Computer vision large model processing method based on knowledge graph

Abstract

The invention discloses a computer vision large model processing method based on a knowledge graph, and relates to the technical field of image processing, and the method comprises the following steps: constructing an initial causal knowledge graph; reasoning a dynamic diagnosis confidence coefficient; a diagnostic feedback closed-loop adaptive calibration is performed. Disease and environment features are separated through a spatial-temporal feature decoupling network, precise diagnosis is realized in combination with a causal knowledge graph, a closed-loop calibration mechanism is utilized to dynamically optimize the model according to expert feedback, and an optimal intervention strategy for comprehensive treatment effect, cost and risk can be generated based on the knowledge graph. The problems that in the prior art, feature decoupling is insufficient, a model cannot be adaptively optimized, and intervention decision scientificity is insufficient are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for processing large computer vision models based on knowledge graphs. Background Technology

[0002] In agricultural production, timely diagnosis and precise control of crop diseases are crucial. With the development of knowledge graph and computer vision technologies, crop disease diagnosis through image recognition combined with knowledge reasoning has become an important direction. By extracting crop image features and combining them with environmental data to construct knowledge models, intelligent analysis and prediction of diseases can be achieved.

[0003] Existing crop image processing methods struggle to effectively distinguish between disease-induced tissue texture evolution and environmental stress-induced morphological changes in crop disease diagnosis due to insufficient feature decoupling capabilities. Furthermore, diagnostic models lack adaptive calibration mechanisms, employ simplistic intervention strategies that are difficult to continuously optimize based on expert feedback, and the generation of intervention strategies does not adequately incorporate multi-dimensional factors, thus requiring improvement in the scientific rigor of decision-making. Summary of the Invention

[0004] The purpose of this invention is to provide a method for processing large computer vision models based on knowledge graphs, which solves the problems existing in the background technology.

[0005] To address the aforementioned technical problems, this invention provides a method for processing large computer vision models based on knowledge graphs, comprising the following steps: Step 1: Construct an initial causal knowledge graph; The historical dataset is obtained, which includes historical crop image sequences, historical environmental state sequences, and historical expert diagnostic annotations. The historical dataset is then processed using a causal inference engine to construct an initial causal knowledge graph. Step two: Reasoning for dynamic diagnostic confidence; Acquire real-time crop image sequences; process the real-time crop image sequences using a spatiotemporal feature decoupling network, which is controlled by trainable model parameters; process the visual primitives representing disease symptoms and the visual primitives representing environmental responses used to decouple the output; use the diagnostic fusion reasoning module to perform comprehensive reasoning on the visual primitives representing disease symptoms and the visual primitives representing environmental responses, combining the initial causal knowledge graph and real-time environmental data, and output a diagnostic confidence vector; Step 3: Perform diagnostic feedback closed-loop adaptive calibration; Obtain the expert's final diagnosis result corresponding to the diagnosis confidence vector; use a closed-loop calibration mechanism to calculate the deviation between the expert's final diagnosis result and the diagnosis confidence vector; calculate the model parameters used to synchronously update the initial causal knowledge graph and the spatiotemporal feature decoupling network to generate the updated knowledge graph and updated model parameters for subsequent diagnostic inference.

[0006] Preferably, the construction of the initial causal knowledge graph in step one further includes: Historical crop image sequences, historical environmental state sequences, and historical expert diagnostic annotations were collected and structured to form a historical dataset. The initial causal knowledge graph is a directed acyclic graph, which contains nodes and weighted directed edges. Nodes are used to represent known disease types, visual symptoms, stress responses, and environmental factors, while weighted directed edges are used to quantify the strength of causal relationships between nodes.

[0007] Preferably, a multi-scale temporal attention mechanism is integrated within the spatiotemporal feature decoupling network; The multi-scale temporal attention mechanism is used to analyze the temporal dimension features of real-time crop image sequences and separate segments with different frequencies of change in the image sequences to distinguish between low-frequency stable tissue texture evolution caused by disease and high-frequency transient morphological changes caused by microenvironmental stress.

