A Method and System for Pest and Disease Image Recognition Based on a Lightweight Neural Network Model

Abstract

This application provides a method and system for pest and disease image recognition based on a lightweight neural network model, relating to the fields of smart agriculture and computer vision. The method includes: acquiring multimodal observation data of the target crop at current and historical moments; extracting texture, edge, and thermal infrared features based on the current moment data and calculating a data complexity index; inputting historical moment data into a time-series model to obtain a disease state sequence and predicting the disease development level to calculate a predicted risk level; determining the target computational depth based on the predicted risk level and the data complexity index; determining the target network branch from multiple feature extraction network branches accordingly; and using the target network branch to extract features from the current moment data to obtain disease characterization features and identify pests and diseases. This solves the technical problem in existing technologies where fixed-structure networks cannot adaptively adjust computational resources according to the dynamic complexity of the disease.

Description

Technical Field

[0001] This application relates to the fields of smart agriculture and computer vision, and in particular to a method and system for identifying pest and disease images based on a lightweight neural network model. Background Technology

[0002] Existing technologies typically rely on fixed-structure deep learning models to extract features and classify acquired crop images, executing the same computational path and inference process regardless of the complexity of the image content or the severity of disease development. However, this fixed-structure inference mode still performs full deep learning computation when facing early or mild cases with simple backgrounds and obvious disease features, resulting in significant waste of computing resources and increased energy consumption, making it difficult to meet the real-time and low-power requirements of mobile devices or edge computing nodes. Furthermore, when facing scenarios with drastic changes in lighting, severe occlusion, or moderate to severe diseases with complex features, fixed-structure models often suffer from insufficient feature extraction due to inadequate computational allocation, leading to decreased recognition accuracy, missed detections, or false detections. Therefore, a method is urgently needed to address the technical problem that existing fixed-structure networks cannot adaptively adjust computational resources according to the dynamic complexity of disease conditions. Summary of the Invention

[0003] This application provides a method and system for disease and pest image recognition based on a lightweight neural network model, which solves the technical problem in the prior art that fixed-structure networks cannot adaptively adjust computing resources according to the dynamic complexity of diseases.

[0004] To achieve the above objectives, this application adopts the following technical solution: In a first aspect, a method for pest and disease image recognition based on a lightweight neural network model is provided, characterized by comprising: acquiring multimodal observation data of the target crop at the current moment and multimodal observation data at historical moments, wherein the multimodal observation data includes visible light images, thermal infrared images, and near-infrared images; extracting texture features, edge features, and thermal infrared features based on the multimodal observation data at the current moment, and calculating the data complexity index at the current moment based on the texture features, edge features, and thermal infrared features; inputting the multimodal observation data at historical moments into a time-series model for processing in chronological order to obtain a disease state sequence; predicting the disease development level corresponding to multiple future moments based on the disease state sequence, and calculating the predicted risk level based on the disease development level corresponding to multiple future moments; determining the target computation depth based on the predicted risk level and the data complexity index; determining the target network branch from multiple feature extraction network branches with different computation depths based on the target computation depth; using the target network branch to extract features from the multimodal observation data at the current moment to obtain disease characterization features; and identifying pests and diseases based on the disease characterization features.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, the target computation depth is determined based on the predicted risk level and the data complexity index, including: normalizing the predicted risk level to obtain a risk score; normalizing the data complexity index to obtain a complexity score; weighting and summing the risk score and the complexity score according to preset weights to obtain a comprehensive scheduling score; comparing the comprehensive scheduling score with scheduling thresholds corresponding to multiple computation depths; and determining the target computation depth based on the comparison results.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves inputting multimodal observation data from historical moments into a time-series model in chronological order to obtain a disease state sequence. This includes: inputting multimodal observation data from multiple consecutive historical moments into the time-series model in chronological order; extracting state transition features between each moment using the time-series model; generating a disease state sequence based on the state transition features; predicting the disease development level corresponding to multiple future moments based on the disease state sequence; calculating the predicted deterioration probability based on the change magnitude of the disease development level, and using the predicted deterioration probability as the predicted risk level.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the risk level satisfies the following formula:

[0008] in, This indicates the disease development level at the current time t. This represents the disease development level corresponding to the k-th time step in the future, where K represents the length of the predicted time step used for risk assessment. This represents the weighting coefficients corresponding to different future time steps; The adjustment coefficient representing the acceleration term of disease development; This represents the complexity adjustment factor; This represents the data complexity index calculated based on the fusion of texture features, edge features, and thermal infrared features at the current moment. This represents the nonlinear adjustment coefficient; This represents a complexity threshold, used to distinguish between low-complexity and high-complexity scenarios. This represents the time corresponding to the s-th time step in the time series.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the calculation of the data complexity index at the current moment includes: processing the visible light image using a texture analysis operator to obtain texture features; processing the visible light image using an edge detection operator to obtain edge features; performing temperature distribution analysis on the thermal infrared image to obtain thermal infrared features; and using the texture features, edge features, and thermal infrared features as data complexity evaluation parameters.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the data complexity index at the current moment is calculated by: calculating the texture variance value of the texture feature, the edge density value of the edge feature, and the temperature dispersion value of the thermal infrared feature respectively; weighting the texture variance value, edge density value, and temperature dispersion value according to preset weights; and using the weighted summation result as the data complexity index.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, before using the target network branch for feature extraction, the method further includes: extracting thermal anomalies from the thermal infrared image; extracting near-infrared reflection anomalies from the near-infrared image; extracting color anomalies from the visible light image; calculating a disease anomaly index based on the thermal anomalies, near-infrared reflection anomalies, and color anomalies; and weighting the multimodal observation data according to the disease anomaly index.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the disease anomaly index is calculated based on thermal anomalies, near-infrared reflectance anomalies, and color anomalies, including: normalizing the thermal anomalies, near-infrared reflectance anomalies, and color anomalies respectively; weighting the normalized thermal anomalies, near-infrared reflectance anomalies, and color anomalies according to their corresponding weights; and using the weighted summation result as the disease anomaly index.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, the method further includes: generating an anomaly distribution map based on the disease anomaly index; extracting connected regions in the anomaly distribution map whose disease anomaly index is greater than a preset threshold; identifying the connected regions as high-confidence candidate regions; and using target network branches to extract features and identify diseases and pests in the high-confidence candidate regions.

[0014] Secondly, a pest and disease image recognition system based on a lightweight neural network model is provided, comprising: a data acquisition module for acquiring multimodal observation data of the target crop at the current moment and multimodal observation data at historical moments; a complexity calculation module for calculating a data complexity index based on the multimodal observation data at the current moment; a time series prediction module for generating a disease state sequence based on the multimodal observation data at historical moments and predicting the risk level; a depth scheduling module for determining the target computation depth based on the predicted risk level and the data complexity index; a network selection module for determining the target network branch based on the target computation depth; and a recognition module for identifying pests and diseases using the target network branch.

