Liquid fertilizer application effect analysis method based on self-supervised learning

Abstract

The invention discloses a liquid fertilizer application effect analysis method based on self-supervised learning. The method comprises the following steps: S1, generating initial feature data; s2, environment feature representation is generated based on a heterogeneous environment self-adaption module; s3, extracting short-term and long-term spatio-temporal characteristics of the liquid fertilizer application effect based on the dual spatio-temporal characteristic decoupling and dynamic fusion model, and generating comprehensive spatio-temporal characteristic representation; s4, generating fertilization characteristics in different situations by utilizing self-supervised learning, and eliminating environmental interference factors through a causal reasoning module; s5, generating a fertilization intention feature, and dynamically adjusting a fertilization strategy; s6, when the difference between the prediction result and the actual effect exceeds a threshold value, triggering a re-learning module to update model parameters; and S7, carrying out integrated optimization, realizing feature sharing and task coordination through multi-task collaborative learning, and carrying out automatic tuning by utilizing Bayesian optimization. The intelligent analysis of the liquid fertilizer application effect is realized through the self-supervised learning and multi-task collaborative optimization method.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent agricultural fertilization, and in particular to a method for analyzing the application effect of liquid fertilizer based on self-supervised learning. Background Art

[0002] In modern agricultural production, liquid fertilizer application is widely used as an efficient fertilization method because it can be absorbed by crops more quickly and improve fertilizer utilization. The accuracy and efficiency of liquid fertilizer application directly affect the growth quality and yield of crops. However, the effect of liquid fertilizer application is complexly affected by environmental conditions (such as climate, soil moisture, soil pH, etc.) and crop growth status, and its application effect may vary significantly under different fertilization conditions. Therefore, in order to ensure the stability of fertilization effects, existing technologies generally rely on manual experience or single sensor monitoring to make fertilization decisions, which makes it difficult to fully consider the diversity of the environment and the needs of crops at different growth stages.

[0003] The traditional liquid fertilizer application method mainly determines the amount, frequency and concentration of fertilizer application through a simple summary analysis of environmental sensor data. The limitation of this method is that it is difficult to adapt to the diverse agricultural environment and the dynamic growth needs of crops. In addition, in the prior art, most fertilization systems only use fixed fertilization parameters and lack intelligent adjustment mechanisms. Even after monitoring environmental changes or crop feedback data, the adjustment of fertilization parameters still depends on fixed preset rules or manual intervention. This lack of intelligent adaptability in fertilization leads to low fertilizer utilization efficiency and unstable fertilization effects. Especially in the face of complex agricultural environments (such as areas with large climate fluctuations and diverse soil types), the fertilization methods of the prior art are difficult to optimize based on real-time feedback.

[0004] In order to solve these problems, some agricultural fertilization methods based on machine learning have emerged in recent years. They provide data support for fertilization by modeling and predicting historical data. Such technologies usually rely on supervised learning algorithms to train models with a large amount of labeled historical data. However, due to the complexity and variability of the agricultural environment, it is very difficult to obtain a large amount of high-quality labeled data. The dependence on supervised learning makes these methods perform poorly in new environments or unlabeled data. At the same time, the lack of effective use of real-time feedback data during the fertilization process makes it difficult to achieve dynamic analysis and adaptive optimization of the effect of liquid fertilizer application.

[0005] Existing intelligent fertilization technologies often lack comprehensive consideration of characteristics at different time scales when dealing with a variety of environmental factors, and are unable to accurately capture the different short-term and long-term effects of liquid fertilizer application. The impact of liquid fertilizer on crops has both immediate short-term effects and slower long-term effects. However, existing technologies usually use a single time series model to analyze the effects of fertilization, ignoring the differences between short-term and long-term effects. In addition, traditional models lack the ability to adaptively adjust environmental data features during feature fusion, resulting in poor generalization capabilities under diverse environmental conditions, which places certain limitations on the scope of application and accuracy of the model.

[0006] In addition, existing technologies lack support for causal reasoning when analyzing the application of liquid fertilizers. The agricultural environment has complex cause-and-effect relationships, and correlation analysis alone cannot accurately eliminate the interference of environmental variables on fertilization effects, which greatly affects the prediction accuracy of fertilization effects. Traditional models often only perform correlation analysis and ignore the in-depth modeling of cause-and-effect relationships, making it difficult to optimize fertilization decisions based on the actual needs of crop growth. The lack of causal analysis will cause the model to misjudge the true relationship between environmental factors and fertilization effects, thereby affecting the accuracy and adaptability of fertilization strategies.

[0007] For the adjustment of fertilization strategies, existing technologies mainly rely on fixed rules and thresholds, lacking dynamic adjustment and intelligent optimization methods. With environmental changes and the continuous updating of crop growth status, fixed fertilization strategies are difficult to meet the needs of dynamic changes. Although some technologies attempt to optimize fertilization strategies through feedback mechanisms, they are often based on simple comparative analysis without more complex feedback control mechanisms. Once the actual fertilization effect does not meet expectations, it is difficult for existing systems to automatically update fertilization strategies, and usually manual intervention or full model retraining is required, resulting in a waste of time costs and computing resources. This inefficient adjustment mechanism is particularly insufficient in actual agricultural environments, especially in situations where the crop growth cycle is long or the environment changes frequently. The existing fertilization model has obvious lags and it is difficult to achieve real-time optimization and updates.

[0008] In the optimization of fertilization strategy parameters, the existing intelligent fertilization system lacks automation and precision means for tuning key parameters, and often adopts manual settings or rough parameter searches, which are inefficient and ineffective. For parameter adjustment in multi-task fertilization scenarios, most traditional methods lack a multi-task coordination mechanism and fail to effectively achieve feature sharing and dynamic collaborative optimization between tasks. In addition, commonly used parameter optimization algorithms, such as grid search and random search, are difficult to find the optimal parameter combination in complex agricultural environments, and the existing machine learning-based parameter tuning algorithms are limited to single-task learning models and fail to comprehensively consider the relationship and potential synergy between different tasks, resulting in limited optimization of the overall fertilization effect.

