Multi-agent collaborative driving residual agricultural film self-adaptive regulation strategy optimization method

Abstract

The present application relates to a kind of multi-agent collaborative driving residual agricultural film self-adaptive regulation strategy optimization method, belong to the cross field of agricultural environment management and intelligent optimization decision, including integration farmland multi-source space-time data constructs multidimensional evaluation index system and parameterized regulation strategy library;Using KAN network establishes environmental feature perception agent, realizes evaluation index weight self-adaptive dynamic adjustment;Using Mamba model constructs time sequence rhythm response agent, generates operation access window and operation intensity boundary dual dynamic constraint;Through multi-agent collaborative calibration mechanism, resource allocation and spatial patch optimization are completed, and resource overload and spatial incompatibility penalty signal are output;By the integrated weight of each plot decision agent, iteration optimization is output, and the optimal regulation strategy is output.The present application can significantly improve the environmental adaptability, time sequence accuracy and regional collaboration of residual agricultural film management, reduce resource waste and operation cost, and is suitable for different climate, soil and tillage system under residual agricultural film precision management.

Description

Technical Field

[0001] This invention relates to an optimization method for adaptive regulation strategy of residual agricultural film driven by multi-agent collaboration, belonging to the interdisciplinary field of agricultural environmental governance and intelligent optimization decision-making. Background Technology

[0002] The control strategies for residual agricultural film are influenced by environmental conditions such as climate and soil, as well as agricultural time patterns such as crop systems and agricultural rhythms. Regional characteristics collectively affect the focus, composition, and timing of strategy optimization. How to intelligently identify regional characteristics in differentiated environments and adaptively optimize control strategies has become a key technological bottleneck for achieving precise management of residual agricultural film.

[0003] Existing optimization methods for residual agricultural film management strategies mostly employ multi-objective optimization algorithms such as NSGA-II, PSO, and AHP-TOPSIS. However, they generally suffer from two shortcomings: First, existing algorithms cannot dynamically adjust the optimization objective according to changes in regional characteristics, making the optimized strategies difficult to adapt to different environmental conditions. Second, existing algorithms lack the ability to model dynamic responses to temporal evolution, neglecting the periodic changes in the residual film's state caused by seasonal agricultural activities, resulting in optimized strategies failing to meet the management requirements of different agricultural stages. Therefore, there is an urgent need to construct a management strategy optimization method that can sense regional environmental differences and respond to seasonal changes, enabling the management strategy to adaptively match local environmental conditions and agricultural production patterns, thereby improving the scientific and intelligent level of residual agricultural film management. Summary of the Invention

[0004] This invention aims to solve the problems of existing residual agricultural film control strategies being unable to dynamically adapt to regional environments for optimization and adjustment, unable to accurately respond to agricultural time rhythms, and prone to conflict in resource allocation and fragmentation of work space. It proposes a multi-agent collaborative driven adaptive control strategy optimization method for residual agricultural film.

[0005] The technical solution of this invention:

[0006] A multi-agent collaborative approach to optimize the adaptive regulation strategy of residual agricultural film involves constructing a plot-based decision-making agent with a single farmland plot as the smallest decision-making unit. A distributed multi-agent collaborative architecture is adopted, where each plot-based decision-making agent integrates an environmental feature perception agent and a temporal rhythm response agent. Global strategy optimization is achieved through a collaborative calibration mechanism among the multiple agents. The method includes the following steps:

[0007] S1. Integration and Standardization of Multi-Source Spatiotemporal Data of Farmland: Construct a three-level spatial hierarchy system of region, farm, and plot; use a geographic information system to extract the geographic vector boundary of the plot and construct the spatial adjacency matrix of the plot; collect the background attribute data of the plot, the time series data of environmental status, the phenology and rhythm data of agricultural crops, and the regional engineering resource data including the number of agricultural machines and the frequency of labor; define the upper limit of regional operation resource capacity based on this; after spatial registration and vectorization processing, interpolate and resample continuous variables, classify and encode discrete variables and perform extreme value normalization processing to form a multi-source spatiotemporal feature vector set;

[0008] S2. Construction of Multidimensional Evaluation Index System and Parametric Coding of Control Strategies: Construct a comprehensive evaluation system with multiple indicators covering process indicators and result indicators, establish a sample set of control strategies including source reduction, process substitution, pollution blocking, precise recycling, classification treatment and resource utilization, and parametrically encode the strategies to form strategy parameter vectors.