[0008] Preferably, the reasoning process of the diagnostic fusion reasoning module further includes: Execute the confidence propagation algorithm on the updated knowledge graph; The visual primitives of environmental response are correlated with real-time environmental data to quantify and isolate the apparent visual features caused by non-pathological environmental stress. Based on the stripping results, when analyzing the visual primitives of symptoms, a net confidence score is generated for each possible disease. The net confidence scores together constitute the diagnostic confidence vector.

[0009] Preferably, it also includes: simulating the evolution trend of the disease; Identify the highest probability disease type from the diagnostic confidence vector; Using a prospective disease evolution simulator, with symptom visual primitives as the initial state and causal subgraphs related to the highest probability disease type in the updated knowledge graph as the evolution model, we simulate and output the future symptom evolution trajectory.

[0010] Preferably, the prospective disease evolution simulator further acquires data on future environmental conditions and the duration of segment predictions of future disease progression during simulation. The disease evolution trend simulation is based on future environmental condition prediction data. Within the predicted time frame of the future disease, a time-series-based stochastic process is performed on the causal subgraph.

[0011] Preferably, it also includes: generating the optimal precision intervention strategy; The diagnostic result determined from the diagnostic confidence vector; Based on the confirmed diagnostic results, the intervention decision optimization engine is invoked to retrieve all relevant candidate interventions from the updated knowledge graph, and the optimal intervention plan is recommended from the candidate interventions through quantitative evaluation.

[0012] Preferably, the quantitative evaluation process of the intervention decision optimization engine further includes: For candidate interventions, perform counterfactual simulations to predict intervention effects; Counterfactual simulation involves simulating the future environment after the intervention and, based on the future environment after the intervention, recalculating the evolution trajectory of symptoms after the intervention. The severity of symptoms is assessed using a symptom severity assessment function. The degree of damage in the future symptom evolution trajectory and the post-intervention symptom evolution trajectory is quantified, and the difference between the two is defined as the expected therapeutic effect of the candidate intervention.

[0013] Preferably, the intervention decision optimization engine further includes the following before recommending the optimal intervention plan: Calculate the overall utility value for candidate interventions; The overall utility value is generated through a utility function, which integrates expected treatment effects, implementation costs, and potential environmental risks, and is balanced by weighting coefficients. The candidate intervention with the highest overall utility value is identified as the optimal intervention plan.

[0014] Preferably, the closed-loop calibration mechanism further obtains a preset learning rate during calculation; The closed-loop calibration mechanism backpropagates the bias based on the learning rate, which is used to synchronously fine-tune the weights of the weighted directed edges in the initial causal knowledge graph and the trainable model parameters of the spatiotemporal feature decoupling network, so as to achieve adaptive iteration of the model.

[0015] Beneficial effects Compared with the prior art, the present invention has the following beneficial effects: 1. By using a spatiotemporal feature decoupling network to separate disease and environmental features, the low-frequency stable tissue texture evolution caused by disease can be distinguished from the high-frequency transient morphological changes caused by environmental stress, thus avoiding interference from environmental factors in diagnosis. Combined with causal knowledge graph reasoning, this effectively improves the accuracy of crop disease diagnosis.

[0016] 2. Through a closed-loop calibration mechanism, the model parameters and graphs are dynamically updated based on expert feedback, enabling the model to continuously learn and optimize in practical applications, thereby avoiding model rigidity and improving the model's adaptability to dynamic scenarios.