[0015] This application provides a method and system for pest and disease image recognition based on a lightweight neural network model. By introducing a core mechanism that uses temporal evolution prediction results to drive dynamic reconstruction of the network structure, and combining the current data complexity index with the predicted risk level to jointly determine the target computational depth, it completely changes the rigid inference mode of fixed input in existing technologies. When the prediction shows a rapid deterioration trend in the pest and the current image complexity is high, the system automatically increases the network depth to activate the deep path, ensuring recognition accuracy in high-risk scenarios and preventing missed detections. When the prediction shows the pest is stabilizing and the scenario is simple, the system automatically switches to the lightweight path, significantly reducing computational load and energy consumption. This achieves adaptive resource allocation of less computation for simple pests and more computation for complex pests, fundamentally solving the technical problem that fixed-structure networks cannot adaptively adjust computational resources according to the dynamic complexity of the pest, and accurately balancing the contradiction between computational power consumption and recognition accuracy on mobile devices in complex scenarios. Meanwhile, by introducing a disease anomaly index based on thermal, water, and color anomalies for regional localization, background noise is effectively filtered and unnecessary computation is reduced. Through cross-modal feature fusion and interpretability risk index calculation, the detection rate of early diseases and the reliability of the system are further improved, realizing a leap from passive identification to active prediction and adaptive resource allocation. Attached Figure Description

[0016] Figure 1 A system architecture diagram of a pest and disease image recognition system based on a lightweight neural network model is provided for embodiments of this application; Figure 2 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model, provided in Embodiment 2 of this application. Figure 1 ; Figure 3 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model provided in Embodiment 3 of this application. Figure 2 ; Figure 4 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model provided in Embodiment 4 of this application. Figure 3 ; Figure 5 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model, provided in Embodiment 5 of this application. Figure 4 ; Figure 6 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model provided in Embodiment Six of this application. Figure 5 ; Figure 7 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model provided in Embodiment 7 of this application. Figure 6. Detailed Implementation

[0017] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0018] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0019] Example 1:

[0020] The pest and disease image recognition method based on a lightweight neural network model provided in this application can be applied to, for example... Figure 1 The pest and disease image recognition system shown is based on a lightweight neural network model, such as... Figure 1 As shown, the system includes: a data acquisition module 101, used to acquire multimodal observation data of the target crop at the current moment and multimodal observation data at historical moments; a complexity calculation module 102, used to calculate the data complexity index based on the multimodal observation data at the current moment; a time series prediction module 103, used to generate a disease status sequence based on the multimodal observation data at historical moments and predict the prediction risk level; a depth scheduling module 104, used to determine the target computation depth according to the prediction risk level and the data complexity index; a network selection module 105, used to determine the target network branch according to the target computation depth; and an identification module 106, used to identify pests and diseases using the target network branch.

[0021] Example 2:

[0022] To address the technical problem in existing technologies where fixed-structure networks cannot adaptively adjust computational resources according to the dynamic complexity of pests, this application provides a pest image recognition method based on a lightweight neural network model. Figure 2 A flowchart illustrating the pest and disease image recognition method based on a lightweight neural network model provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method includes: S201. Obtain the current multimodal observation data and historical multimodal observation data of the target crop.

[0023] Multimodal observation data refers to a collection of data that reflects the multidimensional physiological state of crops, acquired through different physical principles or sensing mechanisms. Multimodal observation data includes visible light images, thermal infrared images, and near-infrared images.

[0024] In this embodiment, the system synchronously acquires multidimensional physiological state data of the target crop through intelligent sensing terminals deployed in the field. The acquisition equipment includes a high-resolution visible light camera, a thermal infrared imager, and a near-infrared spectral sensor. It can also connect to weather stations and soil sensor networks to obtain environmental parameters. The system encapsulates the above heterogeneous data into a unified disease observation vector, breaking the limitation of traditional single-modal recognition that relies solely on visual information.

[0025] It should be noted that the multimodal observation data of historical moments does not refer to a snapshot of a single past moment, but rather to a sequence of data containing multiple consecutive time steps, used to capture the dynamic evolution of diseases.

[0026] S202. Extract texture features, edge features, and thermal infrared features based on the multimodal observation data at the current moment, and calculate the data complexity index at the current moment based on the texture features, edge features, and thermal infrared features.

[0027] Among them, the data complexity index refers to a comprehensive evaluation index used to quantify the disorder of the microstructure of lesions, the clarity of their outlines, and the degree of abnormality in thermal distribution in the current scenario.

[0028] In this embodiment of the application, the system extracts texture features and edge features reflecting the morphology of lesions from visible light images, and extracts thermal infrared features reflecting the abnormal temperature distribution and transpiration of crop canopies from thermal infrared images. The above three types of features are mapped to a unified complexity value through a preset evaluation function.

[0029] It should be noted that the purpose of calculating the data complexity index is to evaluate the difficulty of current image feature extraction in the spatial dimension, thereby providing a static basis for the scheduling of subsequent network depth at the current moment.

[0030] Based on the above steps, a multi-dimensional complexity evaluation system was constructed, which enables the complexity index to comprehensively reflect the microstructure, outline clarity and thermal distribution intensity of lesions, avoiding the one-sidedness of evaluation based on a single visual feature.

[0031] S203. Input the multimodal observation data of historical moments into the time series model in chronological order for processing to obtain the disease state sequence.

[0032] Among them, the time series model refers to the neural network model that can process data with time dependence, and the disease state sequence refers to the vector sequence that characterizes the spatiotemporal evolution of the disease over a continuous time step.

[0033] In this embodiment of the application, the system inputs multimodal observation data from multiple consecutive historical moments into a time series model in chronological order. The model extracts the state transition features between each moment through its internal memory mechanism, thereby generating a disease state sequence that can reflect the dynamic evolution of the disease.

[0034] It should be noted that the macroscopic logic of time series model processing lies in capturing the microscopic change trend of diseases over time, rather than simply encoding the static state at a single moment.

[0035] Based on the above steps, the system achieves accurate modeling of the dynamic evolution of diseases, transforming discrete historical observation data into continuous spatiotemporal evolution feature expressions.

[0036] S204. Based on the disease state sequence, predict the disease development level corresponding to multiple future moments, and calculate the predicted risk level based on the disease development level corresponding to multiple future moments.

[0037] Among them, the disease development level refers to the indicator that quantifies and classifies the severity of the disease, while the predicted risk level refers to the risk quantification value calculated based on the future trend of disease deterioration and used to guide the allocation of network resources in reverse.

[0038] In this embodiment, the system, based on the disease state sequence output by the time-series model, maps the disease development levels corresponding to multiple different time steps in the future through a prediction head, and then calculates the predicted risk level based on the magnitude and trend of these future levels. It should be understood that the calculation logic for the predicted risk level is not limited to simple difference accumulation, but may also include consideration of the acceleration of deterioration or nonlinear trends.

[0039] It should be noted that this step is the core link in establishing a predictive guidance identification closed loop, transforming qualitative trends into quantitative control variables, enabling the system to adjust the current allocation of computing resources in advance based on future risk predictions.

[0040] S205. Determine the target computation depth based on the predicted risk level and data complexity index.

[0041] The target computation depth refers to the computation level or resource consumption corresponding to the neural network inference path that the system decides to invoke at the current moment.

[0042] In this embodiment, the system jointly decides on the predicted risk level in the time dimension and the data complexity index in the spatial dimension, and determines the target computational depth through a qualitative or quantitative mapping mechanism. When the prediction shows that the disease is rapidly deteriorating and the current image complexity is high, the system determines a higher target computational depth; when the prediction shows that the disease is stabilizing and the scene is simple, the system determines a lower target computational depth.