[0009] Therefore, how to provide a liquid fertilizer application effect analysis method based on self-supervised learning is a problem that technicians in this field urgently need to solve. Summary of the invention

[0010] One purpose of the present invention is to propose a method for analyzing the effect of liquid fertilizer application based on self-supervised learning. The present invention adopts self-supervised learning, multi-task collaborative optimization and causal reasoning methods to intelligently analyze and dynamically optimize the effect of liquid fertilizer application. By generating multi-scenario fertilization characteristics, dual spatiotemporal decoupling of short-term and long-term fertilization effects, and a real-time feedback mechanism, it is ensured that the fertilization strategy can be adaptively adjusted according to crop growth status and environmental changes. Bayesian optimization further improves the automatic tuning efficiency of key parameters, so that the present invention has strong environmental adaptability and real-time response capabilities, and significantly improves the application accuracy and resource utilization efficiency of liquid fertilizers.

[0011] A method for analyzing the effect of liquid fertilizer application based on self-supervised learning according to an embodiment of the present invention comprises the following steps:

[0012] S1, collect environmental data and crop images through multiple sensors, and perform cleaning, standardization and feature extraction to generate initial feature data;

[0013] S2, input the initial feature data into the heterogeneous environment adaptive module, dynamically adjust the convolution kernel size and feature fusion strategy according to different environmental characteristics, and generate environmental feature representation;

[0014] S3. Based on the dual spatiotemporal feature decoupling and dynamic fusion model, the environmental feature representation is input into the multi-scale temporal feature extraction module to extract the short-term spatiotemporal features and long-term spatiotemporal features of the liquid fertilizer application effect. The spatiotemporal decoupling mechanism is used to separate feature channels in different time periods, and the short-term spatiotemporal features and long-term spatiotemporal features are fused through a dynamic weighting mechanism. According to the crop growth cycle and the changes in fertilization effect, the fusion ratio is automatically adjusted to generate a comprehensive spatiotemporal feature representation.

[0015] S4. Based on the generated comprehensive spatiotemporal feature representation, self-supervised learning is used to generate fertilization characteristics under different scenarios, and the causal reasoning module is used to eliminate environmental interference factors and extract the real impact characteristics of liquid fertilizer at different growth stages;

[0016] S5. Generate fertilization intention features based on real impact features and crop growth status, dynamically adjust fertilization strategies, and regenerate fertilization intentions when feedback effects are not ideal;

[0017] S6. Collect crop growth change data after fertilization through a real-time feedback loop, calculate the difference between the predicted result and the actual effect, and when the difference exceeds a threshold, trigger the re-learning module to update the fertilization parameters;

[0018] S7. Perform integrated optimization, achieve feature sharing and task coordination through multi-task collaborative learning, and use Bayesian optimization to automatically tune key parameters.

[0019] Furthermore, the S2 specifically includes:

[0020] S21, inputting the initial feature data generated in S1 into a heterogeneous environment adaptation module, wherein the initial feature data includes soil moisture, nutrient content, pH value and climate data;

[0021] S22, in the heterogeneous environment adaptation module, dynamically selecting and initializing convolution kernel parameters, including an initial size and an initial shape of the convolution kernel, according to the soil moisture, nutrient composition, pH value and climate data;

[0022] S23. Based on the multiple statistical characteristics of the initial feature data, the size and shape of the convolution kernel are adaptively adjusted:

[0023]

[0024] Among them, K adjust represents the adjusted convolution kernel size, K init represents the initial convolution kernel size, E var represents the variance of the initial feature data, E avg represents the mean of the initial feature data, E std represents the standard deviation of the initial feature data, E range represents the value range of the initial feature data, E max Indicates the maximum value of the initial feature data, E min Represents the minimum value of the initial feature data, α a , α d , β b , β f , γ c , γ g and δ e represents the control coefficient;

[0025] S24, performing a convolution operation using the adjusted convolution kernel parameters, performing multi-scale feature extraction on the initial feature data, and generating environment feature matrices at different scales;

[0026] S25. In the heterogeneous environment adaptation module, a multi-level feature fusion strategy is applied to fuse the environmental feature matrices of different scales and modes in space, channel and scale;

[0027] S26. The final output is the generated environmental feature representation.

[0028] Furthermore, the S3 specifically includes:

[0029] S31, inputting the generated environmental feature representation into a dual spatiotemporal feature decoupling and dynamic fusion model, and performing temporal decomposition of the environmental feature representation through a multi-scale temporal feature extraction module in the dual spatiotemporal feature decoupling and dynamic fusion model;

[0030] S32, through the short-term spatiotemporal feature extraction channel, the input environmental feature data is sent to the long short-term memory network structure to extract the short-term effect characteristics after fertilization, record the rapid response characteristics of crops within a few days or weeks, and generate short-term spatiotemporal features F short ;

[0031] S33, using the long-term spatiotemporal feature extraction channel, input the environmental feature data into the temporal convolutional network, extract the long-term impact of fertilization effect, and generate the long-term spatiotemporal feature F containing the impact of fertilization in several months or the entire growing season. long ;

[0032] S34, apply the space-time decoupling mechanism to transform the short-term space-time characteristics F short and long-term spatiotemporal characteristics F long Separate into independent characteristic channels in different time periods to form short-term channels and long-term channels;

[0033] S35, based on a dynamic weighting mechanism, setting adaptive weight parameters, acting on the spatiotemporal characteristics of the short-term channel and the long-term channel respectively, to form weighted short-term characteristics and long-term characteristics;

[0034] S36, the weighted short-term features and long-term features are integrated to generate a comprehensive spatiotemporal feature representation:

[0035]

[0036] Among them, F fusion represents the comprehensive spatiotemporal feature representation, exp represents the exponential function, α represents the weighting parameter of the short-term feature, β represents the weighting parameter of the long-term feature, γ represents the power control coefficient, δ represents the exponential control coefficient, η represents the transformation parameter of the long-term feature, sinh represents the hyperbolic sine function, θ represents the coefficient of the hyperbolic sine term, and ζ represents the normalization parameter.

