features such as leaf morphology, color, and texture. Indicates at time The collected training dataset; The parameter update function is a non-linear function that describes how the model parameters are adaptively updated based on the current state and data. It is usually implemented based on optimization algorithms such as gradient descent.

[0082] Step 1.3: Construct a coupled system model, define the observation function and feedback function, and establish the coupling relationship between the two dynamic systems:

[0083] ;

[0084] ;

[0085] in, This represents the training dataset collected at time t; Let be the observation function, representing the process of obtaining data from the plant state; This represents a plant state vector, which includes features such as leaf morphology, color, and texture. Observational noise represents random errors during the data acquisition process; This represents the environmental condition vector for the next time step. This represents the environmental condition vector at the current moment; The feedback function represents the process of intervening in the environment based on the model's recognition results; This represents the model parameter vector, which contains all trainable parameters of the deep learning model.

[0086] Step 1.4, Discretization: The continuous model is converted into discrete update rules to facilitate computer implementation.

[0087] Plant status update:

[0088] ;

[0089] in, This represents the plant state vector at the next moment; This represents the plant state vector at the current moment; The time step is determined based on the actual application scenario; The plant state transition function describes the mathematical laws governing how the plant state dynamically changes with environmental conditions. This represents the environmental condition vector at the current moment.

[0090] Model parameter update:

[0091] ;

[0092] in, Represents the model parameter vector at the next time step; This represents the model parameter vector at the current moment; Functions for updating model parameters; This represents the training dataset collected at the current moment; The time step is determined based on the actual application scenario; This represents the plant state vector at the current moment.

[0093] like Figure 2 The figure shows a comparison between the method of this invention and traditional methods in five key performance indicators, including recognition accuracy, early detection capability, environmental adaptability, long-term stability, and data efficiency. It is clearly evident from the figure that the method of this invention significantly outperforms the traditional method in all dimensions, with its advantage being most prominent in long-term stability.

[0094] Step 2: Construct a growth stage-aware feature extractor. Through a multi-scale feature extraction network and growth stage conditional coding, the feature extractor can adapt to the differences in cherry blossom leaf features at different growth stages.

[0095] Specifically, the following steps are included:

[0096] Step 2.1: Design the basic structure of the multi-scale feature extraction network, constructing a deep convolutional neural network using convolutional layers, pooling layers, and normalization layers:

[0097] Input layer: Receives RGB images of cherry blossom leaves, normalizes them to 224×224 pixels, and applies standard data preprocessing operations, including channel mean subtraction and standard deviation normalization.

[0098] Feature extraction backbone network: Pre-trained backbone networks such as ResNet-50 or DenseNet-121 are used as the basis for feature extraction.

[0099] In the cherry blossom leaf spot disease identification scenario, the ResNet-50 network can provide five feature maps of different depths: C1 (56×56×64), C2 (28×28×256), C3 (14×14×512), C4 (7×7×1024), and C5 (7×7×2048). These feature maps correspond to shallow texture features and deep semantic features, respectively.

[0100] Multi-scale feature extraction module: Extracts feature maps from multiple intermediate layers of the backbone network, specifically including:

[0101] Low-level features are extracted from the C2 layer, mainly capturing basic visual features such as leaf edges, textures, and color distribution.

[0102] Intermediate features are extracted from layer C3 to capture the structural patterns and local morphological features of the blade.

[0103] High-level features are extracted from layer C4 to identify semantic features and global patterns of diseases;

[0104] Extract the highest-level features from layer C5 to encode abstract category information and discriminative features;

[0105] Feature Pyramid Network (FPN): fuses features at different scales through top-down paths and lateral connections.

[0106] Upsample the C5 feature map to the same spatial dimension as C4;

[0107] The upsampled feature map is fused with the C4 feature map through a lateral connection (1×1 convolution);

[0108] Repeat the above process to fuse C3 and C2 features in sequence;

[0109] The fused features at each scale are smoothed using 3×3 convolutions;

[0110] The final result is a multi-scale feature pyramid {P2, P3, P4, P5}, which have the same channel dimension (256) but different spatial resolutions.

[0111] Step 2.2 introduces growth stage conditional encoding, encoding the cherry blossom growth stage information into conditional vectors and injecting them into the feature extraction process:

[0112] Growth stage encoding: The growth stages of cherry blossoms are encoded as conditional vectors, specifically including:

[0113] Establish a growth stage classification system: Divide the annual growth cycle of cherry blossoms into 6 main stages (initial stage, leaf unfolding stage, maturity stage, initial senescence stage, senescence stage, and dormancy stage).