[0043] It should be noted that this logic of jointly controlling the computation depth by combining the temporal evolution prediction results with the current complexity completely changes the rigid mode of fixed input in existing technologies.

[0044] Based on the above steps, the system achieves a unified quantitative decision-making process for both time-dimensional risk and spatial-dimensional complexity, ensuring a precise match between the allocation of computing resources and the actual dynamic needs of disease.

[0045] S206. Based on the target computation depth, determine the target network branch from multiple feature extraction network branches with different computation depths.

[0046] Among them, the feature extraction network branch refers to a parallel inference path with different network layers, number of parameters or computational complexity, while the target network branch refers to a specific inference path that is about to be activated and executed and matches the currently determined target computational depth.

[0047] In this embodiment, the system pre-deploys multiple feature extraction network branches with different computational depths. These branches share some low-level features to reduce storage overhead, but have different depths and accuracies in high-level semantic extraction. Based on the target computational depth output in S205, the system dynamically routes and activates the corresponding target network branch, enabling the network structure to adaptively reconstruct according to disease risk and complexity.

[0048] It should be noted that the macroscopic mechanism of multi-branch network selection realizes the "on-demand allocation" computing strategy, avoiding the waste of computing power in simple scenarios and the lack of accuracy in complex scenarios by a single fixed network.

[0049] As an example, multiple feature extraction network branches with different computational depths can include a lightweight path with the fewest parameters, a standard path that balances accuracy and efficiency, and a deep path that includes more convolutional layers and attention modules.

[0050] Based on the above steps, the system implements an adaptive resource allocation strategy that requires less computation for simple diseases and more computation for complex diseases, thus resolving the contradiction between efficiency and accuracy in a fixed model.

[0051] S207. Use the target network branch to extract features from the multimodal observation data at the current moment to obtain the disease characterization features.

[0052] Among them, the disease characterization features refer to the multimodal fusion feature vectors that can highly summarize the semantic information of the disease after deep processing of the target network branches.

[0053] In this embodiment, the system inputs the current multimodal observation data into the target network branch determined in S206. This branch performs layer-by-layer abstraction and cross-modal fusion of the input data according to its own computational depth, and finally outputs the disease characterization features. It should be understood that the feature extraction process is not limited to simple convolution operations, but may also include mechanisms such as cross-modal attention fusion to highlight abnormal areas of temperature, moisture, and color.

[0054] It should be noted that since the target network branches are dynamically selected based on predicted risk and complexity, the extracted disease characterization features are highly adapted to the current scenario requirements in terms of both accuracy and efficiency.

[0055] Based on the above steps, the system completes feature extraction under a dynamically reconstructed network architecture, ensuring sufficiency of feature extraction in high-risk and complex scenarios and computational efficiency in low-risk and simple scenarios.

[0056] S208. Identify pests and diseases based on their symptom characteristics.

[0057] Among them, pest and disease identification refers to the classification process of determining the type, degree of occurrence, and confidence level of a disease based on the extracted disease characteristics.

[0058] In this embodiment, the system inputs the disease characterization features output by the target network branch into the classifier, which outputs the disease category probability distribution and the current level determination result. It should be understood that the identification process can further include interpretability analysis, generating temperature, color, and moisture contribution maps to visually demonstrate the determination criteria.

[0059] It should be noted that the identification results not only serve as the final output to guide prevention and control decisions, but their confidence and level information can also be used as feedback signals for time series prediction updates at the next moment.

[0060] As an example, the classifier uses a combination of fully connected layers and the Softmax activation function to output specific judgment results such as "suspected early stage of Fusarium head blight" or "severe outbreak of Fusarium head blight" and the corresponding confidence percentage.

[0061] This embodiment introduces a core mechanism that uses the results of temporal evolution prediction to drive the dynamic reconstruction of the network structure. Combined with the current data complexity index, the system can intelligently allocate computing resources according to the dynamic trend of disease development and the complexity of the current scenario. This achieves an adaptive resource allocation strategy that requires less computation for simple diseases and more computation for complex diseases. It resolves the contradiction between the waste of computing power in simple scenarios and the insufficient accuracy in complex scenarios for fixed models, and accurately balances the contradiction between the computing power consumption and recognition accuracy of edge devices in complex scenarios.

[0062] Example 3:

[0063] In one possible implementation of the embodiments of this application, Figure 3 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model provided in Embodiment 3 of this application. Figure 2 , combined Figure 2 ,like Figure 3 As shown, the above S202 can be specifically implemented through the following S301 to S304, which are explained in detail below: S301. The visible light image is processed using a texture analysis operator to obtain texture features.

[0064] Among them, texture analysis operators refer to mathematical transformation tools that can extract from images the regularity of surface microstructure and the periodicity of gray-level distribution.

[0065] In this embodiment, the system calls a texture analysis operator to perform two-dimensional spatial filtering or local pattern encoding on the microscopic attributes of the lesion region in a visible light image, such as roughness, directionality, and local gray-level variation, thereby outputting a texture feature vector that can characterize the texture disorder of the lesion. It should be understood that the specific implementation of the texture analysis operator is not unique, and the system can adaptively select the most matching operator family according to the lesion type and image resolution.

[0066] It should be noted that texture features directly reflect the degree of microstructural disorder of cell necrosis and pigment deposition inside lesions, and are an important visual dimension for assessing the difficulty of extracting features from early micro lesions.

[0067] As one example, the system can use the Gabor filter family to perform multi-scale and multi-directional convolution processing on visible light images to capture texture responses at different frequencies and directions; as another example, the system can also use local binary mode operators to extract the relative gray-level changes in the pixel neighborhood to efficiently encode the local texture microstructure.

[0068] Based on the above steps, the system extracts texture features that reflect the disorder of the microstructure through flexible texture analysis operators, providing accurate visual structural dimension input for subsequent complex quantification.

[0069] S302. The visible light image is processed using an edge detection operator to obtain edge features.

[0070] Among them, the edge detection operator refers to the gradient calculation tool used to identify the set of pixels in an image where the gray level changes drastically, thereby outlining the contour and shape boundary of the lesion.

[0071] In this embodiment, the system applies gradient operations or first- or second-order differential operations to the visible light image to accurately locate the gray-level transition boundary between lesions and healthy tissue, and outputs an edge feature map that reflects the clarity of the contour and the complexity of the shape.

[0072] It should be noted that edge features reflect the clarity and complexity of the lesion outline. When the lesion edge is blurred or intertwined with the background, the difficulty of extracting edge features increases significantly. This is precisely the spatial dimension factor that the data complexity index needs to consider.

[0073] As one example, the system uses the Canny operator combined with a dual-threshold hysteresis detection mechanism to extract continuous and fine lesion edges to maintain high positioning accuracy under complex lighting conditions. As another example, the system uses the Sobel operator to quickly calculate gradient magnitude through local difference to achieve extremely fast edge feature extraction in edge scenarios with limited computing power.

[0074] Based on the above steps, the system extracts edge features that reflect the clarity of the contour by using edge detection operators adapted to different scenario requirements, providing an objective basis for the shape boundary dimension for complexity assessment.

[0075] S303. Perform temperature distribution analysis on the thermal infrared image to obtain thermal infrared characteristics.