[0037] Furthermore, the S4 specifically includes:

[0038] S41, inputting the generated comprehensive spatiotemporal feature representation into a self-supervised learning module to generate fertilization feature representations under multiple different scenarios, wherein the self-supervised learning module generates multi-scenario fertilization features by constructing an unlabeled data comparison task and constructs a multi-scenario fertilization feature library;

[0039] S42. In the self-supervised learning process, the target context feature F is created by contrastive learning method target and contrast situational features Fcontrast , and use the contrast loss function L contrast Indicates the difference between situations:

[0040]

[0041] Among them, exp represents the exponential function, sim represents the similarity function, and F j contrast represents the jth contrast situation feature, F positive represents the positive sample feature representation, ∥·∥ represents the norm operation of the vector, κ1 and κ2 represent the weight coefficients of the norm term, and τ represents the parameter regulating the similarity distribution in contrastive learning;

[0042] S43, inputting the generated multi-scenario fertilization feature library into a causal reasoning module, establishing a causal relationship between the features through a structural equation model, eliminating environmental interference factors, and the causal reasoning module identifies the direct causal relationship between the fertilization features and the crop growth;

[0043] S44. Use potential confounding variable models to correct fertilization characteristics and crop growth characteristics, identify the impact of potential environmental variables and eliminate noise interference:

[0044]

[0045] Among them, Y represents the impact of crop growth, X represents the fertilization characteristics, Z represents the potential environmental variables, and α x , β z , γ p , δ s and θ x represents the weight coefficient, η y represents the normalization coefficient;

[0046] S45. Apply counterfactual reasoning methods to create counterfactual situation features F counter , simulate the crop growth effects under different fertilization conditions, adjust the fertilization characteristics through intervention operations, and generate crop responses under hypothetical scenarios;

[0047] S46, by comparing the actual situation characteristics F actual and counterfactual situation feature F counter , extract the real impact characteristics of liquid fertilizer in different growth stages, eliminate the interference of potential environmental variables, and finally obtain the real fertilization effect during crop growth.

[0048] Furthermore, the S5 specifically includes:

[0049] S51, the generated real impact feature F true The current growth status S of the crop is input into the fertilization intention generation module to generate the initial fertilization intention feature Iinitial , the initial fertilization intention feature I initial Used to indicate the best strategy for liquid fertilization under current circumstances;

[0050] S52, in the process of generating fertilization intention, based on the real impact feature F true The fertilization intention function is constructed by dynamically mapping the current growth state S of the crop. The fertilization intention function integrates the weights of different features to generate the initial fertilization intention feature I adapted to different environments and crop requirements. initial ;

[0051] S53, based on the generated initial fertilization intention feature I initial , generate a set of fertilization strategy parameters P, including the fertilization amount P amount , Fertilization frequency P freq and fertilizer concentration P conc , to achieve the best fertilization effect;

[0052] S54. During the fertilization process, based on the real-time feedback data of crop growth actual Monitor the effect of fertilization and compare the feedback results with the expected effect expected Compare and if the deviation exceeds the set threshold, the fertilization intention update mechanism is triggered:

[0053]

[0054] Among them, D represents the deviation between the actual effect and the expected effect, α growth represents the growth-related regulation coefficient, β env represents the weight coefficient of environmental characteristics, γ fert The adjustment coefficient representing the fertilization characteristics, δ crop represents the adjustment coefficient of crop characteristics, ∈ represents the smoothing term;

[0055] S55, the fertilization intention update mechanism adjusts the initial fertilization intention feature I according to the deviation D between the actual effect and the expected effect initial :

[0056]

[0057] Among them, I update represents the updated fertilization intention feature, exp represents the exponential function, and λ growth represents the growth regulation coefficient, λ env represents the environmental adjustment coefficient, θ fert represents the fertilization characteristic adjustment coefficient, ζ crop represents the crop characteristic adjustment coefficient, η growth represents the gain control coefficient;

[0058] S56, according to the updated fertilization intention feature I update , regenerate the fertilization strategy parameter set P and apply it to the liquid fertilizer application process.

[0059] Furthermore, the S7 specifically includes:

[0060] S71. Based on the multi-task collaborative learning mechanism, a unified feature vector set is constructed to achieve feature sharing and form a stable multi-task fertilization model.

[0061] S72. On the basis of feature sharing, an independent task representation is generated for each fertilization task, and each fertilization task retains its own characteristics while sharing features; based on the design of task-specific encoding modules, feature subsets suitable for specific tasks are automatically identified and utilized when dealing with different tasks;

[0062] S73. In multi-task collaborative learning, an interaction module between tasks is introduced to capture the potential correlation between different tasks, generate a joint feature representation, and through the weighting mechanism of the interaction module, the learning process of a task is fed back to the feature update of other tasks;

[0063] S74. Based on the multi-task learning framework, the Bayesian optimization method is used to automatically tune the key parameters, construct the parameter space, and gradually update the parameter distribution according to the feedback data of the fertilization task;

[0064] S75. Using a proxy model in Bayesian optimization to quickly evaluate different parameter combinations, simulating feedback during fertilization through the proxy model, wherein the proxy model is updated based on historical data and learns the optimal parameter configuration from less experimental data;

[0065] S76. After the parameter tuning is completed, the optimal parameter configuration set is applied to the multi-task collaborative model to achieve the overall optimization of the fertilization strategy. Through the integrated optimized multi-task collaborative learning model, each fertilization task can obtain the best fertilization effect under different environments.

[0066] The beneficial effects of the present invention are:

[0067] First, by introducing a self-supervised learning module, the present invention can generate multi-scenario fertilization features based on unlabeled data, significantly reducing the reliance on labeled data, reducing the training cost of the model, and improving the adaptability of the model in new environments. At the same time, the introduction of dual spatiotemporal feature decoupling and dynamic fusion mechanisms enables the model to accurately capture the short-term response and long-term effects of liquid fertilizer application, meet the nutrient requirements of crops at different growth stages, avoid the limitations of a single time series feature model in a complex agricultural environment, and thus improve the accuracy of the fertilization strategy.