[0017] 3. By generating optimal intervention strategies that comprehensively consider treatment effects, costs, and risks through knowledge graphs, and combining utility functions to integrate multi-dimensional factors to calculate comprehensive utility values, an optimal intervention plan is generated, realizing an intelligent closed loop from diagnosis to decision-making, providing scientific decision support, and improving the scientific nature of intervention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] Example 1: Please see Figure 1 This invention provides a method for processing large computer vision models based on knowledge graphs, comprising the following steps: Step 1: Construct an initial causal knowledge graph; The historical dataset is obtained, which includes historical crop image sequences, historical environmental state sequences, and historical expert diagnostic annotations. The historical dataset is then processed using a causal inference engine to construct an initial causal knowledge graph. Step two: Reasoning for dynamic diagnostic confidence; Acquire real-time crop image sequences; process the real-time crop image sequences using a spatiotemporal feature decoupling network, which is controlled by trainable model parameters; process the visual primitives representing disease symptoms and the visual primitives representing environmental responses used to decouple the output; use the diagnostic fusion reasoning module to perform comprehensive reasoning on the visual primitives representing disease symptoms and the visual primitives representing environmental responses, combining the initial causal knowledge graph and real-time environmental data, and output a diagnostic confidence vector; Step 3: Perform diagnostic feedback closed-loop adaptive calibration; Obtain the expert's final diagnosis result corresponding to the diagnosis confidence vector; use a closed-loop calibration mechanism to calculate the deviation between the expert's final diagnosis result and the diagnosis confidence vector; calculate the model parameters used to synchronously update the initial causal knowledge graph and the spatiotemporal feature decoupling network to generate the updated knowledge graph and updated model parameters for subsequent diagnostic inference. In this embodiment, a computer vision large model processing method based on knowledge graph is implemented. The method first processes the real-time acquired crop leaf image time series through a spatiotemporal feature decoupling network controlled by trainable model parameters. The core function of the spatiotemporal feature decoupling network is to achieve the decoupling of visual features. It decomposes the input image sequence into stable texture primitives representing diseases and instantaneous morphological primitives representing environmental responses. This process is expressed by the following formula: ; Stable texture primitives representing diseases, or simply symptom primitives; The instantaneous form of the environmental response is referred to as the environmental element; This represents a spatiotemporal feature decoupling network used to separate components of different frequencies in image features; Represents a time series of crop leaf images; Represents the trainable model parameters of a spatiotemporal feature decoupling network; This method of decoupling features provides high-quality input for subsequent accurate diagnosis, effectively distinguishing between fundamental tissue changes caused by disease and temporary apparent changes caused by environmental stress, thus laying the foundation for improving diagnostic accuracy.

[0021] Example 2: Step one, constructing the initial causal knowledge graph, further includes: Historical crop image sequences, historical environmental state sequences, and historical expert diagnostic annotations were collected and structured to form a historical dataset. The initial causal knowledge graph is a directed acyclic graph, which contains nodes and weighted directed edges. Nodes are used to represent known disease types, visual symptoms, stress responses, and environmental factors, while weighted directed edges are used to quantify the strength of causal relationships between nodes. In this embodiment, the construction process of the initial causal knowledge graph is described in detail. Before the system is deployed, the target greenhouse is monitored for a long time, and historical data is collected and structured to form a comprehensive historical dataset. This historical dataset consists of a series of structured data tuples, each of which contains a crop image sequence at a specific time point, a synchronized environmental state sequence, and authoritative expert diagnosis and event annotation. The historical dataset is composed of the following: ; ; This represents a historical dataset, which includes image sequences, environmental data, and expert annotations. Represents a data tuple; It represents an image sequence, which is a sequence of visual changes in crops captured by a fixed-position camera, recording the dynamic process of crops from health to disease, as well as under different environmental stresses; This represents a sequence of environmental conditions, which includes real-time monitoring values ​​of environmental factors such as temperature, humidity, light intensity, and soil nutrients, collected synchronously with the image sequence from sensors in the greenhouse's microenvironment. This indicates that expert diagnoses and event annotations are authoritative annotations formed by agricultural technology experts after retrospective analysis of data, including symptom descriptions at specific time points, the finally diagnosed disease type, known environmental stress events or human intervention measures, etc. Subsequently, a causal inference engine based on a Bayesian network structure learns from this historical dataset to generate an initial causal knowledge graph. This knowledge graph is a directed acyclic graph, in which nodes represent entities such as disease types, visual symptoms, stress responses, and environmental factors, while weighted directed edges quantitatively describe the strength of causal relationships between these entities. ; Represents the initial causal knowledge graph; This represents a causal inference engine based on Bayesian networks; This represents a historical dataset, which includes image sequences, environmental data, and expert annotations. This step, through in-depth analysis of historical data, establishes a prior knowledge base containing domain knowledge, providing a solid logical foundation and reasoning basis for the system's subsequent real-time diagnosis.