[0076] Temperature distribution analysis refers to the process of spatial statistics and difference calculation of the physical temperature values ​​mapped by pixel grayscale in thermal infrared images.

[0077] In this embodiment, the system converts the grayscale matrix of the thermal infrared image into a temperature matrix, and then extracts thermal infrared features reflecting the distribution range and temperature gradient of canopy temperature anomalies using spatial statistical methods. It should be understood that temperature distribution analysis is not limited to global statistics; it can also focus on extracting temperature differences in local anomalous areas to more sensitively reflect the obstructed transpiration and abnormal metabolic heat production caused by disease.

[0078] It should be noted that thermal infrared features reflect the degree of anomalousness in the thermal distribution of crop canopy. When disease causes local temperature to rise and forms a significant temperature difference with the surrounding healthy tissue, the dispersion of thermal infrared features increases sharply. This anomalous distribution in a non-visual dimension cannot be replaced by single visible light analysis.

[0079] As an example, the system uses temperature histogram statistics to calculate the variance and kurtosis of the temperature distribution in the whole map or a local area, in order to quantify the dispersion trend of the overall thermal distribution.

[0080] Based on the above steps, the system extracts thermal infrared features that reflect the degree of thermal distribution anomaly through a multi-dimensional temperature distribution analysis method, introducing a cross-modal physiological anomaly dimension index for complexity assessment.

[0081] S304. Use texture features, edge features, and thermal infrared features as parameters for evaluating data complexity.

[0082] Among them, the data complexity evaluation parameter refers to the multi-dimensional feature set that constitutes the basis for calculating the final data complexity index and covers visual structure and physiological abnormalities.

[0083] In this embodiment, the system encapsulates the extracted texture features, edge features, and thermal infrared features in a structured manner to form a multi-dimensional set of data complexity evaluation parameters. This set will be directly input into the subsequent quantization calculation module for fusion mapping.

[0084] It should be noted that multimodal features construct a multi-dimensional complexity evaluation system, avoiding the one-sidedness of a single visual feature. In the early stage of disease, visible light texture and edges may not be obvious, but thermal infrared anomalies have already increased significantly; in the late stage of disease, the edges of visible light lesions are extremely complex, while the thermal infrared temperature difference may tend to be gradual. Only by combining all three can the true difficulty of feature extraction be fully assessed.

[0085] As an example, the system can concatenate the texture feature vector, edge feature vector, and thermal infrared feature vector into a joint feature matrix as an evaluation parameter.

[0086] Based on the above steps, the system constructs a multi-dimensional complexity evaluation parameter system by fusing texture, edge and thermal infrared features to ensure that the data complexity index can comprehensively and evenly reflect the overall extraction difficulty of lesion microstructure, outline clarity and thermal distribution.

[0087] This application embodiment extracts texture, edge, and thermal infrared features and constructs a multi-dimensional complexity evaluation system to avoid the one-sidedness of a single visual feature, providing a comprehensive and accurate feature-level input for the subsequent quantitative calculation of data complexity indicators.

[0088] Example 4:

[0089] Furthermore, in one possible implementation of the embodiments of this application, Figure 4 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model provided in Embodiment 4 of this application. Figure 3 , combined Figure 2 ,like Figure 4 As shown, the quantitative calculation of the data complexity index at the current moment in S202 can be specifically implemented through the following S401, S402, and S403, which are explained in detail below: S401. Calculate the texture variance value of the texture feature, the edge density value of the edge feature, and the temperature dispersion value of the thermal infrared feature, respectively.

[0090] Among them, texture variance refers to the average of the squares of the deviations of each element in the texture feature vector from its mean; edge density refers to the proportion of effective edge pixels in the edge feature map to the total number of pixels; and temperature dispersion refers to the degree of dispersion of the distribution of each temperature sampling point in the thermal infrared feature vector.

[0091] In this embodiment, the system further compresses qualitative feature vectors or feature maps into scalarized numerical indicators. Specifically, the system calculates the variance of each component within the texture feature vector to quantify the disorder of the lesion's microstructure; a larger variance indicates more chaotic texture details. It also calculates the proportion of pixels in the edge feature map whose gradient magnitude exceeds a preset threshold to quantify the outline clarity; a higher density indicates a more complex boundary between the lesion and the background. Finally, it calculates the standard deviation or range of temperature sampling values ​​within the thermal infrared feature vector to quantify the anomaly of thermal distribution; a larger dispersion indicates more severe local temperature differences in the canopy. It should be understood that this dimensionality reduction mapping from high-dimensional feature vectors to low-dimensional scalar values ​​not only preserves the most crucial discriminative information but also significantly reduces the complexity of subsequent fusion calculations.

[0092] It should be noted that texture variance, edge density, and temperature dispersion values ​​quantify the difficulty of feature extraction from three independent dimensions: microstructure, morphological boundary, and physiological abnormality, respectively. The numerical ranges of the three values ​​may differ significantly. For example, temperature dispersion values ​​may be distributed in a large range due to environmental temperature differences, while edge density values ​​are usually limited by the pixel ratio and are distributed between 0 and 1.

[0093] Based on the above steps, the system transforms qualitative multidimensional feature vectors into quantitative three-dimensional numerical indicators by calculating variance, density, and dispersion values. This intuitively and accurately reflects the specific degree of disorder in the microstructure of lesions, the clarity of their outlines, and the abnormality of their thermal distribution.

[0094] S402. The texture variance value, edge density value and temperature dispersion value are weighted and summed according to the preset weights.

[0095] Among them, the preset weight refers to the coefficients assigned by the system to the texture variance value, edge density value and temperature dispersion value, which reflect the proportion of each dimension's contribution to the overall complexity at the current stage of the disease.

[0096] In this embodiment of the application, the system multiplies the three scalar values ​​by their respective preset weights and then sums them up to output a comprehensive value that integrates visual structure and physiological abnormalities.

[0097] It should be noted that the preset weights are not fixed static parameters, but rather have a strategy space for dynamic adjustment based on the disease development stage and crop type. In the early stages of disease, when visible light texture and edge features are not yet obvious, and thermal infrared temperature anomalies appear before visual features, the system can adaptively increase the weight of temperature dispersion values ​​to enhance sensitivity to early physiological anomalies. In the late stages of disease outbreak, when the visual morphology of lesions is extremely complex, the system can increase the weight of texture variance and edge density values ​​to accurately assess the computational power requirements for visual feature extraction.

[0098] As an example, in the early warning scenario of wheat scab, the preset weights can be configured as texture variance value weight 0.2, edge density value weight 0.2, and temperature dispersion value weight 0.6 to highlight the early indication role of thermal infrared anomalies. As another example, in the scenario of severe disease outbreak, the preset weights can be adjusted to texture variance value weight 0.4, edge density value weight 0.4, and temperature dispersion value weight 0.2 to regress the core assessment of visual morphological complexity.

[0099] Based on the above steps, the system achieves the integration of multi-dimensional quantitative indicators and the focus on dominant characteristics through a weighted summation mechanism with dynamic adjustment space.

[0100] S403. Use the weighted summation result as a data complexity indicator.

[0101] Among them, the data complexity index refers to the final unified value obtained after multi-dimensional feature quantification and weighted fusion, which is used to directly participate in the deep scheduling decision of subsequent target calculation.