[0114] Constructing the basic encoding: Encode these 6 stages into a 6-dimensional one-hot vector E;

[0115] Constructing continuous transition codes: In order to capture the smooth transition between adjacent growth stages, a Gaussian mixture model is used to generate continuous code values, that is, when the leaf is in the transition period between two stages, the code values ​​of the two adjacent stages are both greater than 0 and the sum is 1;

[0116] Enhanced coding: The basic coding vector is mapped to a high-dimensional embedding space (128 dimensions) through a fully connected layer to form a rich conditional representation;

[0117] Conditional batch normalization layer: Implements conditional parameter modulation in the feature extraction network.

[0118] In each residual block of ResNet, the standard batch normalization layer is replaced with a conditional batch normalization layer.

[0119] The scaling and offset parameters of CBN are dynamically generated through conditional vectors:

[0120] ;

[0121] ;

[0122] in, The scaling parameters for the conditional batch normalization layer; The offset parameters for the conditional batch normalization layer; This is the encoding vector for the growth stage; To generate fully connected layers with scaling parameters; To generate a fully connected layer with offset parameters.

[0123] Conditional batch normalization operation:

[0124] ;

[0125] in, The output features are normalized. For input features; The scaling parameters for the conditional batch normalization layer; The offset parameters for the conditional batch normalization layer; The mean of the characteristic batches; The standard deviation of the characteristic batch; To prevent small constants from being divided by zero.

[0126] In practical applications, for cherry blossom leaves in their early stages, the CBN parameter tends to increase the weight of fine texture features; while for mature leaves, the CBN parameter focuses more on color and shape features.

[0127] Feature map modulation: Channel-level modulation of convolutional feature maps based on growth stage encoding:

[0128] Learn a specific channel attention vector for each growth stage:

[0129] ;

[0130] in, This is the channel attention vector, used to modulate the feature importance of different channels; The sigmoid activation function compresses the output value to between 0 and 1; This is a fully connected layer used to convert the growth stage encoding into channel attention vectors; This is the encoding vector for the growth stage, containing information about the current growth stage.

[0131] Apply the channel attention vector to the feature map:

[0132] ;

[0133] in, The modulated feature map; This is the original feature map; This is the channel attention vector, used to modulate the feature importance of different channels; This indicates element-wise multiplication.

[0134] This modulation method enables the network to adaptively emphasize or suppress the characteristics of specific channels according to different growth stages.

[0135] In the application examples, the system emphasizes the light green channel for newly emerging cherry blossom leaves, while for mature leaves, the system focuses more on the dark green and texture channels.

[0136] Step 2.3: Construct a feature adaptation module to enable the feature extractor to adaptively adjust according to the growth stage.

[0137] Dynamic feature transformation matrix: Constructing a parameterized feature transformer that depends on the growth stage:

[0138] Learn a specific feature transformation matrix for each growth stage ,in, For feature dimensions.

[0139] Based on the current growth stage coding Calculate the weighted transformation matrix:

[0140] ;

[0141] in, This is a weighted transformation matrix; Encoding vectors for growth stages The Middle The encoded value of each stage (between 0 and 1); For the first Feature transformation matrix corresponding to each growth stage; This indicates a weighted summation of the six growth stages.

[0142] Apply the transformation matrix to the eigenvector:

[0143] ;

[0144] in, The transformed feature vector; This is the original feature vector; It is a weighted transformation matrix, which is a weighted combination of the feature transformation matrices of different growth stages.

[0145] This mechanism allows the model to adaptively project features into the feature space that best suits the current stage, depending on the different growth stages.

[0146] Dynamic convolution kernel generation network: Generates customized convolution parameters based on the growth stage.

[0147] Constructing a kernel-generated network:

[0148] ;

[0149] in, This is the set of convolution parameters, including kernel weights and bias terms; It is a multilayer perceptron used to convert the growth stage encoding into convolutional parameters; This is the encoding vector for the growth stage, containing information about the current growth stage.

[0150] Dynamic convolutional layers:

[0151] ;

[0152] in, This is the output feature map of the convolutional layer; The input feature map for the convolutional layer; This is the set of convolution parameters, including kernel weights and bias terms; This represents the convolution operation.