[0022] Example 3: The spatiotemporal feature decoupling network integrates a multi-scale temporal attention mechanism. The multi-scale temporal attention mechanism is used to analyze the temporal dimension features of real-time crop image sequences and separate segments with different frequencies of change in the image sequence to distinguish between low-frequency stable tissue texture evolution caused by disease and high-frequency transient morphological changes caused by microenvironmental stress. In this embodiment, the core mechanism of the spatiotemporal feature decoupling network, namely the multi-scale temporal attention mechanism, is specifically described. This mechanism is designed to process real-time crop image sequences and achieves the separation of visual signals of different properties by analyzing their dynamic features in the time dimension. Specifically, the multi-scale temporal attention mechanism can identify and separate signal components with different frequencies in the image sequence. Tissue texture evolution caused by disease, such as the formation and expansion of lesions, is identified as a low-frequency stable signal, while wilting caused by microenvironmental stress, such as instantaneous water imbalance, is identified as a high-frequency transient signal. Through this frequency-based separation, the network can decouple mixed visual information into independent symptom visual primitives and environmental response visual primitives. The integration of this mechanism enables the model to penetrate the appearance and accurately capture the essential features of disease development, avoiding the interference of environmental factors on the diagnostic results and significantly improving the robustness and effectiveness of feature extraction.

[0023] Example 4: The reasoning process of the diagnostic fusion reasoning module further includes: Execute the confidence propagation algorithm on the updated knowledge graph; The correlation analysis between the visual primitives of environmental response and real-time environmental data is used to quantify and isolate the apparent visual features caused by non-pathological environmental stress. Based on the stripping results, when analyzing the visual primitives of symptoms, a net confidence score is generated for N possible diseases, and the N built-in confidence scores together constitute the diagnostic confidence vector. In this embodiment, the workflow of the diagnostic fusion reasoning module is explained. This module receives symptom visual primitives and environmental response visual primitives output by the spatiotemporal feature decoupling network, and performs reasoning on the updated knowledge graph in conjunction with real-time environmental data. The core of the reasoning is a confidence propagation algorithm. First, the module performs correlation analysis between the environmental response visual primitives and the real-time environmental data to quantify and strip away the apparent visual features caused by non-pathological environmental stresses, such as temporary wilting. After stripping, the module focuses on analyzing the pure symptom visual primitives and calculates a net confidence score for each possible disease based on the causal relationships defined in the knowledge graph. All these scores together constitute a normalized diagnostic confidence vector. This process can be summarized by the following formula: ; This represents a normalized diagnostic confidence vector; This represents the diagnostic fusion reasoning module, used for confidence propagation calculation; Represents the symptom primitives in the current image; Represents the environmental response primitives of the current image; Representing a causal knowledge graph; Indicates real-time environmental parameters; This reasoning process achieves precise focusing on disease signals by actively removing environmental interference, making the diagnostic results more reliable and accurate.

[0024] Example 5: A knowledge graph-based method for processing large computer vision models also includes the step of: simulating the evolution trend of disease progression; Identify the highest probability disease type from the diagnostic confidence vector; Using a prospective disease evolution simulator, with the visual primitives of symptoms as the initial state and the causal subgraphs related to the highest probability disease type in the updated knowledge graph as the evolution model, a future disease evolution trajectory is simulated and output. In this embodiment, after generating the diagnostic confidence vector, a step of simulating the disease evolution trend is introduced. The system first determines the disease type with the highest probability from the diagnostic confidence vector. Subsequently, a prospective disease evolution simulator is activated. The simulator uses the current symptom visual primitives as the initial state of the simulation and extracts the causal subgraph related to the highest probability disease type from the updated knowledge graph as the evolution model. Combining the prediction data of future environmental conditions, the simulator extrapolates within a set time span, simulates and outputs a future symptom evolution trajectory. The simulation process is represented as follows: ; Indicates the future trajectory of symptom evolution; This represents a prospective disease evolution simulator; Indicates the type of disease with the highest probability of being diagnosed; Indicates the current symptom primitive; Representing a causal knowledge graph; Indicates predicted future environmental parameters; Indicates the predicted duration; By introducing this forward-looking simulation step, the system can not only diagnose current problems but also foresee future risks, providing managers with the ability to upgrade their decision-making from passive response to proactive intervention.