[0102] In this embodiment, the system directly assigns the weighted summation result to the data complexity index variable, making it a precise control parameter representing the spatial dimension computing power requirement at the current moment. This index exists in a single continuous numerical form, completely eliminating the ambiguity of qualitative feature descriptions and the redundancy of multi-dimensional vector expressions, and can be directly compared numerically with the scheduling threshold.

[0103] This application's embodiments achieve a closed-loop transformation from qualitative features to quantitative indicators by using parallel numerical calculations of texture variance, edge density, and temperature dispersion values, along with a weighted summation mechanism with dynamic adjustment space. This quantitative path not only intuitively reflects the specific degree of disorder in the microstructure of lesions, the clarity of their outlines, and the abnormality of their thermal distribution, but also focuses on the dominant abnormal features at the current stage through flexible allocation of weights.

[0104] Example 5:

[0105] In one possible implementation of the embodiments of this application, Figure 5 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model, provided in Embodiment 5 of this application. Figure 4 , combined Figure 2 ,like Figure 5 As shown, the above S203 can be specifically implemented through the following S501 to S505, which are explained in detail below: S501. Input multimodal observation data from multiple consecutive historical moments into the time series model in chronological order.

[0106] Among them, multiple consecutive historical moments refer to multiple frames of data slices with a fixed time step span extracted from the time series using a sliding window mechanism.

[0107] In this embodiment, the system does not indiscriminately input all historical data into the model. Instead, it selects multimodal observation data from multiple recent consecutive historical moments as the input sequence using a sliding window mechanism. This mechanism allows the model to focus on the most relevant evolutionary segments in the recent past, while avoiding computational redundancy and gradient vanishing problems caused by excessively long input sequences. It should be understood that the length of the sliding window can be adaptively adjusted according to the incubation and outbreak characteristics of different diseases, and is not a fixed constant.

[0108] It should be noted that the sliding window input mechanism preserves the strict chronological order of data in the time dimension, which is a necessary structural prerequisite for time series models to capture dynamic evolution patterns.

[0109] As an example, the system can set the sliding window length to 4, that is, select multimodal observation data of the past 4 consecutive time steps as input to reflect the continuous changing trend of the disease in the short term.

[0110] Based on the above steps, the system achieves effective filtering and structured input of historical data through a sliding window mechanism, ensuring that the time series model can extract evolutionary features based on the most timely continuous segments.

[0111] S502. Use a time series model to extract the state transition features between different time points.

[0112] Among them, state transition characteristics refer to the hidden layer vectors that characterize the microscopic changes in spatial morphology and physiological indicators of diseases at different time steps. Their physical meaning corresponds to biophysical mechanisms such as the rate of lesion expansion, the magnitude of abnormal temperature rise, or the rate of chlorophyll decline.

[0113] In this embodiment, the time-series model uses internal memory units and gating mechanisms to perform differential and correlation calculations on multimodal observation data at adjacent time points, extracting transition features that reflect the transition of disease states. These transition features are no longer limited to static representations at a single time point, but capture the microscopic mechanism of lesions expanding from tiny yellowing spots to large-area necrotic areas, as well as the dynamic process of canopy temperature evolving from localized abnormal heating to global thermal imbalance.

[0114] It should be noted that the extraction of state transition features is the core of time series prediction. It transforms discrete static observations into continuous dynamic evolution descriptions, enabling the system to perceive the acceleration and directionality of disease development.

[0115] As an example, the temporal model uses a convolutional long short-term memory network to extract spatial features at each time step through convolution operations and extract state transition features between adjacent time steps through a long short-term memory gating mechanism.

[0116] Based on the above steps, the system accurately captures the microscopic mechanisms of lesion expansion speed and abnormal changes in physiological indicators by extracting state transition features, providing dynamic evolutionary feature support for subsequent sequence generation and trend prediction.

[0117] S503. Generate disease state sequence based on state transition features.

[0118] Among them, the disease state sequence refers to the set of spatiotemporal evolution vectors output by the time series model at each time step, which integrates historical memory and current transition features.

[0119] In this embodiment, the system fuses and updates the extracted state transition features between different time points with the historical memory state, and sequentially outputs the disease state vector corresponding to each historical time point, thereby forming a complete disease state sequence. This sequence not only contains the static disease representation at each time point, but also implicitly contains the dynamic transition logic between time points, providing a comprehensive characterization of the disease evolution process.

[0120] It should be noted that the generation of disease state sequences is a process of mapping low-dimensional observation data to a high-dimensional hidden state space, which amplifies and makes explicit the subtle deterioration trends that are difficult to perceive intuitively in the original data in the sequence vector.

[0121] As an example, the generated disease state sequence can be represented as a feature matrix, where each row of the matrix corresponds to a hidden state vector at a historical moment, and the vector dimension reflects the richness of the spatiotemporal evolution of the disease.

[0122] Based on the above steps, the system transforms discrete historical observation data into continuous spatiotemporal evolution feature expressions, enabling the dynamic evolution law of the disease to be fully serialized and characterized in the latent state space.

[0123] S504. Predict the disease development level at multiple future moments based on the disease state sequence.

[0124] Among them, the disease development level refers to the index that quantifies and classifies the severity of the disease, such as discrete levels or continuous severity scores such as mild, moderate and severe.

[0125] In this embodiment, the system inputs the generated disease state sequence into the prediction head network, and maps the disease development level corresponding to multiple different time steps in the future through regression or classification operations. It should be understood that the prediction time step can be flexibly set according to the lead time requirements of prevention and control decisions, covering short-term early warning and medium- and long-term trend inference.

[0126] It should be noted that multi-step prediction based on state sequences enables the system to output a complete future evolution trajectory, rather than a single future state point, providing sufficient information for assessing the acceleration and persistence of the deterioration trend.

[0127] Based on the above steps, the system realizes multi-step prediction from historical sequence to future trajectory, providing a forward-looking multi-timescale level prediction for assessing the trend of disease deterioration.

[0128] S505. Calculate the predicted deterioration probability based on the change range of the disease development level, and use the predicted deterioration probability as the predicted risk level.

[0129] Among them, the predicted deterioration probability refers to the probability value that quantifies the risk of disease outbreak, calculated based on the increasing rate and accelerating trend of disease development levels at multiple future times.

[0130] In this embodiment, the system calculates the predicted changes in disease development levels relative to the current level at multiple future moments and converts these changes into a predicted probability of deterioration. This conversion logic is not limited to simple linear mapping; it can also employ nonlinear functions such as exponential mapping to better reflect the objective law of exponential growth in disease deterioration during the outbreak phase. For example, when the change is small, linear mapping can sensitively capture early, subtle trends; when the change is large, exponential mapping can drastically amplify high-risk signals, ensuring the system triggers the deep calculation path in a timely manner. Finally, the system directly uses this predicted probability of deterioration as the predicted risk level for subsequent deep scheduling decisions.

[0131] Preferably, the risk level satisfies the following formula:

[0132] in, This indicates the disease development level at the current time t. This represents the disease development level corresponding to the k-th time step in the future, where K represents the length of the predicted time step used for risk assessment. This represents the weighting coefficients corresponding to different future time steps; The adjustment coefficient representing the acceleration term of disease development; This represents the complexity adjustment factor; This represents the data complexity index calculated based on the fusion of texture features, edge features, and thermal infrared features at the current moment. This represents the nonlinear adjustment coefficient; This represents a complexity threshold, used to distinguish between low-complexity and high-complexity scenarios. This represents the time corresponding to the s-th time step in the time series.