[0068] Secondly, the present invention can effectively eliminate the interference of environmental variables through the use of causal reasoning modules, thereby performing a more realistic and reliable analysis of the fertilization effect. Compared with the limitations of traditional methods based only on correlation analysis, the present method can identify the real impact of liquid fertilizers at different growth stages, avoid the misleading of environmental variables on the prediction of fertilization effects, and fundamentally improve the scientificity and reliability of fertilization decisions. In addition, the design of dynamically adjusting the fertilization intention characteristics and fertilization strategies enables the present invention to respond to changes in the growth status of crops and the external environment in real time, ensuring the flexibility of fertilization strategies. When the fertilization effect does not meet expectations, the present method can automatically adjust the fertilization intention and optimize the fertilization parameters according to the feedback data, thereby avoiding manual intervention, improving the autonomy and intelligence of the fertilization process, and further improving the utilization efficiency of liquid fertilizers.

[0069] In addition, the combination of multi-task collaborative learning and Bayesian optimization adopted by the present invention ensures the integrated optimization of fertilization strategies and the automatic tuning of parameters. The multi-task collaborative learning mechanism enables feature sharing and task coordination between different fertilization tasks, and enhances the stability of the model in a complex and changeable agricultural environment. At the same time, Bayesian optimization can efficiently search for the optimal combination in the multidimensional parameter space, provide accurate parameter configuration for each fertilization scenario, and ensure the best effect of the fertilization strategy under different environmental and crop conditions. Through integrated optimization, the fertilization model of the present invention has excellent generalization ability and adaptability, which can not only effectively cope with the changes of various environmental variables, but also can continuously provide optimal fertilization decisions under different crops and growth stages, thereby realizing the intelligentization, refinement and efficiency of the application effect of liquid fertilizers in agricultural production. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0071] Figure 1 A flow chart of a method for analyzing the effect of liquid fertilizer application based on self-supervised learning proposed by the present invention;

[0072] Figure 2 This is a structural schematic diagram of a dual spatiotemporal feature decoupling and dynamic fusion model of a liquid fertilizer application effect analysis method based on self-supervised learning proposed in the present invention. DETAILED DESCRIPTION

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0074] refer to Figure 1 and Figure 2 , a liquid fertilizer application effect analysis method based on self-supervised learning, comprising the following steps:

[0075] S1, collect environmental data and crop images through multiple sensors, and perform cleaning, standardization and feature extraction to generate initial feature data;

[0076] S2, input the initial feature data into the heterogeneous environment adaptive module, dynamically adjust the convolution kernel size and feature fusion strategy according to different environmental characteristics, and generate environmental feature representation;

[0077] S3. Based on the dual spatiotemporal feature decoupling and dynamic fusion model, the environmental feature representation is input into the multi-scale temporal feature extraction module to extract the short-term spatiotemporal features and long-term spatiotemporal features of the liquid fertilizer application effect. The spatiotemporal decoupling mechanism is used to separate feature channels in different time periods, and the short-term spatiotemporal features and long-term spatiotemporal features are fused through a dynamic weighting mechanism. According to the crop growth cycle and the changes in fertilization effect, the fusion ratio is automatically adjusted to generate a comprehensive spatiotemporal feature representation.

[0078] S4. Based on the generated comprehensive spatiotemporal feature representation, self-supervised learning is used to generate fertilization characteristics under different scenarios, and the causal reasoning module is used to eliminate environmental interference factors and extract the real impact characteristics of liquid fertilizer at different growth stages;

[0079] S5. Generate fertilization intention features based on real impact features and crop growth status, dynamically adjust fertilization strategies, and regenerate fertilization intentions when feedback effects are not ideal;

[0080] S6. Collect crop growth change data after fertilization through a real-time feedback loop, calculate the difference between the predicted result and the actual effect, and when the difference exceeds a threshold, trigger the re-learning module to update the fertilization parameters;

[0081] S7. Perform integrated optimization, achieve feature sharing and task coordination through multi-task collaborative learning, and use Bayesian optimization to automatically tune key parameters.

[0082] In this implementation, S2 specifically includes:

[0083] S21, inputting the initial feature data generated in S1 into a heterogeneous environment adaptation module, wherein the initial feature data includes soil moisture, nutrient content, pH value and climate data;

[0084] S22, in the heterogeneous environment adaptation module, dynamically selecting and initializing convolution kernel parameters, including an initial size and an initial shape of the convolution kernel, according to the soil moisture, nutrient composition, pH value and climate data;

[0085] S23. Based on the multiple statistical characteristics of the initial feature data, the size and shape of the convolution kernel are adaptively adjusted:

[0086]

[0087] Among them, K adjust represents the adjusted convolution kernel size, K init represents the initial convolution kernel size, E var represents the variance of the initial feature data, E avg represents the mean of the initial feature data, E std represents the standard deviation of the initial feature data, E range represents the value range of the initial feature data, E max Indicates the maximum value of the initial feature data, E min Represents the minimum value of the initial feature data, α a , α d , β b , β f , γ c , γ g and δ e represents the control coefficient;

[0088] S24, performing a convolution operation using the adjusted convolution kernel parameters, performing multi-scale feature extraction on the initial feature data, and generating environment feature matrices at different scales;

[0089] S25. In the heterogeneous environment adaptation module, a multi-level feature fusion strategy is applied to fuse the environmental feature matrices of different scales and modes in space, channel and scale;

[0090] S26. The final output is the generated environmental feature representation.