[0025] Example 6: The prospective disease evolution simulator further acquires data on future environmental conditions and the duration of future disease predictions during simulation. The disease evolution trend simulation is based on future environmental condition prediction data. Within the predicted time of the disease, a time-series-based stochastic process is performed on the causal subgraph. In this embodiment, the specific execution mechanism of the prospective disease evolution simulator is defined in detail. When executing the simulation, the simulator needs to obtain two key inputs: a set of predicted data on future environmental conditions, such as the temperature and humidity curve for the next 24 hours; and a future disease duration for defining the prediction range. The simulation of the disease evolution trend is based on the acquired future environmental data and performs a time-series-based stochastic process extrapolation on the causal subgraph related to the specific disease within this specified duration. This simulation method based on specific environmental prediction and time-series extrapolation makes disease prediction no longer a static inference, but a dynamic evolutionary process closely coupled with environmental changes, which greatly improves the realism and practical value of the prediction trajectory.

[0026] Example 7: A knowledge graph-based computer vision large model processing method includes the following step after step two: generating the optimal and precise intervention strategy; The diagnostic result determined from the diagnostic confidence vector; Based on the confirmed diagnostic results, the intervention decision optimization engine is invoked to retrieve all relevant candidate interventions from the updated knowledge graph, and the optimal intervention plan is recommended from the candidate interventions through quantitative evaluation. In this embodiment, after completing the diagnostic confidence inference, a step of generating the optimal and precise intervention strategy is added. The system determines the final diagnostic result based on the diagnostic confidence vector. Based on this result, an intervention decision optimization engine is invoked. The engine first retrieves all relevant candidate intervention measures from the updated knowledge graph, around the identified disease type, such as adjusting water and fertilizer, applying specific agents, etc. Subsequently, the engine performs a quantitative evaluation of each candidate intervention measure and finally recommends an optimal intervention plan from all candidates. This step extends the system's function from diagnosis and early warning to decision support, forming a complete closed loop from "discovering the problem" to "solving the problem," aiming to achieve scientific and optimized decision-making.

[0027] Example 8: The quantitative evaluation process of the intervention decision optimization engine further includes: For candidate interventions, perform counterfactual simulations to predict intervention effects; Counterfactual simulation involves simulating the future environment after the intervention and, based on the future environment after the intervention, recalculating the evolution trajectory of symptoms after the intervention. The severity of symptoms is assessed using a symptom severity assessment function. The degree of damage in the future symptom evolution trajectory and the post-intervention symptom evolution trajectory is quantified, and the difference between the two is defined as the expected therapeutic effect of the candidate intervention. In this embodiment, the core of the quantitative evaluation of the intervention decision optimization engine is counterfactual simulation. For each candidate intervention, the engine performs a counterfactual simulation to predict its intervention effect. This simulation process first quantifies the intervention as an adjustment to the prediction parameters of the future environment, thereby generating a "post-intervention" future environmental condition. Then, based on this new environmental condition, the engine recalculates a post-intervention symptom evolution trajectory. Finally, a predefined symptom severity assessment function is used to quantify the cumulative damage degree of the evolution trajectory without intervention and the post-intervention evolution trajectory, respectively. The difference between the two is defined as the expected therapeutic effect of the candidate intervention. The expected treatment effect is calculated as follows: ; This indicates the expected therapeutic effect of candidate intervention a; Functions representing the severity assessment of symptoms; This indicates the trajectory of symptom evolution without intervention; This indicates the trajectory of symptom evolution after intervention; Through this rigorous quantitative evaluation, the system can objectively compare the potential effectiveness of different intervention measures, providing solid data support for selecting the optimal plan.