[0009] S3. Construction of an adaptive weighting model for evaluation indicators based on an environmental feature-sensing agent: Using the base attribute data of the land parcel and the time series data of the environmental state as input, an environmental feature-sensing agent is constructed using a Kolmogorov-Arnold KAN network. A learnable mapping relationship is established between spatiotemporally heterogeneous environmental features and evaluation indicator weights, and a dynamic indicator weight vector is generated. The dynamic indicator weight vector serves as the evaluation basis for land parcel regulation strategies.

[0010] S4. Construction of a dynamic constraint model for regulation strategy based on time-series rhythmic response agent: Using environmental state time-series data and agricultural phenological rhythm data as input, a time-series rhythmic response agent is constructed using the selective state space model Mamba, the implicit working environment state is extracted, and the working admission window in the time dimension and the working intensity boundary in the physical dimension are generated as the time-series and physical dual dynamic constraints of the regulation strategy.

[0011] S5. Resource allocation and spatial contiguous calibration based on multi-agent collaborative calibration mechanism: Using regional engineering resource data and land parcel spatial adjacency matrix as input, a multi-agent collaborative calibration mechanism is constructed, including a resource allocation consistency model and a spatial consistency penalty term. Through the message passing mechanism between multiple agents, staggered operation is achieved, and the resource overload penalty term and spatial consistency penalty term are output as dynamic calibration signals, which are input into the comprehensive scoring function of the land parcel decision-making agent in real time.

[0012] S6. Multi-agent-driven comprehensive decision-making and optimal output of regulation strategy: Each plot decision agent integrates dynamic index weight vector, temporal and physical dual dynamic constraints and multi-agent collaborative calibration mechanism, calculates comprehensive score and iteratively converges, and outputs the optimal residual agricultural film adaptive regulation strategy at the plot level.

[0013] Specifically, in step S1,

[0014] The baseline attribute data of the land parcel includes slope, elevation, baseline values ​​of soil physical and chemical properties, and climate zone characteristics, which are non-time-varying features;

[0015] The environmental status time series data includes daily average temperature, precipitation, wind speed, real-time soil moisture content, and frost depth;

[0016] The agricultural phenological and rhythmic data include crop sowing period, harvesting period, cropping system and duration of agricultural film mulching;

[0017] The regional engineering resource data includes the amount of agricultural machinery in stock, the efficiency of machinery operation, and the frequency of labor dispatch.

[0018] Spatial registration adopts the WGS84 / UTM unified coordinate system.

[0019] Specifically, in step S2, process indicators include membrane material properties, laying density, recycling frequency and energy consumption intensity; result indicators include removal rate, treatment cost, resource utilization rate and carbon emission reduction benefits. The process indicators and result indicators are normalized based on the multi-source spatiotemporal feature vector set obtained in step S1.

[0020] Specifically, in step S3, the Kolmogorov-Arnold KAN network constructs an environmental feature-aware agent as follows:

[0021] S31. Utilizing the characteristic of KAN networks to configure learnable nonlinear activation functions on the edges, the activation functions are parameterized by using a linear combination of basis functions and B-splines:

[0022]

[0023] in, For trainable weights, For B-spline basis functions, Let be the input scalar independent variables on the network edges in the KAN network architecture. The control coefficient is k, which is the upper limit of the number index of B-spline basis functions. Its value is determined by the preset mesh density and spline order.

[0024] Using the land parcel baseline attribute data and environmental state time series data from step S1 as input, a nonlinear mapping model from input environmental characteristics to indicator influence is constructed based on the superposition theorem:

[0025]

[0026] in, For input features, To represent the influence of the indicator, q is the index of the output layer neuron, p is the index of the input layer neuron, and N is the index of the input layer neuron.in The environmental feature input dimension, i.e., the total number of neurons in the input layer, N. out The output dimension for the indicator's influence is the total number of neurons in the output layer, x. p Input environment feature vector The p-th feature component in (Xenv) is the overall nonlinear mapping function constructed by the KAN network from the environmental feature space to the influence space of evaluation indicators;

[0027] Automatically capture the nonlinear dependency relationship between environmental factors and agricultural film management evaluation indicators, and construct an environmental and indicator response matrix;

[0028] S32. Extract the nonlinear function features on each edge of the KAN network, quantitatively analyze the marginal contribution of the current environmental factors to each evaluation index, and convert the environmental perception and identification results into an evaluation index weight vector through a normalization calculation formula. The evaluation index weight vector is used as the evaluation basis input for the subsequent decision-making agent.

[0029] The normalized calculation formula for the transformation is as follows:

[0030]

[0031] in, For the first in the response matrix The response value of each indicator, To smooth the temperature coefficient, For the final generated first The weight of each indicator is denoted by n, which represents the total number of indicators participating in the regulation and evaluation.