[0133] It should be noted that the system calculates the final predicted risk level by comprehensively considering the cumulative difference between the disease development level at multiple future time steps and the current time, the accelerating deterioration trend of disease development, and the nonlinear coupling effect of the current data complexity index. Specifically, by accumulating the difference between the level at each future time step and the current level, the system reflects the overall deterioration of the disease in the time dimension; by calculating the cumulative change in level at adjacent future time steps, the system captures the accelerating trend of disease deterioration, i.e., whether the disease is rapidly accelerating its outbreak; and by introducing the current spatial dimension data complexity index in a nonlinear step manner, when the complexity exceeds a certain threshold, it has a sharp amplification effect on the risk level, ensuring that the system can sensitively improve the warning level in complex scenarios where feature extraction is extremely difficult.

[0134] Based on the above steps, the system uses quantitative logic to transform the magnitude of changes in disease development levels into the probability of predicted deterioration, achieving a leap from qualitative trend perception to quantitative risk control, and providing direct and precise control variables for reverse-driven network structure adjustment.

[0135] This application embodiment uses a sliding window input mechanism and state transition feature extraction to accurately capture the microscopic mechanisms of lesion expansion speed and abnormal changes in physiological indicators. By converting the magnitude of change into a quantitative logic of deterioration probability, the qualitative deterioration trend is transformed into a quantitative predicted risk level, providing a forward-looking and numerical control variable for subsequent reverse-driven deep network scheduling.

[0136] Example 6:

[0137] In one possible implementation of the embodiments of this application, Figure 6 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model provided in Embodiment Six of this application. Figure 5 , combined Figure 2 ,like Figure 6As shown, the above S205 can be implemented by the following S601 to S605, which are explained in detail below: S601. Normalize the predicted risk level to obtain a risk score.

[0138] Normalization refers to the mathematical operation of mapping raw data with different dimensions or numerical ranges to a unified dimensionless interval.

[0139] In the embodiments of this application, the system transforms the absolute value of the predicted risk level, which reflects the deterioration trend over time, into a standardized range of 0 to 1 through linear scaling or nonlinear mapping, thereby eliminating the difference between its original physical dimensions and numerical range.

[0140] It should be noted that normalization is not limited to Min-Max scaling; other mathematical transformations such as Z-score standardization or Sigmoid function compression can also be used. Eliminating dimensional differences is a necessary prerequisite for joint decision-making based on multi-dimensional indicators. If unnormalized absolute risk values ​​are directly mixed with complexity values ​​in calculations, the indicator with the larger numerical range will dominate the decision result, thus masking the true impact of the other dimension.

[0141] As an example, when the original values ​​of the predicted risk level are distributed in the integer range of 0 to 100, the system can linearly normalize them to a risk score between 0 and 1 by dividing by the maximum value boundary of 100.

[0142] S602. Normalize the data complexity index to obtain the complexity score.

[0143] The complexity score refers to the dimensionless quantized value of the difficulty of extracting spatial dimension image features at the current moment after standardization.

[0144] In this embodiment, the system also maps the data complexity index to the same 0-1 range as the risk score using a preset normalization function, ensuring that the two are aligned on a numerical scale. This alignment operation makes the subsequent weighted fusion calculation mathematically rigorous, avoiding decision bias caused by scale imbalance.

[0145] S603. The risk score and complexity score are weighted and summed according to the preset weights to obtain the comprehensive scheduling score.

[0146] Among them, the preset weights refer to the coefficients pre-configured by the system to adjust the proportion of different dimension scores in the final decision. The comprehensive scheduling score refers to the final control variable that integrates time risk and space complexity.

[0147] In this embodiment, the system multiplies the risk score and complexity score by their respective preset weights and then adds them together to output a comprehensive scheduling score.

[0148] It should be noted that the preset weights are not fixed constants, but have room for dynamic adjustment. The system can adaptively adjust the weight allocation strategy according to different crop growth stages or disease types. For example, in the early warning stage of diseases, since the visible light characteristics are not obvious, the system can increase the weight of the risk score to enhance the sensitivity to the trend of temporal deterioration; while during the outbreak of diseases, the system can increase the weight of the complexity score to ensure sufficient computing power allocation in complex scenarios.

[0149] It should be noted that a comprehensive scheduling score is used instead of a single indicator for decision-making because neither time risk nor space complexity can fully reflect the actual computing power requirements of the current scenario. Low-risk but high-complexity background interference scenarios, as well as high-risk but low-complexity early deterioration scenarios, both require comprehensive scoring to accurately trigger corresponding computational depth adjustments.

[0150] Based on the above steps, the system achieves quantitative unification of time dimension risk and spatial dimension complexity through a weighted summation mechanism, so that the final comprehensive scheduling score can reflect the comprehensive computing power requirements of disease dynamic evolution and current scene feature extraction in a balanced and accurate manner.

[0151] S604. Compare the comprehensive scheduling score with the scheduling thresholds corresponding to multiple computation depths.

[0152] The scheduling threshold refers to the critical value used to divide the trigger intervals of different computing depths.

[0153] In this embodiment of the application, the system pre-configures a set of scheduling thresholds corresponding to multiple calculation depth levels, and compares the calculated comprehensive scheduling score with these thresholds one by one to determine which threshold range it falls into.

[0154] It should be noted that, through the threshold comparison mechanism, the system discretizes the continuous comprehensive scheduling score into specific computation depth levels. This discretization not only simplifies the subsequent network branch selection logic, but also avoids frequent switching of network branches and computing power jitter caused by small fluctuations in the score.

[0155] As an example, the system can set three computing depths: lightweight, standard, and deep, and configure two scheduling thresholds accordingly: a first scheduling threshold of 0.3 and a second scheduling threshold of 0.7. When the overall scheduling score is below 0.3, it is mapped to the lightweight level; when the score is between 0.3 and 0.7, it is mapped to the standard level; and when the score is above 0.7, it is mapped to the deep level.

[0156] Based on the above steps, the system establishes a clear mapping relationship between the comprehensive scheduling score and multiple computation depths, providing a clear and stable decision boundary for determining the final target computation depth.

[0157] S605. Determine the target calculation depth based on the comparison results.

[0158] The target computation depth refers to the computation level or resource consumption corresponding to the neural network inference path that the system decides to invoke at the current moment.

[0159] In this embodiment, the system directly determines the computation depth level corresponding to the interval into which the comprehensive scheduling score falls as the target computation depth based on the interval determination result. This determination process is the final execution link of the numerical scheduling chain, directly converting the aforementioned quantitative evaluation results into network structure reconstruction instructions.

[0160] It should be noted that the determination of the target computing depth relies entirely on the quantitative drive of the comprehensive scheduling score, rather than on static rules set by humans. This enables the system to achieve adaptive and precise allocation of computing power without human intervention.

[0161] As an example, if the overall scheduling score falls into the depth range after comparison, the system determines the target computation depth to be at the depth level and will activate the depth path branch containing more convolutional layers and attention modules; if it falls into the lightweight range, the system determines the target computation depth to be at the lightweight level and will activate the lightweight path branch with the fewest parameters.