[0091] In this implementation, S3 specifically includes:

[0092] S31, inputting the generated environmental feature representation into a dual spatiotemporal feature decoupling and dynamic fusion model, and performing temporal decomposition of the environmental feature representation through a multi-scale temporal feature extraction module in the dual spatiotemporal feature decoupling and dynamic fusion model;

[0093] S32, through the short-term spatiotemporal feature extraction channel, the input environmental feature data is sent to the long short-term memory network structure to extract the short-term effect characteristics after fertilization, record the rapid response characteristics of crops within a few days or weeks, and generate short-term spatiotemporal features F short ;

[0094] S33, using the long-term spatiotemporal feature extraction channel, input the environmental feature data into the temporal convolutional network, extract the long-term impact of fertilization effect, and generate the long-term spatiotemporal feature F containing the impact of fertilization in several months or the entire growing season. long ;

[0095] S34, apply the space-time decoupling mechanism to transform the short-term space-time characteristics F short and long-term spatiotemporal characteristics F long Separate into independent characteristic channels in different time periods to form short-term channels and long-term channels;

[0096] S35, based on a dynamic weighting mechanism, setting adaptive weight parameters, acting on the spatiotemporal characteristics of the short-term channel and the long-term channel respectively, to form weighted short-term characteristics and long-term characteristics;

[0097] S36, the weighted short-term features and long-term features are integrated to generate a comprehensive spatiotemporal feature representation:

[0098]

[0099] Among them, F fusion represents the comprehensive spatiotemporal feature representation, exp represents the exponential function, α represents the weighting parameter of the short-term feature, β represents the weighting parameter of the long-term feature, γ represents the power control coefficient, δ represents the exponential control coefficient, η represents the transformation parameter of the long-term feature, sinh represents the hyperbolic sine function, θ represents the coefficient of the hyperbolic sine term, and ζ represents the normalization parameter.

[0100] In this implementation, S4 specifically includes:

[0101] S41, inputting the generated comprehensive spatiotemporal feature representation into a self-supervised learning module to generate fertilization feature representations under multiple different scenarios, wherein the self-supervised learning module generates multi-scenario fertilization features by constructing an unlabeled data comparison task and constructs a multi-scenario fertilization feature library;

[0102] S42. In the self-supervised learning process, the target context feature F is created by contrastive learning method target and contrast situational features F contrast , and use the contrast loss function L contrast Indicates the difference between situations:

[0103]

[0104] Among them, exp represents the exponential function, sim represents the similarity function, represents the jth contrast situation feature, F positiverepresents the positive sample feature representation, ∥·∥ represents the norm operation of the vector, κ1 and κ2 represent the weight coefficients of the norm term, and τ represents the parameter regulating the similarity distribution in contrastive learning;

[0105] S43, inputting the generated multi-scenario fertilization feature library into a causal reasoning module, establishing a causal relationship between the features through a structural equation model, eliminating environmental interference factors, and the causal reasoning module identifies the direct causal relationship between the fertilization features and the crop growth;

[0106] S44. Use potential confounding variable models to correct fertilization characteristics and crop growth characteristics, identify the impact of potential environmental variables and eliminate noise interference:

[0107]

[0108] Among them, Y represents the impact of crop growth, X represents the fertilization characteristics, Z represents the potential environmental variables, and α x , β z , γ p ,δ s and θ x represents the weight coefficient, η y represents the normalization coefficient;

[0109] S45. Apply counterfactual reasoning methods to create counterfactual situation features F counter , simulate the crop growth effects under different fertilization conditions, adjust the fertilization characteristics through intervention operations, and generate crop responses under hypothetical scenarios;

[0110] S46, by comparing the actual situation characteristics F actual and counterfactual situation feature F counter , extract the real impact characteristics of liquid fertilizer in different growth stages, eliminate the interference of potential environmental variables, and finally obtain the real fertilization effect during crop growth.

[0111] In this implementation manner, S5 specifically includes:

[0112] S51, the generated real impact feature F true The current growth status S of the crop is input into the fertilization intention generation module to generate the initial fertilization intention feature I initial , the initial fertilization intention feature I initial Used to indicate the best strategy for liquid fertilization under current circumstances;

[0113] S52, in the process of generating fertilization intention, based on the real impact feature F true The fertilization intention function is constructed by dynamically mapping the current growth state S of the crop. The fertilization intention function integrates the weights of different features to generate the initial fertilization intention feature I adapted to different environments and crop requirements.initial ;

[0114] S53, based on the generated initial fertilization intention feature I initial , generate a set of fertilization strategy parameters P, including the fertilization amount P amount , Fertilization frequency P freq and fertilizer concentration P conc , to achieve the best fertilization effect;

[0115] S54. During the fertilization process, based on the real-time feedback data of crop growth actual Monitor the effect of fertilization and compare the feedback results with the expected effect expected Compare and if the deviation exceeds the set threshold, the fertilization intention update mechanism is triggered:

[0116]

[0117] Among them, D represents the deviation between the actual effect and the expected effect, α growth represents the growth-related regulation coefficient, β env represents the weight coefficient of environmental characteristics, γ fert The adjustment coefficient representing the fertilization characteristics, δ crop represents the adjustment coefficient of crop characteristics, ∈ represents the smoothing term;

[0118] S55, the fertilization intention update mechanism adjusts the initial fertilization intention feature I according to the deviation D between the actual effect and the expected effect initial :

[0119]

[0120] Among them, I update represents the updated fertilization intention feature, exp represents the exponential function, and λ growth represents the growth regulation coefficient, λ env represents the environmental adjustment coefficient, θ fert represents the fertilization characteristic adjustment coefficient, ζ crop represents the crop characteristic adjustment coefficient, η growth represents the gain control coefficient;

[0121] S56, according to the updated fertilization intention feature I update , regenerate the fertilization strategy parameter set P and apply it to the liquid fertilizer application process.

[0122] In this implementation manner, the S7 specifically includes:

[0123] S71. Based on the multi-task collaborative learning mechanism, a unified feature vector set is constructed to achieve feature sharing and form a stable multi-task fertilization model.

[0124] S72. On the basis of feature sharing, an independent task representation is generated for each fertilization task, and each fertilization task retains its own characteristics while sharing features; based on the design of task-specific encoding modules, feature subsets suitable for specific tasks are automatically identified and utilized when dealing with different tasks;

[0125] S73. In multi-task collaborative learning, an interaction module between tasks is introduced to capture the potential correlation between different tasks, generate a joint feature representation, and through the weighting mechanism of the interaction module, the learning process of a task is fed back to the feature update of other tasks;

[0126] S74. Based on the multi-task learning framework, the Bayesian optimization method is used to automatically tune the key parameters, construct the parameter space, and gradually update the parameter distribution according to the feedback data of the fertilization task;

[0127] S75. Using a proxy model in Bayesian optimization to quickly evaluate different parameter combinations, simulating feedback during fertilization through the proxy model, wherein the proxy model is updated based on historical data and learns the optimal parameter configuration from less experimental data;

[0128] S76. After the parameter tuning is completed, the optimal parameter configuration set is applied to the multi-task collaborative model to achieve the overall optimization of the fertilization strategy. Through the integrated optimized multi-task collaborative learning model, each fertilization task can obtain the best fertilization effect under different environments.