[0032] Specifically, step S4 includes:

[0033] S41. Using the environmental state time-series data and agricultural phenological rhythm data from step S1 as input vectors, the selective scanning mechanism (SSM) of Mamba is used to extract nonlinear features from long-term agricultural information. Recursive calculations are performed using the state evolution equation to transform the cumulative effect of historical time-series information into a deep semantic description of the current plot's physical environment, outputting the work suitability score at time t in real time. The state evolution equation is:

[0034]

[0035]

[0036] in, This represents the dynamic feature vector input at time t. Indicates the state from the previous moment. With the current input The hidden layer representation generated by the interaction is as follows: At, Bt, and C are the selective system parameter matrices, and yt is the job suitability score output at time t.

[0037] S42. Based on the changing trajectory of the job suitability score, automatically determine the feasible time interval of the strategy by setting a judgment threshold:

[0038]

[0039] in, This indicates the job access window. As a feasibility threshold; combining static soil properties and real-time environmental conditions, the maximum operational intensity within the current window is generated through a mapping function:

[0040]

[0041] in, This refers to the static characteristic data of the land parcel. This serves as the boundary for work intensity.

[0042] S43. The generated operation admission window and operation intensity boundary are used as dual constraints and input into the land parcel decision-making agent in real time.

[0043] Specifically, in step S5, the system utilizes a resource allocation consistency model to achieve rational resource allocation among multiple agents. By reading the existing resource capacity limit within the region, it performs real-time conflict detection on the governance plans proposed by the decision-making agents of each plot. The function that quantitatively describes the resource overload penalty... The calculation formula is as follows:

[0044]

[0045] in, Indicates the first The intelligent agent of each plot of land Parameters of the preliminary treatment plan to be formulated at any time; Represents the resource demand function; This indicates the upper limit of the operation resource capacity defined in the area engineering resource definition in step S1; This indicates the total number of plots of land within the collaborative operation area; This means that the overflow is calculated only when the total demand exceeds the limit; otherwise, the penalty is 0.

[0046] when At that time, the collaborative calibration module uses the message passing mechanism between multiple agents to feed back the global resource occupancy penalty signal to the decision-making agent of a specific plot that has resource competition or workspace conflict, driving it to stagger its operation or switch to a non-scarce resource model.

[0047] Specifically, the formula for calculating the space consistency penalty term in step S5 is as follows:

[0048]

[0049] in, Represents a plot of land as determined by the NetworkX topology map. The first-order neighborhood set; This represents the topological adjacency weight. If the parcels are physically adjacent and unobstructed, the weight is 1; otherwise, it is 0. This represents the weighted compatibility distance of the policy vectors between adjacent land parcels.

[0050] Specifically, step S6 includes:

[0051] S61. The land parcel decision-making agent extracts candidate solutions from the strategy parameter vector in step S2 and instantiates them as the execution plan for the current land parcel; it calls the internally integrated regional feature perception agent from step S3 to obtain the evaluation index weight vector, and calls the internally integrated temporal rhythm response agent from step S4 to obtain the operation admission window and operation intensity boundary. Combined with the dynamic calibration signal from step S5, a comprehensive scoring function is constructed:

[0052] in, Indicates the candidate strategy solution in the th... Standardized performance values ​​on each evaluation index; This indicates the generation of the first generation by the KAN network in step S3. The dynamic weight vector of each indicator; This indicates an internal physical constraint penalty when the preset job time deviates from the job admission window. Or the intensity of the work exceeds the boundary of the work intensity. At that time, this item increased; Indicates a collaborative calibration penalty term. This represents a resource competition penalty. When the required agricultural machinery or human resources exceed the regional inventory limit defined in step S1, this item reduces the score through exponential growth, causing the agent to operate off-peak.

[0053] S62. Each plot of land decision-making agent independently runs the comprehensive scoring function to score and correct candidate strategies. Through information interaction among multiple agents, the optimization direction is dynamically adjusted until the control strategies of all agents converge to the optimal solution that achieves the best comprehensive governance effect and is compatible with engineering resources.

[0054] S63. The site decision-making agent locks the strategy combination with the highest comprehensive score, decodes it into standardized operation instruction output, and obtains the specific execution plan of the optimal operation time window, membrane material and equipment selection, and rated operation intensity parameters for the corresponding site.