[0153] To reduce computational overhead, dynamic convolutions are used only in key layers, such as the fusion layer of a feature pyramid network.

[0154] In practical applications, dynamic convolutional kernels tend to learn fine edge filters for nascent leaves, while learning pattern filters with a larger receptive field for mature leaves.

[0155] Dual attention modulation module: integrates spatial and channel attention, dynamically adjusting feature maps according to the growth stage.

[0156] Channel attention generation:

[0157] ;

[0158] in, This is the channel attention vector, used to modulate the feature importance of different channels; The sigmoid activation function compresses the output value to between 0 and 1; It is a multilayer perceptron used for feature transformation and dimension mapping; This is an average pooling operation used to compress the spatial dimension of the feature map; Input feature map; This is the encoding vector for the growth stage, containing information about the current growth stage.

[0159] Spatial attention generation:

[0160] ;

[0161] in, This is a spatial attention map, used to modulate the importance of different spatial locations in the feature map; The sigmoid activation function compresses the output value to between 0 and 1; This represents the convolution operation; For the feature concatenation operation, the feature map and the growth stage are encoded and concatenated along the channel dimension; Input feature map; To facilitate repeated operations, the growth stages are encoded. The spatial dimension is repeatedly extended to have the same height and width as the feature map; This is the encoding vector for the growth stage, containing information about the current growth stage.

[0162] Attention application:

[0163] ;

[0164] in, The output feature map after applying attention; Input feature map; This is the channel attention vector, used to modulate the feature importance of different channels; This is a spatial attention map, used to modulate the importance of different spatial locations in the feature map; This indicates element-wise multiplication.

[0165] In the early stages of cherry blossom leaf spot disease, spatial attention is focused on the leaf edges and tiny spots; while in the later stages, attention shifts to the expanding lesion area and changes in the overall leaf morphology.

[0166] Step 2.4: Implement cross-stage feature mapping and establish the mapping relationship between feature spaces at different growth stages.

[0167] Feature alignment loss function: Design multiple loss functions to ensure feature alignment at different growth stages:

[0168] Absolute feature alignment loss:

[0169] ;

[0170] in, This is the absolute feature alignment loss, used to measure the degree of difference before and after feature mapping; Indicates from the growth stage Mapping to growth stage eigenvectors; Indicates growth stage The original feature vector; This represents the square of the L2 norm, used to calculate the Euclidean distance between two eigenvectors.

[0171] The goal of this loss function is to minimize the difference between the mapped features and the original features in the target stage.

[0172] Relative feature preservation loss:

[0173] ;

[0174] in, This is a relative feature preservation loss, used to preserve the relative relationship between samples before and after feature mapping; This indicates that the anchor sample has grown from the growth stage. Mapping to growth stage The feature vector after; This indicates that positive samples are in the growth stage. The original feature vector; This indicates that negative samples are in the growth phase. The original feature vector; Anchor samples are the samples that need to be feature mapped. A positive sample is a sample from a different viewpoint or time point that belongs to the same leaf as the anchor sample. A negative sample is a sample that does not belong to the same leaf as the anchor sample. This is the interval parameter, used to control the minimum distance difference between positive and negative sample pairs; This indicates the positive operation, i.e. ; This represents the square of the L2 norm, used to calculate the Euclidean distance between eigenvectors.

[0175] Combination loss:

[0176] ;

[0177] in, This is the total loss function, used to optimize the entire feature mapping network; This is the absolute feature alignment loss, used to measure the degree of difference before and after feature mapping; This is a relative feature preservation loss, used to preserve the relative relationship between samples before and after feature mapping; This is a balancing parameter used to adjust the relative importance of the two loss terms and control their contribution to the total loss.

[0178] Growth stage encoder-decoder network: Constructing a feature space transformation architecture:

[0179] Stage-specific encoder : Phase The features are mapped to a shared latent space;

[0180] Stage-specific decoder Decoding latent spatial features into stages Feature representation;

[0181] Cross-stage mapping:

[0182] ;

[0183] in, Indicates from the growth stage Mapping to growth stage eigenvectors; Indicates growth stage The original feature vector; For stage-specific encoders, features are mapped to a shared latent space; It is a stage-specific decoder that decodes latent spatial features into feature representations of the target stage.

[0184] This design enables the system to establish feature mapping relationships between different growth stages, thereby achieving knowledge transfer.