[0129] Embodiment 1:

[0130] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a certain experimental field project. The experimental field covers an area of ​​20 hectares, and the crop planted is wheat. Agricultural production in the region faces diverse climate changes and complex soil types. Traditional fertilization technology is difficult to effectively adapt to the needs of different regions and different growth stages, resulting in uneven crop growth and unstable fertilization effects. Traditional liquid fertilizer application relies on fixed fertilization parameters, which cannot be dynamically adjusted when environmental factors change, resulting in low fertilizer utilization and serious soil nutrient loss. Through the method of the present invention, an attempt is made to solve the problem of intelligent application of liquid fertilizers in multiple environments, in order to achieve efficient and accurate fertilization effects.

[0131] In this embodiment, the system of the present invention is equipped with a variety of environmental sensors, including soil moisture sensors, temperature sensors, pH sensors, and nitrogen, phosphorus, and potassium content sensors. These sensors collect data every 30 minutes and transmit the data to the central processing system through a wireless network. After cleaning, standardization, and feature extraction, the data is input into the heterogeneous environment adaptive module. The module adaptively adjusts the size of the convolution kernel and the feature fusion strategy according to the soil and climate characteristics of different regions to generate an adaptable environmental feature representation. In the experiment, the initial data of soil moisture was distributed between 10-30%, the pH value was between 6.0-7.5, and the temperature fluctuated between 15-30°C due to seasonal changes. These environmental features are processed by the heterogeneous environment adaptive module and converted into standard features that the system can process.

[0132] Next, the system uses a dual spatiotemporal feature decoupling and dynamic fusion model to perform multi-scale feature extraction on the environmental feature representation. The short-term spatiotemporal feature extraction module captures the immediate effects after fertilization, such as short-term growth performance such as changes in wheat leaf color and stem growth height within a few days; the long-term spatiotemporal features are used to analyze the continuous impact of liquid fertilizers throughout the growing season, such as long-term performance such as increased plant height and leaf area index. In the first month of the initial trial, changes in short-term features showed the rapid response of wheat growth to liquid fertilizers, while as the growing season progressed, long-term features captured the effects of liquid fertilizers in improving crop disease resistance and increasing stem strength. These short-term and long-term features are dynamically weighted and fused in the dual spatiotemporal feature decoupling and dynamic fusion model to generate comprehensive spatiotemporal features to ensure that fertilization strategies can adapt to the actual needs of crops at different growth stages.

[0133] In order to further improve the accuracy of fertilization, the system generates a fertilization feature library under multiple scenarios through the self-supervised learning module. The fertilization data is analyzed using the causal reasoning module to eliminate environmental interference factors and identify the real impact of liquid fertilizers at different growth stages. In the experimental field, the system eliminated the random effects of climate and soil structure through causal reasoning, and obtained the actual impact of fertilizer application on leaf thickness, plant height and grain weight. For example, when soil moisture is low (10%-15%), the positive effect of fertilization on wheat leaf thickness and grain weight is more significant, while when the humidity is high (25%-30%), this effect is weakened. These causal characteristics provide a scientific basis for the system's fertilization strategy.

[0134] During the experiment, the system will adjust the fertilization strategy in real time based on the crop growth feedback data. If there is a large difference between the actual fertilization effect and the expected effect, the system will automatically regenerate the fertilization intention and optimize the fertilization parameters. In actual application, monitoring data in the experimental field showed that as wheat entered the heading period, the system automatically increased the proportion of nitrogen fertilizer to promote grain development. At the same time, if it is observed that the expected plant height growth rate after fertilization does not reach the set target (such as a 10% increase), the system will trigger the re-learning module, readjust the fertilization parameters, and apply the optimized parameters in the subsequent fertilization process.

[0135] The test data show that the method of the present invention can significantly improve the application efficiency of liquid fertilizer and the growth effect of wheat. After the fertilization process throughout the growing season, the test field using the method of the present invention increased the average plant height by 15%, the leaf area index by 12%, the grain weight by 10%, and the fertilizer usage by 12% compared with the control field using the traditional fertilization method. In addition, the fertilization strategy through dynamic optimization and feedback control also significantly improved the fertilizer utilization rate and reduced the soil salinization and nutrient loss caused by excessive fertilization. The following is some experimental data of the present invention in the wheat test field.

[0136] Table 1 Monthly average data of environmental conditions in wheat experimental fields

[0137]

[0138] Table 2 Wheat growth feedback data statistics

[0139] month Average plant height (cm) Leaf Area Index Stem strength (N) Grain weight (g / ear) March 12 1.5 5.8 1.2 April 24 2.1 6.5 1.5 May 42 2.8 7.2 1.8 June 55 3.2 8.0 2.2

[0140] Table 3 Dynamic adjustment record of liquid fertilizer application strategy

[0141]

[0142] In this experimental data, Table 1 records the monthly average data of the environmental conditions of the wheat experimental field, reflecting the changes in soil moisture, pH value, nitrogen, phosphorus, potassium content and temperature in the experimental field in different months. From March to June, soil moisture increased month by month, gradually increasing from 15% to 28%, indicating that the experimental field remained relatively moist in the middle of the growing season. At the same time, the pH value of the soil changed little, basically stabilizing between 6.8 and 7.2, which is in the pH range suitable for wheat growth. The nitrogen, phosphorus and potassium content also gradually increased under the action of seasonal fertilization, effectively supporting the growth needs of wheat. The temperature gradually rose from 16°C to 26°C in June, providing suitable temperature conditions for the rapid growth of wheat.

[0143] Table 2 summarizes the growth feedback data of wheat in different months, including important growth indicators such as plant height, leaf area index, stem strength and grain weight. From March to June, the average plant height of wheat increased from 12 cm to 55 cm, and the leaf area index also increased from 1.5 to 3.2, reflecting the improvement of the leaf coverage area and photosynthesis capacity of the plant. The stem strength increased from 5.8N in March to 8.0N in June, indicating that the crop's resistance to lodging has increased. At the same time, the continued increase in grain weight indicates that the precise application of liquid fertilizer has played a positive role in effectively increasing yield. These data further support the beneficial effects of the method of the present invention in the process of crop growth.