[0028] Example 9: Before recommending the optimal intervention plan, the intervention decision optimization engine further includes: Calculate the overall utility value for candidate interventions; The overall utility value is generated through a utility function, which integrates expected treatment effects, implementation costs, and potential environmental risks, and is balanced by weighting coefficients. The candidate intervention with the highest overall utility value is identified as the optimal intervention plan; In this embodiment, before determining the optimal intervention plan, the intervention decision optimization engine calculates a comprehensive utility value for each candidate intervention measure. This comprehensive utility value is generated through a dedicated utility function. This function aims to achieve a balance of multiple objectives, integrating not only the expected therapeutic effect calculated through counterfactual simulation, but also considering the implementation cost of the measure and potential environmental risks. These three indicators are balanced through weight coefficients that can be dynamically adjusted by the user or the system. The utility function is expressed in the following form: ; This represents the overall utility value of candidate intervention a; , , Weighting coefficients representing treatment effectiveness, cost, and environmental risk; This indicates the expected therapeutic effect of candidate intervention a; Indicates intervention measures Execution costs; Indicates intervention measures Potential environmental risks; , , It is quantified and converted to a uniform utility scale before calculation; Ultimately, the candidate intervention with the highest overall utility value is identified as the optimal intervention plan and recommended to the user; this mechanism ensures that the final decision is not only technically effective, but also economically and environmentally sound.

[0029] Example 10: The closed-loop calibration mechanism further obtains the preset learning rate during calculation; The closed-loop calibration mechanism backpropagates the bias based on the learning rate, which is used to synchronously fine-tune the weights of the weighted directed edges in the initial causal knowledge graph and the trainable model parameters of the spatiotemporal feature decoupling network, so as to achieve adaptive iteration of the model. In this embodiment, the adaptive iterative process of the closed-loop calibration mechanism is specified. After receiving the final diagnosis result from the agricultural technology expert, the mechanism compares it with the previously output diagnostic confidence vector of the system and calculates the deviation between the two. The mechanism backpropagates the calculated deviation based on a preset learning rate. This backpropagation process simultaneously acts on two aspects: first, it fine-tunes the weights of the relevant weighted directed edges in the initial causal knowledge graph to correct the quantification strength of the causal relationship; second, it updates the trainable model parameters in the spatiotemporal feature decoupling network to optimize the feature extraction capability. The adaptive iterative process of the model is defined by the following formula: ; , This represents the knowledge graph before and after the update. , Indicates the model parameters before and after the update; This indicates a closed-loop calibration mechanism; This indicates the expert's final diagnosis; This represents the diagnostic confidence vector output by the system. Indicates the preset learning rate; Through this mechanism, the model can continuously self-calibrate and optimize using expert knowledge in ongoing real-world applications, thereby achieving adaptive iteration and continuous improvement in performance, ensuring the long-term effectiveness and accuracy of the model.

[0030] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for processing large computer vision models based on knowledge graphs, characterized in that, Includes the following steps: Step 1: Construct an initial causal knowledge graph; The historical dataset is obtained, which includes historical crop image sequences, historical environmental state sequences, and historical expert diagnostic annotations. The historical dataset is then processed using a causal inference engine to construct an initial causal knowledge graph. Step two: Reasoning for dynamic diagnostic confidence; Acquire real-time crop image sequences; process the real-time crop image sequences using a spatiotemporal feature decoupling network, which is controlled by trainable model parameters; process the visual primitives representing disease symptoms and the visual primitives representing environmental responses used to decouple the output; use the diagnostic fusion reasoning module to perform comprehensive reasoning on the visual primitives representing disease symptoms and the visual primitives representing environmental responses, combining the initial causal knowledge graph and real-time environmental data, and output a diagnostic confidence vector; Step 3: Perform diagnostic feedback closed-loop adaptive calibration; Obtain the expert's final diagnostic result corresponding to the diagnostic confidence vector; Using a closed-loop calibration mechanism, calculations are performed based on the deviation between the expert's final diagnostic result and the diagnostic confidence vector. The model parameters used to synchronously update the initial causal knowledge graph and the spatiotemporal feature decoupling network are calculated to generate the updated knowledge graph and updated model parameters, which are then used to perform subsequent diagnostic inference.