[0185] Periodic consistency constraint: Ensures the invertibility of feature mapping.

[0186] Forward mapping: ;

[0187] in, Indicates from the growth stage Mapping to growth stage eigenvectors; Indicates growth stage The original feature vector; Indicates from the growth stage to growth stage The mapping function.

[0188] Reverse mapping:

[0189] ;

[0190] in, Indicates from the growth stage Mapping back to the growth stage eigenvectors; Indicates growth stage The original feature vector; Indicates from the growth stage to growth stage The mapping function.

[0191] Periodic consistency constraints:

[0192] ;

[0193] ;

[0194] in, Indicates growth stage The original feature vector; Indicates growth stage The original feature vector; Indicates from the growth stage to growth stage The mapping function; Indicates from the growth stage to growth stage The mapping function; This indicates that both sides of the constraint should be approximately equal.

[0195] The corresponding loss function:

[0196] ;

[0197] in, Periodic consistency loss is used to measure the difference between the features after bidirectional mapping and the original features; The L1 norm is used to calculate the absolute difference between eigenvectors; the first term... Measurement phase Reconstruction error of features after bidirectional mapping; second term Measurement phase Reconstruction error of features after bidirectional mapping; Indicates growth stage The original feature vector; Indicates growth stage The original feature vector; Indicates from the growth stage to growth stage The mapping function; Indicates from the growth stage to growth stage The mapping function.

[0198] This constraint ensures bidirectional consistency of feature mapping and prevents information loss.

[0199] Identity Preservation Mechanism: Ensures that leaf identity information is preserved during cross-stage feature mapping.

[0200] Discriminative constraint: Train an identity discriminator Verify whether the features before and after mapping belong to the same leaf;

[0201] Loss of identity preservation:

[0202] ;

[0203] in, The identity preservation loss measures the degree to which the leaf identity information is preserved before and after feature mapping. This represents the expectation operation, which takes the average over all training samples; It serves as an identity discriminator, used to determine whether two feature vectors originate from the same leaf. Indicates growth stage The original feature vector; Indicates from the growth stage Mapping to growth stage eigenvectors; It is the natural logarithm function.

[0204] Contrastive learning: Construct positive and negative sample pairs to ensure that the characteristics of the same leaf remain similar at different growth stages;

[0205] This mechanism ensures that the individual identification information of the blades is not lost during the feature mapping process.

[0206] like Figure 3 As shown, the model update frequencies of traditional methods and the method of this invention are compared at different growth stages of cherry blossoms. It can be seen that traditional methods require frequent model updates to adapt to changes in different growth stages, especially in the budding and senescence stages. In contrast, the method of this invention, due to its bio-technological co-adaptive framework, maintains a low update frequency at all growth stages, demonstrating the system's adaptive capability and a significant reduction in maintenance costs.

[0207] Step 3: Construct a progressive knowledge base, which stores and manages feature prototypes and recognition rules from different periods through a hierarchical knowledge storage structure and knowledge update mechanism;

[0208] Specifically, the following steps are included:

[0209] Step 3.1: Establish a hierarchical knowledge storage structure, organizing knowledge into different levels:

[0210] Instance layer: Stores feature vectors and label information of typical samples;

[0211] Prototype layer: Stores feature prototypes of different categories and growth stages;

[0212] Rule layer: Stores recognition decision rules and relationships between categories;

[0213] Meta-knowledge layer: Stores knowledge update strategies and priority information.

[0214] Step 3.2: Implement knowledge caching and persistence management;

[0215] Short-term cache: Stores recently acquired samples and features in memory;

[0216] Long-term storage: Storing persistent knowledge in a database format;

[0217] Indexing system: Establishes a multi-dimensional index structure to support efficient retrieval;

[0218] Compression representation: Feature compression technology is used to reduce storage overhead.

[0219] Step 3.3: Construct a knowledge update and verification mechanism;

[0220] Incremental update: The knowledge base is incrementally updated after new samples pass the credibility assessment;

[0221] Conflict resolution: Set up conflict detection and resolution strategies to handle inconsistencies between old and new knowledge;

[0222] Knowledge validation: Regularly perform knowledge consistency checks and remove outdated or incorrect knowledge;

[0223] Version control: Maintains historical versions of the knowledge base and supports rollback operations.