[0144] Table 3 shows the dynamic adjustment of liquid fertilizer application strategy during the trial. The experimental field adjusted the fertilizer amount, nitrogen, phosphorus, potassium ratio and fertilization frequency according to real-time feedback data to meet the nutritional needs of wheat at various growth stages. In March, the system set the nitrogen fertilizer ratio to 40% to promote early leaf growth; as the crop entered the vigorous growth period in April, the nitrogen fertilizer ratio was gradually increased to 45%. In the critical growth stages of May and June, the nitrogen fertilizer ratio was further increased to 50% and 55% to support wheat growth and grain development. The fertilization frequency was also increased accordingly, from the initial 2 times / month to 3 times / month to ensure a continuous supply of nutrients. The ratio of phosphorus and potassium fertilizers was also dynamically adjusted according to soil and crop needs, effectively supporting the development of the root system and the improvement of disease resistance.

[0145] From the above data analysis, it can be seen that the liquid fertilizer application effect analysis method of the present invention effectively improves the accuracy and adaptability of fertilization through real-time optimization of environmental characteristics, crop feedback and fertilization strategies. Various growth data show that the dynamically adjusted fertilization strategy significantly promotes the growth and yield of wheat, especially in terms of plant height, leaf area index, stalk strength and grain weight. At the same time, the feedback tuning mechanism of the present invention can be finely adjusted according to environmental changes and crop growth feedback, so as to achieve efficient use of resources and ensure that crops obtain optimal nutritional support at different growth stages. This shows that the present invention has good adaptability under multiple environmental conditions and can efficiently improve fertilizer utilization and crop yield in agricultural production.

[0146] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for analyzing the effect of liquid fertilizer application based on self-supervised learning, characterized in that: The steps include: S1, collect environmental data and crop images through multiple sensors, and perform cleaning, standardization and feature extraction to generate initial feature data; S2, input the initial feature data into the heterogeneous environment adaptive module, dynamically adjust the convolution kernel size and feature fusion strategy according to different environmental characteristics, and generate environmental feature representation; S3. Based on the dual spatiotemporal feature decoupling and dynamic fusion model, the environmental feature representation is input into the multi-scale temporal feature extraction module to extract the short-term spatiotemporal features and long-term spatiotemporal features of the liquid fertilizer application effect. The spatiotemporal decoupling mechanism is used to separate feature channels in different time periods, and the short-term spatiotemporal features and long-term spatiotemporal features are fused through a dynamic weighting mechanism. According to the crop growth cycle and the changes in fertilization effect, the fusion ratio is automatically adjusted to generate a comprehensive spatiotemporal feature representation. S4. Based on the generated comprehensive spatiotemporal feature representation, self-supervised learning is used to generate fertilization characteristics under different scenarios, and the causal reasoning module is used to eliminate environmental interference factors and extract the real impact characteristics of liquid fertilizer at different growth stages; S5. Generate fertilization intention features based on real impact features and crop growth status, dynamically adjust fertilization strategies, and regenerate fertilization intentions when feedback effects are not ideal; S6. Collect crop growth change data after fertilization through a real-time feedback loop, calculate the difference between the predicted result and the actual effect, and when the difference exceeds a threshold, trigger the re-learning module to update the fertilization parameters; S7. Perform integrated optimization, achieve feature sharing and task coordination through multi-task collaborative learning, and use Bayesian optimization to automatically tune key parameters.

2. The method for analyzing the effect of liquid fertilizer application based on self-supervised learning according to claim 1, characterized in that: The S2 specifically includes: S21, inputting the initial feature data generated in S1 into a heterogeneous environment adaptation module, wherein the initial feature data includes soil moisture, nutrient content, pH value and climate data; S22, in the heterogeneous environment adaptation module, dynamically selecting and initializing convolution kernel parameters, including an initial size and an initial shape of the convolution kernel, according to the soil moisture, nutrient composition, pH value and climate data; S23. Based on the multiple statistical characteristics of the initial feature data, the size and shape of the convolution kernel are adaptively adjusted: Among them, K adjust represents the adjusted convolution kernel size, K init represents the initial convolution kernel size, E var represents the variance of the initial feature data, E avg represents the mean of the initial feature data, E std represents the standard deviation of the initial feature data, E range represents the value range of the initial feature data, E max Indicates the maximum value of the initial feature data, E min Represents the minimum value of the initial feature data, α a , α d , β b , β f , γ c , γ g and δ e represents the control coefficient; S24, performing a convolution operation using the adjusted convolution kernel parameters, performing multi-scale feature extraction on the initial feature data, and generating environment feature matrices at different scales; S25. In the heterogeneous environment adaptation module, a multi-level feature fusion strategy is applied to fuse the environmental feature matrices of different scales and modes in space, channel and scale; S26. The final output is the generated environmental feature representation.

3. The method for analyzing the effect of liquid fertilizer application based on self-supervised learning according to claim 1, characterized in that: The S3 specifically includes: S31, inputting the generated environmental feature representation into a dual spatiotemporal feature decoupling and dynamic fusion model, and performing temporal decomposition of the environmental feature representation through a multi-scale temporal feature extraction module in the dual spatiotemporal feature decoupling and dynamic fusion model; S32, through the short-term spatiotemporal feature extraction channel, the input environmental feature data is sent to the long short-term memory network structure to extract the short-term effect characteristics after fertilization, record the rapid response characteristics of crops within a few days or weeks, and generate short-term spatiotemporal features F short ; S33, using the long-term spatiotemporal feature extraction channel, input the environmental feature data into the temporal convolutional network, extract the long-term impact of fertilization effect, and generate the long-term spatiotemporal feature F containing the impact of fertilization in several months or the entire growing season. long ; S34, apply the space-time decoupling mechanism to transform the short-term space-time characteristics F short and long-term spatiotemporal characteristics F long Separate into independent characteristic channels in different time periods to form short-term channels and long-term channels; S35, based on a dynamic weighting mechanism, setting adaptive weight parameters, acting on the spatiotemporal characteristics of the short-term channel and the long-term channel respectively, to form weighted short-term characteristics and long-term characteristics; S36, the weighted short-term features and long-term features are integrated to generate a comprehensive spatiotemporal feature representation: Among them, F fusion represents the comprehensive spatiotemporal feature representation, exp represents the exponential function, α represents the weighting parameter of the short-term feature, β represents the weighting parameter of the long-term feature, γ represents the power control coefficient, δ represents the exponential control coefficient, η represents the transformation parameter of the long-term feature, sinh represents the hyperbolic sine function, θ represents the coefficient of the hyperbolic sine term, and ζ represents the normalization parameter.