2. The method for processing large computer vision models based on knowledge graphs according to claim 1, characterized in that, Step one, constructing the initial causal knowledge graph, further includes: Historical crop image sequences, historical environmental state sequences, and historical expert diagnostic annotations were collected and structured to form a historical dataset. The initial causal knowledge graph is a directed acyclic graph, which contains nodes and weighted directed edges. Nodes are used to represent known disease types, visual symptoms, stress responses, and environmental factors, while weighted directed edges are used to quantify the strength of causal relationships between nodes.

3. The method for processing large computer vision models based on knowledge graphs according to claim 1, characterized in that, The spatiotemporal feature decoupling network integrates a multi-scale temporal attention mechanism. The multi-scale temporal attention mechanism is used to analyze the temporal dimension features of real-time crop image sequences and separate segments with different frequencies of change in the image sequences to distinguish between low-frequency stable tissue texture evolution caused by disease and high-frequency transient morphological changes caused by microenvironmental stress.

4. The method for processing large computer vision models based on knowledge graphs according to claim 1, characterized in that, The reasoning process of the diagnostic fusion reasoning module further includes: Execute the confidence propagation algorithm on the updated knowledge graph; The visual primitives of environmental response are correlated with real-time environmental data to quantify and isolate the apparent visual features caused by non-pathological environmental stress. Based on the stripping results, when analyzing the visual primitives of symptoms, a net confidence score is generated for each possible disease. The net confidence scores together constitute the diagnostic confidence vector.

5. The method for processing large computer vision models based on knowledge graphs according to claim 1, characterized in that, Also includes: Simulate the disease progression trend; Identify the highest probability disease type from the diagnostic confidence vector; Using a prospective disease evolution simulator, with symptom visual primitives as the initial state and causal subgraphs related to the highest probability disease type in the updated knowledge graph as the evolution model, we simulate and output the future symptom evolution trajectory.

6. The method for processing large computer vision models based on knowledge graphs according to claim 4, characterized in that, The prospective disease evolution simulator further acquires data on future environmental conditions and the duration of future disease predictions during simulation. The disease evolution trend simulation is based on future environmental condition prediction data. Within the predicted time frame of the future disease, a time-series-based stochastic process is performed on the causal subgraph.

7. The method for processing large computer vision models based on knowledge graphs according to claim 1, characterized in that, It also includes: generating optimal and precise intervention strategies; The diagnostic result determined from the diagnostic confidence vector; Based on the confirmed diagnostic results, the intervention decision optimization engine is invoked to retrieve all relevant candidate interventions from the updated knowledge graph, and the optimal intervention plan is recommended from the candidate interventions through quantitative evaluation.

8. The method for processing large computer vision models based on knowledge graphs according to claim 6, characterized in that, The quantitative evaluation process of the intervention decision optimization engine further includes: For candidate interventions, perform counterfactual simulations to predict intervention effects; Counterfactual simulation involves simulating the future environment after the intervention and, based on the future environment after the intervention, recalculating the evolution trajectory of symptoms after the intervention. The severity of symptoms is assessed using a symptom severity assessment function. The degree of damage in the future symptom evolution trajectory and the post-intervention symptom evolution trajectory is quantified, and the difference between the two is defined as the expected therapeutic effect of the candidate intervention.

9. The method for processing large computer vision models based on knowledge graphs according to claim 7, characterized in that, Before recommending the optimal intervention plan, the intervention decision optimization engine further includes: Calculate the overall utility value for candidate interventions; The overall utility value is generated through a utility function, which integrates expected treatment effects, implementation costs, and potential environmental risks, and is balanced by weighting coefficients. The candidate intervention with the highest overall utility value is identified as the optimal intervention plan.

10. The method for processing large computer vision models based on knowledge graphs according to claim 1, characterized in that, The closed-loop calibration mechanism further obtains the preset learning rate during calculation; The closed-loop calibration mechanism backpropagates the bias based on the learning rate, which is used to synchronously fine-tune the weights of the weighted directed edges in the initial causal knowledge graph and the trainable model parameters of the spatiotemporal feature decoupling network, so as to achieve adaptive iteration of the model.

Images