[0224] Step 3.4: Implement knowledge management strategies related to the biological cycle;

[0225] Periodic organization: Organizing knowledge units according to the cherry blossom growth cycle;

[0226] Seasonal activation: Activate relevant knowledge based on the current season and growth stage;

[0227] Historical experience integration: Integrating data from the same period over many years to form periodic knowledge patterns;

[0228] Predictive caching: Preloading knowledge related to the upcoming growth stage.

[0229] like Figure 4 The figure shows a comparison of the disease detection rates of the method of this invention and the traditional method at different number of days after the onset of the disease. As can be seen from the figure, the method of this invention can detect disease signs as early as the 2nd to 3rd day after the onset of the disease, while the traditional method requires waiting until the 4th to 5th day to begin detection. By the 7th day, the method of this invention can detect nearly 95% of the diseases, reaching the same detection level about 3 days earlier than the traditional method, demonstrating the advantage of this invention in early disease detection.

[0230] Step 4: Construct a dual memory system, which distinguishes between long-term stable knowledge and short-term adaptive knowledge through long-term memory and short-term memory subsystems;

[0231] Specifically, the following steps are included:

[0232] Step 4.1 Constructing the long-term memory subsystem:

[0233] Core Feature Storage: Preserves the core identification features of cherry blossom leaf spot disease;

[0234] Invariant knowledge: storing identification rules that remain stable across growth cycles;

[0235] Long-term statistical model: A statistical distribution model built based on data from multiple years;

[0236] Category prototype: Long-term stable feature prototypes for each category.

[0237] Step 4.2, construct the short-term memory subsystem;

[0238] Environmental adaptation parameters: Adaptability parameters based on current environmental conditions;

[0239] Seasonal characteristics: Adjustments to characteristics of the current season and growth stage;

[0240] Recent Sample Cache: Features and labels of recently observed samples;

[0241] Provisional hypothesis model: A hypothesis model based on short-term observations.

[0242] Step 4.3: Implement the memory interaction and fusion mechanism;

[0243] Two-way information flow: a two-way information exchange channel between long and short-term memory;

[0244] Knowledge distillation: extracting stable portions from short-term memory for long-term memory;

[0245] Memory reset gate: A gating mechanism that controls the updating and resetting of short-term memory;

[0246] Long-term guidance: The constraint and guidance that long-term memory exerts on short-term memory.

[0247] Step 4.4, construct a memory update strategy;

[0248] Stability assessment: Assessing the time stability of knowledge to determine its attribution;

[0249] Importance-weighted: Assigning and updating weights based on the importance of knowledge;

[0250] Forgetting mechanisms: achieving controlled decay of short-term memory;

[0251] Consolidation mechanism: Transferring frequently activated short-term knowledge to long-term memory.

[0252] like Figure 5 The diagram shows a comparison of the accuracy of the proposed method and traditional methods in identifying the six main growth stages of cherry blossoms. It can be seen that the proposed method maintains a high accuracy of over 90% across all growth stages, with a fluctuation range of only 1.7%; while the accuracy of the traditional method varies significantly with the growth stage, reaching its lowest point during the senescence stage, with a fluctuation range as high as 14.4%. This verifies that the growth stage-aware feature extractor and dual-memory system of the proposed method can effectively adapt to feature changes at different growth stages.

[0253] Step 5: Construct a concept drift detector to monitor the long-term trends of disease manifestations and plant characteristics through multi-granularity feature distribution monitoring and multi-timescale change detection.

[0254] Specifically, the following steps are included:

[0255] Step 5.1: Construct a multi-granularity feature distribution monitoring system;

[0256] Original feature distribution: Monitors changes in the statistical distribution of low-level visual features (such as color and texture);

[0257] Semantic feature distribution: Monitoring changes in the statistical distribution of high-level semantic features;

[0258] Discrimination boundary: Monitor changes in the classification decision boundary;

[0259] Prediction confidence: Monitor changes in the confidence distribution of prediction results.

[0260] Step 5.2: Implement multi-timescale change detection;

[0261] Short-term fluctuation detection: Detects daily or weekly fluctuations in data distribution;

[0262] Mid-term trend detection: Detects the changing trends in data distribution each month or quarter;

[0263] Long-term cycle detection: Detecting cyclical change patterns over an annual or multi-year period;

[0264] Non-periodic mutation detection: Detects mutation events that do not conform to historical patterns.