4. The method for analyzing the effect of liquid fertilizer application based on self-supervised learning according to claim 1, characterized in that: The S4 specifically includes: S41, inputting the generated comprehensive spatiotemporal feature representation into a self-supervised learning module to generate fertilization feature representations under multiple different scenarios, wherein the self-supervised learning module generates multi-scenario fertilization features by constructing an unlabeled data comparison task and constructs a multi-scenario fertilization feature library; S42. In the self-supervised learning process, the target context feature F is created by contrastive learning method target and contrast situational features F contrast , and use the contrast loss function L contrast Indicates the difference between situations: Among them, exp represents the exponential function, sim represents the similarity function, and F j contrast represents the jth contrast situation feature, F positive represents the positive sample feature representation, ∥·∥ represents the norm operation of the vector, κ1 and κ2 represent the weight coefficients of the norm term, and τ represents the parameter regulating the similarity distribution in contrastive learning; S43, inputting the generated multi-scenario fertilization feature library into a causal reasoning module, establishing a causal relationship between the features through a structural equation model, eliminating environmental interference factors, and the causal reasoning module identifies the direct causal relationship between the fertilization features and the crop growth; S44. Use potential confounding variable models to correct fertilization characteristics and crop growth characteristics, identify the impact of potential environmental variables and eliminate noise interference: Among them, Y represents the impact of crop growth, X represents the fertilization characteristics, Z represents the potential environmental variables, and α x , β z , γ p , δ s and θ x represents the weight coefficient, η y represents the normalization coefficient; S45. Apply counterfactual reasoning methods to create counterfactual situation features F counter , simulate the crop growth effects under different fertilization conditions, adjust the fertilization characteristics through intervention operations, and generate crop responses under hypothetical scenarios; S46, by comparing the actual situation characteristics F actual and counterfactual situation feature F counter , extract the real impact characteristics of liquid fertilizer in different growth stages, eliminate the interference of potential environmental variables, and finally obtain the real fertilization effect during crop growth.

5. The method for analyzing the effect of liquid fertilizer application based on self-supervised learning according to claim 1, characterized in that: The S5 specifically includes: S51, the generated real impact feature F true The current growth status S of the crop is input into the fertilization intention generation module to generate the initial fertilization intention feature I initial , the initial fertilization intention feature I initial Used to indicate the best strategy for liquid fertilization under current circumstances; S52, in the process of generating fertilization intention, based on the real impact feature F true The fertilization intention function is constructed by dynamically mapping the current growth state S of the crop. The fertilization intention function integrates the weights of different features to generate the initial fertilization intention feature I adapted to different environments and crop requirements. initial ; S53, based on the generated initial fertilization intention feature I initial , generate a set of fertilization strategy parameters P, including the fertilization amount P amount , Fertilization frequency P freq and fertilizer concentration P conc , to achieve the best fertilization effect; S54. During the fertilization process, based on the real-time feedback data of crop growth actual Monitor the effect of fertilization and compare the feedback results with the expected effect expected Compare and if the deviation exceeds the set threshold, the fertilization intention update mechanism is triggered: Among them, D represents the deviation between the actual effect and the expected effect, α growth represents the growth-related regulation coefficient, β env represents the weight coefficient of environmental characteristics, γ fert The adjustment coefficient representing the fertilization characteristics, δ crop represents the adjustment coefficient of crop characteristics, ∈ represents the smoothing term; S55, the fertilization intention update mechanism adjusts the initial fertilization intention feature I according to the deviation D between the actual effect and the expected effect initial : Among them, I update represents the updated fertilization intention feature, exp represents the exponential function, and λ growth represents the growth regulation coefficient, λ env represents the environmental adjustment coefficient, θ fert represents the fertilization characteristic adjustment coefficient, ζ crop represents the crop characteristic adjustment coefficient, η growth represents the gain control coefficient; S56, according to the updated fertilization intention feature I update , regenerate the fertilization strategy parameter set P and apply it to the liquid fertilizer application process.

6. The method for analyzing the effect of liquid fertilizer application based on self-supervised learning according to claim 1, characterized in that: The S7 specifically includes: S71. Based on the multi-task collaborative learning mechanism, a unified feature vector set is constructed to achieve feature sharing and form a stable multi-task fertilization model. S72. On the basis of feature sharing, an independent task representation is generated for each fertilization task, and each fertilization task retains its own characteristics while sharing features; based on the design of task-specific encoding modules, feature subsets suitable for specific tasks are automatically identified and utilized when dealing with different tasks; S73. In multi-task collaborative learning, an interaction module between tasks is introduced to capture the potential correlation between different tasks, generate a joint feature representation, and through the weighting mechanism of the interaction module, the learning process of a task is fed back to the feature update of other tasks; S74. Based on the multi-task learning framework, the Bayesian optimization method is used to automatically tune the key parameters, construct the parameter space, and gradually update the parameter distribution according to the feedback data of the fertilization task; S75. Using a proxy model in Bayesian optimization to quickly evaluate different parameter combinations, simulating feedback during fertilization through the proxy model, wherein the proxy model is updated based on historical data and learns the optimal parameter configuration from less experimental data; S76. After the parameter tuning is completed, the optimal parameter configuration set is applied to the multi-task collaborative model to achieve the overall optimization of the fertilization strategy. Through the integrated optimized multi-task collaborative learning model, each fertilization task can obtain the best fertilization effect under different environments.

Images