[0162] This application's embodiments eliminate the dimensional differences between time-dimension risk and spatial-dimension complexity through a normalized and weighted summation-based numerical scheduling mechanism, and achieve quantitative unification of the two through comprehensive scheduling scoring. This comprehensive scoring decision-making mechanism overcomes the one-sidedness of single-index evaluation, enabling the system to accurately balance the evolution trend and the comprehensive computing power requirements of the current complexity, and stably determine the target computational depth through threshold mapping, providing a reliable numerical basis for the dynamic selection of subsequent network branches.

[0163] Example 7:

[0164] In one possible implementation of the embodiments of this application, Figure 7 A flowchart illustrating a pest and disease image recognition method based on a lightweight neural network model provided in Embodiment 7 of this application. Figure 6 , combined Figure 2 ,like Figure 7 As shown, prior to S207 above, the pest and disease image recognition method based on the lightweight neural network model also includes the following implementations S701 to S705, which are described in detail below: S701. Extract thermal anomalies from thermal infrared images, extract near-infrared reflection anomalies from near-infrared images, and extract color anomalies from visible light images.

[0165] Among them, thermal anomaly values ​​refer to quantitative values ​​that reflect the local temperature of the crop canopy deviating from the normal range, near-infrared reflectance anomaly values ​​refer to quantitative values ​​that reflect the chlorophyll content and water status of vegetation deviating from normal physiological indicators, and color anomaly values ​​refer to quantitative values ​​that reflect the color characteristics of lesion areas deviating from the color of healthy tissue.

[0166] In this embodiment, the system is based on the principles of plant pathology, namely that disease occurrence is usually accompanied by a decrease in chlorophyll, disordered cell metabolism, and obstructed water transport, and extracts corresponding outliers from three modalities of images. The system uses the difference method to extract outliers, that is, to calculate the absolute difference between the current pixel value and the historical healthy mean of the same area or the global healthy mean; alternatively, it can use the ratio method to extract outliers, that is, to calculate the ratio of the current pixel value to a reference baseline value to reflect the relative rate of change.

[0167] It should be noted that extracting three outliers from multimodal data is based on the objective law that diseases exhibit different dominant physiological abnormalities at different stages of development. Anomalies in a single modality often cannot fully characterize the comprehensive risk of early latent diseases.

[0168] As an example, for thermal infrared images, the system calculates the difference between the average temperature of the suspected lesion area and the average temperature of the surrounding healthy background area as a thermal anomaly value; for near-infrared images, the system calculates the ratio of the current reflectance to the standard healthy reflectance benchmark as a near-infrared reflectance anomaly value.

[0169] Based on the above steps, the system extracts multimodal outliers and transforms discrete image pixel information into quantitative indicators that reflect deviations in the multidimensional physiological state of crops, providing accurate numerical input for subsequent comprehensive anomaly assessment based on biophysical mechanisms.

[0170] S702. Normalize the thermal anomaly values, near-infrared reflection anomaly values, and color anomaly values ​​respectively. Then, perform a weighted summation on the normalized thermal anomaly values, near-infrared reflection anomaly values, and color anomaly values ​​according to their corresponding weights. Use the weighted summation result as the disease anomaly index.

[0171] Normalization refers to the mathematical operation of mapping outliers with different physical dimensions and numerical ranges to a unified dimensionless interval. The corresponding weight is a coefficient assigned by the system to outliers of different modes, reflecting the proportion of contribution of each physiological abnormality to the comprehensive diagnosis at the current stage of the disease.

[0172] In this embodiment, the system first maps the three extracted absolute outliers to a standardized range of 0 to 1 through linear scaling or nonlinear compression, eliminating the dimensional and scale differences between temperature, reflectivity, and color grayscale. The system then weights and sums the three normalized outliers according to preset corresponding weights, outputting a comprehensive disease anomaly index.

[0173] Based on the above steps, the system transforms multimodal physiological abnormalities into a comprehensive index with unified dimensions through a weighted summation mechanism of normalization and dynamic weights, thereby improving the detection sensitivity of early latent diseases and the adaptability of diagnosis at different stages of development.

[0174] S703. Generate an abnormal distribution map based on the disease abnormality index.

[0175] Among them, the abnormal distribution map refers to a heat map or grayscale map that intuitively reflects the spatial distribution of abnormal diseases after mapping the disease abnormality index of each pixel or local area in the image into a two-dimensional spatial matrix.

[0176] In this embodiment, the system uses a sliding window to traverse the spatial dimensions of multimodal observation data, calculates the disease anomaly index pixel by pixel or block by block, and arranges these index values ​​according to their original spatial coordinates to generate an anomaly distribution map corresponding to the original image size. This distribution map not only contains information on the numerical magnitude of the anomalies but also preserves the spatial clustering locations and topological relationships of the anomalies, providing an intuitive visual and data foundation for subsequent regional positioning.

[0177] It should be noted that the method of generating the anomaly distribution map integrates the anomaly information that was originally scattered in three independent modal images into a unified map, making the cross-modal anomaly overlapping areas visible and greatly simplifying the search complexity of subsequent candidate regions.

[0178] As an example, the system directly maps the values ​​of the disease anomaly index to the color gradient of the heat map. The higher the index, the more prominent the color of the area, thus generating an intuitive heat map of the anomaly distribution.

[0179] Based on the above steps, the system generates anomaly distribution maps, which condense multidimensional discrete anomaly indices into a continuous visual representation with spatial topological structure, providing a clear search space and boundary basis for the extraction of high-confidence regions.

[0180] S704. Extract connected regions from the anomaly distribution map where the anomaly index is greater than a preset threshold, and identify the connected regions as high-confidence candidate regions.

[0181] Among them, the preset threshold is a critical value used to distinguish between valid disease anomalies and random background noise; the connected region is a set of pixels in the anomaly distribution map that meet the threshold condition and are spatially adjacent to each other; and the high-confidence candidate region is a local image region that is determined to have a high probability of disease based on biophysical mechanisms and is worth investing in depth computing resources.

[0182] In this embodiment, the system first performs thresholding on the anomaly distribution map, marking pixels with an anomaly index greater than a preset threshold as foreground and the rest as background. The system then extracts spatially connected pixel sets from the binarized foreground image to form initial connected regions. To eliminate fragmented noise points and fill internal voids, the system applies morphological operations to the binarized image. Specifically, the system uses erosion to remove isolated small noise points and dilation to merge adjacent real lesion regions and fill internal breaks, thereby obtaining connected regions with smooth boundaries and complete structures. Finally, the system identifies these morphologically optimized connected regions as high-confidence candidate regions.

[0183] It should be noted that extracting high-confidence candidate regions based on thresholds and connectivity essentially utilizes the biophysical principle that diseases have localized spatial clustering, effectively distinguishing between systemic environmental disturbances and localized disease lesions.

[0184] As an example, the system can set a preset threshold of 0.6, retain only the core regions with high anomaly indices, and perform morphological opening and closing operations using 3x3 erosion kernels and 5x5 expansion kernels in sequence to extract high-confidence candidate regions with clear boundaries.

[0185] Based on the above steps, the system accurately locates local spatial positions with a high probability of disease by threshold segmentation and morphological connected region extraction, effectively filtering non-disease background noise and significantly reducing the number of invalid pixels that subsequent network branches need to process.