[0265] Step 5.3: Construct the distributed offset quantization model;

[0266] Statistical distance calculation: using KL divergence, Wasserstein distance, etc. to measure distribution change;

[0267] Rate of change assessment: Calculate the rate and acceleration of change in the distribution;

[0268] Threshold Adaptive: Dynamically adjusts the drift determination threshold based on historical data;

[0269] Confidence interval: Establish confidence intervals for normal fluctuations and detect abnormal deviations.

[0270] Step 5.4, design the drift response and early warning mechanism;

[0271] Drift classification: The detected drift is classified as periodic change, temporary change caused by the environment, or essential change in disease characteristics;

[0272] Severity assessment: Assessing the potential impact of drift on recognition performance;

[0273] Warning Trigger: Different levels of warnings are triggered based on the drift type and severity;

[0274] Update recommendations: Generate model update recommendations for different types of drift.

[0275] like Figure 6 The figure shows the percentage improvement in performance across five key metrics compared to traditional methods. It can be seen that the most significant improvements are in reduced maintenance costs, data efficiency, and early detection, reaching 77%, 78%, and 28.3%, respectively; there are also significant improvements in recognition accuracy and environmental adaptability. These data vividly demonstrate the technical effectiveness and practical value of the present invention.

[0276] Step 6: Construct a meta-learning optimizer, which learns the optimal adaptation strategy from the historical adaptation process through meta-parameter space optimization and historical experience learning;

[0277] Specifically, the following steps are included:

[0278] Step 6.1: Construct the meta-parameter space and optimization objective;

[0279] Meta-parameter definition: Defines the set of meta-parameters that control the model update process, including learning rate scheduling, memory update threshold, knowledge transfer rate, etc.

[0280] Multi-objective function: Defines a comprehensive optimization objective that includes multiple dimensions such as accuracy, stability, and computational efficiency;

[0281] Constraints: Set constraints on stability and reliability to prevent over-optimization of a single metric;

[0282] Reward function: Design a reward function that reflects long-term performance and avoids the trap of short-term metrics.

[0283] Step 6.2: Constructing data collection and representation for the meta-learning process;

[0284] Experience trajectory: Records the status, actions, and results during the historical update process;

[0285] Context representation: Encode the current environment, task, and model state into a context vector;

[0286] Performance metrics: Collect multidimensional performance change metrics before and after model updates;

[0287] Metadata annotation: Add meta tags to the collected experience data, such as growth stage, seasonal information, etc.

[0288] Step 6.3: Implement meta-policy learning based on historical experience;

[0289] Pattern extraction: Extracting successful and failed patterns from historical update experiences;

[0290] Condition mapping: Learning the mapping relationship from environmental conditions to the optimal update strategy;

[0291] Policy Network: Train a policy network to output the optimal meta-parameter settings for a given scenario;

[0292] Reinforcement learning: Continuously optimize and update strategies through reinforcement learning algorithms to maximize long-term rewards.

[0293] Step 6.4: Construct an adaptive execution and evaluation mechanism;

[0294] Exploration-Utilization Balance: Automatically adjusts the ratio of exploration to utilization based on the degree of certainty;

[0295] Meta-parameter scheduling: dynamically adjusting meta-parameters during execution;

[0296] A / B testing: Conducting comparative experiments in a local environment to verify the effectiveness of optimization strategies;

[0297] Feedback loop: Construct a feedback mechanism for model update results to continuously improve the meta-learning system.

[0298] like Figure 7As shown, the long-term operational stability of the system of this invention and the traditional method is compared over a 12-month period. It can be seen that the method of this invention maintains a consistently high accuracy rate throughout the year, with a fluctuation range of only 1.8%; while the accuracy rate of the traditional method exhibits significant seasonal fluctuations, dropping to its lowest point (approximately 72%) in the summer (June-July), with a fluctuation range as high as 19.7%. This verifies that the bio-technological co-adaptation framework of this invention can effectively solve the stability problem in long-term applications.

[0299] Step 7: Integrate the biotechnology co-adaptation framework to achieve long-term synergy between the cherry blossom leaf spot disease identification system and the cherry blossom growth cycle through system architecture and biological clock synchronization mechanism;

[0300] Specifically, the following steps are included:

[0301] Step 7.1: Construct the overall system architecture;

[0302] Multi-level design: Organize the system into a data layer, feature layer, knowledge layer, and decision layer;

[0303] Modular structure: It adopts a loosely coupled modular design, which facilitates flexible configuration and upgrades;

[0304] Communication interface: Defines standardized inter-module communication interfaces and data formats;

[0305] Event-driven: Adopts an event-driven architecture to achieve asynchronous collaboration between modules.

[0306] Step 7.2: Implement the biological clock synchronization mechanism;

[0307] Growth cycle modeling: Constructing a mathematical model of the cherry blossom growth cycle, including leaf morphological development and seasonal changes;

[0308] Model cycle synchronization: Synchronize the model update strategy with the plant growth cycle;

[0309] Predictive adjustment: Based on the growth model, the identification strategy is adjusted in advance to adapt to upcoming changes;

[0310] Periodic reset: Perform partial model reset at appropriate growth nodes to avoid accumulating errors.

[0311] Step 7.3, Build an adaptive reasoning engine;

[0312] Multi-model fusion: Dynamically fuses multiple expert models based on the current environment and growth stage;

[0313] Context awareness: Considering time, space, and environmental context during reasoning;

[0314] Confidence assessment: For each prediction result, a reliability assessment is performed to generate a confidence interval;

[0315] Decision adjustment: Adjust decision thresholds based on risk appetite and application scenario.

[0316] Step 7.4: Achieve system self-reflection and optimization;

[0317] Performance monitoring: Continuously monitor key system performance indicators, such as accuracy, stability, and efficiency;

[0318] Error analysis: Automatically analyzes failure cases and identifies system weaknesses;

[0319] Improvement suggestions: Generate targeted improvement suggestions to guide system optimization;

[0320] Adaptive adjustment: Automatically adjusts system configuration parameters based on long-term performance trends.

[0321] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for identifying cherry blossom leaf spot disease based on deep convolutional networks, characterized in that, Includes the following steps: Establish a dual dynamic system mathematical model of plant growth and model evolution to uniformly represent the plant growth process and the model evolution process; The mathematical model of the dual dynamic system of plant growth and model evolution includes: Plant state dynamic evolution model: used to describe the changing pattern of plant state over time, and calculate the rate of change of plant state based on the current state of the plant and environmental conditions; Identify the dynamic evolution model of model parameters: This model is used to describe the update pattern of model parameters over time and calculates the rate of change of model parameters based on the current model parameters, plant status, and collected data. System coupling model: used to describe the acquisition process of plant status observation data and the dynamic update mechanism of environmental conditions. The update of environmental conditions takes into account the influence of model parameters and plant status. A growth stage-aware feature extractor is constructed. Through a multi-scale feature extraction network and growth stage conditional coding, the feature extractor can adapt to the differences in cherry blossom leaf features at different growth stages. Construct a progressive knowledge base, which stores and manages feature prototypes and recognition rules from different periods through a hierarchical knowledge storage structure and knowledge update mechanism; Construct a dual memory system that distinguishes between long-term stable knowledge and short-term adaptive knowledge through long-term memory and short-term memory subsystems; A concept drift detector was constructed to monitor the long-term trends of disease manifestations and plant characteristics through multi-granularity feature distribution monitoring and multi-timescale change detection. A meta-learning optimizer is constructed, which learns the optimal adaptation strategy from the historical adaptation process through meta-parameter space optimization and historical experience learning; By integrating a biotechnology co-adaptation framework and through system architecture and biological clock synchronization mechanisms, a long-term synergy between the cherry blossom leaf spot disease identification system and the cherry blossom growth cycle can be achieved.

2. The method for identifying cherry blossom leaf spot disease based on deep convolutional networks according to claim 1, characterized in that, The feature extractor for growth stage perception includes: Multi-scale feature extraction network: includes a backbone network, a multi-scale feature extraction module, and a feature pyramid network, which extracts and fuses features from different scales; Growth stage conditional encoding: This includes growth stage encoding, conditional batch normalization layer and feature map modulation, which encodes cherry blossom growth stage information and injects it into the feature extraction process; Feature adaptation module: includes dynamic feature transformation matrix, dynamic convolutional kernel generation network and dual attention modulation module, which enables the feature extractor to adaptively adjust according to the growth stage; Cross-stage feature mapping: This includes feature alignment loss function, encoder-decoder network for different growth stages, periodic consistency constraint and identity preservation mechanism, to establish the mapping relationship between feature spaces at different growth stages.