[0186] S705. Based on the disease anomaly index, the multimodal observation data is weighted and processed, and the target network branch is used to extract features and identify diseases and pests in high-confidence candidate regions.

[0187] Among them, weighted processing refers to the data preprocessing operation that uses the disease anomaly index as a weight coefficient or mask to enhance the signal of the abnormal and significant region in the multimodal observation data and suppress the signal of the background region. The cropping input mechanism refers to the computational optimization strategy of only inputting the local image slices corresponding to the high confidence candidate regions into the target network branch instead of the whole image input.

[0188] In this embodiment, before sending the data into the target network branch, the system first weights and enhances the multimodal observation data at the current moment based on the disease anomaly index, thereby improving the signal-to-noise ratio of physiologically abnormal regions highly correlated with disease occurrence at the data level. Based on the boundary coordinates of high-confidence candidate regions, the system crops corresponding local image slices from the weighted multimodal observation data, using only these slices as input data for feature extraction and disease / pest identification in the target network branch.

[0189] As an example, the system can crop three corresponding local slices from the original visible light, thermal infrared and near-infrared images based on the bounding rectangle of the high-confidence candidate region, and then stitch or fuse these slices and input them into the lightweight path or deep path of the target network branch for fine recognition.

[0190] This application embodiment extracts multimodal outliers based on biophysical mechanisms and calculates disease anomaly indices, generates anomaly distribution maps and extracts high-confidence candidate regions. Finally, through weighted processing and input pruning mechanisms, it achieves accurate region positioning and data augmentation, effectively filters background noise, reduces more than 70% of invalid computation, and provides a high signal-to-noise ratio data foundation for the detection of early latent diseases.

Claims

1. A method for identifying pest and disease images based on a lightweight neural network model, characterized in that, include: Acquire multimodal observation data of the target crop at the current moment and multimodal observation data at historical moments, wherein the multimodal observation data includes visible light images, thermal infrared images and near-infrared images; Based on the multimodal observation data at the current moment, extract texture features, edge features, and thermal infrared features, and calculate the data complexity index at the current moment based on the texture features, edge features, and thermal infrared features; The multimodal observation data at the historical moments are input into a time series model in chronological order for processing to obtain a disease state sequence; Based on the disease state sequence, predict the disease development level corresponding to multiple future moments, and calculate the predicted risk level based on the disease development level corresponding to the multiple future moments. The target computation depth is determined based on the predicted risk level and the data complexity index; Based on the target computation depth, the target network branch is determined from multiple feature extraction network branches with different computation depths; The target network branch is used to extract features from the multimodal observation data at the current moment to obtain disease characterization features; Disease and pest identification is performed based on the aforementioned disease symptom characteristics.

2. The method for pest and disease image recognition based on a lightweight neural network model according to claim 1, characterized in that, The step of determining the target computation depth based on the predicted risk level and the data complexity index includes: The predicted risk level is normalized to obtain a risk score; The data complexity index is normalized to obtain a complexity score; The risk score and the complexity score are weighted and summed according to preset weights to obtain a comprehensive scheduling score; The comprehensive scheduling score is compared with scheduling thresholds corresponding to multiple computation depths; The target computation depth is determined based on the comparison results.

3. The method for pest and disease image recognition based on a lightweight neural network model according to claim 1, characterized in that, The multimodal observation data at the historical moments are input into a time-series model in chronological order for processing to obtain a disease state sequence, including: The multimodal observation data from multiple consecutive historical moments are input into the time series model in chronological order; The state transition features between each time step are extracted using the aforementioned time series model; Generate a disease state sequence based on the aforementioned state transition features; Predict the disease development level at multiple future moments based on the disease state sequence; The predicted deterioration probability is calculated based on the change range of the disease development level, and the predicted deterioration probability is used as the predicted risk level.

4. The method for pest and disease image recognition based on a lightweight neural network model according to claim 3, characterized in that, The risk level satisfies the following formula: ； in, This indicates the disease development level at the current time t. This represents the disease development level corresponding to the k-th time step in the future, where K represents the length of the predicted time step used for risk assessment. This represents the weighting coefficients corresponding to different future time steps; The adjustment coefficient representing the acceleration term of disease development; This represents the complexity adjustment factor; This represents the data complexity index calculated based on the fusion of texture features, edge features, and thermal infrared features at the current moment. This represents the nonlinear adjustment coefficient; This represents a complexity threshold, used to distinguish between low-complexity and high-complexity scenarios. This represents the time corresponding to the s-th time step in the time series.

5. The method for pest and disease image recognition based on a lightweight neural network model according to claim 1, characterized in that, The data complexity index for calculating the current moment includes: The visible light image is processed using a texture analysis operator to obtain texture features; The visible light image is processed using an edge detection operator to obtain edge features; Temperature distribution analysis is performed on the thermal infrared image to obtain thermal infrared features; The texture features, edge features, and thermal infrared features are used as parameters for evaluating data complexity.

6. The method for pest and disease image recognition based on a lightweight neural network model according to claim 5, characterized in that, The calculation of the data complexity index at the current moment also includes: Calculate the texture variance value of the texture feature, the edge density value of the edge feature, and the temperature dispersion value of the thermal infrared feature, respectively. The texture variance value, edge density value, and temperature dispersion value are weighted and summed according to preset weights. The weighted summation result is used as an indicator of data complexity.

7. The method for pest and disease image recognition based on a lightweight neural network model according to claim 1, characterized in that, Before using the target network branch for feature extraction, the method further includes: Extract thermal anomaly values ​​from the thermal infrared image; Extract near-infrared reflectance anomalies from the near-infrared image; Extract color anomalies from the visible light image; The disease anomaly index is calculated based on the aforementioned thermal anomaly value, near-infrared reflectance anomaly value, and color anomaly value. The multimodal observation data are weighted according to the disease anomaly index.

8. The method for pest and disease image recognition based on a lightweight neural network model according to claim 7, characterized in that, The calculation of the disease anomaly index based on the thermal anomaly value, near-infrared reflectance anomaly value, and color anomaly value includes: The thermal anomaly values, near-infrared reflectance anomaly values, and color anomaly values ​​are normalized respectively; The normalized thermal anomalies, near-infrared reflectance anomalies, and color anomalies are weighted and summed according to their respective weights. The weighted summation result is used as the disease abnormality index.

9. The method for pest and disease image recognition based on a lightweight neural network model according to claim 7, characterized in that, Also includes: An anomaly distribution map is generated based on the aforementioned disease anomaly index; Extract connected regions from the abnormal distribution map where the disease abnormality index is greater than a preset threshold. The connected regions are identified as high-confidence candidate regions; The target network branch is used to extract features and identify pests and diseases in the high-confidence candidate region.

10. A pest and disease image recognition system based on a lightweight neural network model, characterized in that, include: The data acquisition module is used to acquire multimodal observation data of the target crop at the current moment and multimodal observation data at historical moments; The complexity calculation module is used to calculate the data complexity index based on the multimodal observation data at the current moment; The time-series prediction module is used to generate a disease status sequence based on the multimodal observation data of the historical time and predict the predicted risk level; A deep scheduling module is used to determine the target computation depth based on the predicted risk level and the data complexity index; The network selection module is used to determine the target network branch based on the target computation depth; The identification module is used to identify pests and diseases using the target network branch.

Images