3. The method for identifying cherry blossom leaf spot disease based on deep convolutional networks according to claim 1, characterized in that, The progressive knowledge base includes: Hierarchical knowledge storage structure: knowledge is organized into an instance layer, a prototype layer, a rule layer, and a meta-knowledge layer; Knowledge caching and persistence management includes short-term caching, long-term storage, indexing systems, and compressed representation; Knowledge update and verification mechanisms include incremental updates, conflict resolution, knowledge verification, and version control. Knowledge management strategies related to biological cycles include periodic organization, seasonal activation, integration of historical experience, and predictive caching.

4. The method for identifying cherry blossom leaf spot disease based on deep convolutional networks according to claim 1, characterized in that, Dual memory systems include: Long-term memory subsystem: stores core features, invariant knowledge, long-term statistical models, and category prototypes; Short-term memory subsystem: storage environment adaptation parameters, seasonality characteristics, recent sample caching, and temporary hypothesis model; Memory interaction and fusion mechanisms include bidirectional information flow, knowledge distillation, memory reset gates, and long-term guidance; Memory update strategies include stability assessment, importance weighting, forgetting mechanisms, and consolidation mechanisms.

5. The method for identifying cherry blossom leaf spot disease based on deep convolutional networks according to claim 1, characterized in that, Concept drift detectors include: Multi-granularity feature distribution monitoring system: monitors the original feature distribution, semantic feature distribution, discrimination boundary, and prediction confidence; Multi-timescale change detection: including short-term fluctuation detection, medium-term trend detection, long-term cycle detection, and non-periodic mutation detection; Distribution offset quantification model: including statistical distance calculation, rate of change assessment, threshold adaptation, and confidence interval; Drift response and early warning mechanism: including drift classification, severity assessment, early warning triggering, and update recommendations.

6. The method for identifying cherry blossom leaf spot disease based on deep convolutional networks according to claim 1, characterized in that, Meta-learning optimizers include: Meta-parameter space and optimization objectives: including meta-parameter definitions, multi-objective functions, constraints, and reward functions; Data collection and representation in the meta-learning process: including empirical trajectories, contextual representations, performance metrics, and metadata annotation; Meta-policy learning based on historical experience includes pattern extraction, conditional mapping, policy networks, and reinforcement learning. Adaptive execution and evaluation mechanisms include exploratory balancing, metaparameter scheduling, A / B testing, and feedback loops.

7. The method for identifying cherry blossom leaf spot disease based on deep convolutional networks according to claim 1, characterized in that, The biotechnology co-adaptation framework includes: System overall architecture: including multi-level design, modular structure, communication interfaces, and event-driven architecture; Biological clock synchronization mechanisms include growth cycle modeling, model cycle synchronization, predictive adjustment, and periodic reset. Adaptive reasoning engine: including multi-model fusion, context awareness, confidence assessment, and decision adjustment; System self-reflection and optimization: including performance monitoring, error analysis, improvement suggestions, and adaptive adjustments.

8. A cherry blossom leaf spot disease identification system based on deep convolutional networks, used to execute the cherry blossom leaf spot disease identification method based on deep convolutional networks as described in any one of claims 1-7, characterized in that, include: Image acquisition module: used to acquire images of cherry blossom leaves; Data preprocessing module: used to standardize, enhance, and perform quality control on the acquired images; Dual dynamic system mathematical model: used to uniformly represent the plant growth process and the model evolution process; Biotechnology Co-adaptation Framework: A comprehensive process for identifying cherry blossom leaf spot disease, which achieves long-term synergy between the cherry blossom leaf spot disease identification system and the cherry blossom growth cycle through system architecture and biological clock synchronization mechanism; Results display module: used to present the recognition results in a visual way; Early warning and intervention module: used to generate early warning information and provide intervention suggestions based on the identification results; Environmental sensing module: used to collect environmental parameters and provide environmental condition vectors for the model; Edge computing unit: used for preprocessing and initial identification at the data acquisition site; Cloud servers: used to perform complex calculations, store long-term knowledge bases, and coordinate multi-location systems; Human-computer interface: used to receive expert feedback and manually adjust system parameters; API interface: Used for integration with other agricultural management systems.

Citation Information

Patent Citations

  • Multi-disease detection method in crop intercropping mode

    CN118314462A

  • Cotton leaf disease identification method and system based on multilayer context feature screening

    CN119251813A