[0055] The beneficial effects of this invention are:

[0056] This invention proposes a multi-agent collaborative approach to optimize the adaptive regulation strategy of residual agricultural film. The method constructs a decision-making agent centered on farmland plots. By integrating a KAN network and a Mamba model, it achieves a deep mapping from environmental feature perception to precise adaptation to agricultural rhythms. Furthermore, by combining a multi-agent interactive and collaborative mechanism, it solves the problems of resource allocation conflicts and fragmented operational space. The core advantages of this invention are mainly reflected in the following aspects:

[0057] (1) A method for environmental feature perception and adaptive weighting of indicators based on KAN network is proposed. Addressing the problem of inaccurate governance evaluation caused by the high heterogeneity of farmland environment, this method utilizes the characteristic of Kolmogorov-Arnold (KAN) network to configure learnable nonlinear activation functions on edges, achieving adaptive mapping from plot background attributes and real-time environmental states to evaluation indicator weights. Compared to fixed weight allocation, this method can self-learn the differences in climate, soil, and management conditions in different regions, dynamically optimizing the target weight vector, enabling the evaluation model to accurately identify micro-environmental pressures such as a surge in soil adhesion during the rainy season or an increased risk of photodegradation due to high temperatures. This method significantly improves the adaptability of the evaluation system to complex farmland environments, ensuring that the generation logic of control strategies is always rooted in the real-time physical boundary constraints of the plot, thus enhancing the scientific nature of decision-making.

[0058] (2) A dynamic constraint method for temporal rhythm response and work intensity based on the Mamba model is proposed. Addressing the engineering failure problem caused by fuzzy agricultural work windows and fixed work intensity, the selective scanning mechanism of the Selective State-Space Model (Mamba) is utilized to achieve efficient feature extraction of long-period temporal information. Through deep scanning of agricultural phenological rhythms and environmental state flows, the system can accurately pinpoint the optimal work window. It also dynamically adjusts the upper limit of physical execution intensity based on real-time semantic states such as soil moisture and temperature. Compared to traditional models such as recurrent neural networks, this method effectively filters out redundant interference signals during non-operational periods, achieving precise synchronization between operation timing and agricultural rhythms. Simultaneously, through a real-time negative feedback mechanism, it avoids physical damage to the tillage structure caused by excessive operation under extreme environmental fluctuations, ensuring the operability of the management measures.

[0059] (3) A multi-agent collaborative resource allocation and spatial operation contiguous optimization method is proposed. Addressing the resource competition conflicts and fragmented operation paths that may arise from independent decision-making on a single plot, an iterative interaction mechanism among multiple agents is constructed to achieve closed-loop calibration of the control strategy at the regional level. The system uses plot agents as the decision-making bodies, introducing resource consistency constraints and spatial collaboration penalties during the optimization process. Message passing between agents solves the problem of wasted scheduling of engineering resources such as agricultural machinery and manpower. Compared to isolated plot optimization schemes, this method, while respecting the differences in plot environments, induces adjacent plots to achieve process path compatibility, significantly improving the spatial aggregation and large-scale operation efficiency of the governance strategy. This invention successfully realizes the regionalized analysis of residual agricultural film governance, which is of great significance for reducing regional governance costs and improving agricultural film recycling efficiency. Attached Figure Description

[0060] Figure 1 This is a flowchart of the method of the present invention;

[0061] Figure 2 A heatmap showing the dynamic evolution of the weights of the evaluation indicators for land parcel intelligent agents throughout the year in accordance with agricultural rhythms;

[0062] Figure 3 Weight mapping diagram of environmental feature perception and evaluation indicators for land parcel intelligent agents in the first week of July;

[0063] Figure 4 A schematic diagram of the selective mechanism of the Mamba algorithm for the temporal response of land parcel intelligent agents;

[0064] Figure 5 A diagram showing the access window for intelligent entity residual film treatment operations on land parcels;

[0065] Figure 6 Real-time negative feedback response diagram of the task intensity boundary for intelligent agents to perceive environmental evolution;

[0066] Figure 7 The optimization effect of policy distribution before multi-agent spatial collaborative constraints;

[0067] Figure 8 The diagram shows the optimization effect of policy distribution after multi-agent spatial collaborative constraints. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of this invention, and not all of them.

[0069] Example 1:

[0070] This embodiment provides a multi-agent collaborative driven adaptive regulation strategy optimization method for residual agricultural film. It constructs a distributed optimization architecture integrating environmental feature perception, temporal rhythm response, and multi-agent collaborative optimization. By fusing KAN weight mapping and Mamba state space modeling, it achieves integrated regulation of dynamic indicator weighting, operational access window identification, and regional resource collaborative scheduling. This provides an innovative solution for intelligent decision-making and spatiotemporal consistency optimization of residual agricultural film management schemes. Figure 1 As shown, the method of the present invention includes:

[0071] S1: Integration and feature standardization of multi-source spatiotemporal data of farmland

[0072] This invention, based on regional farmland distribution maps and plot vector boundary data, utilizes ArcGIS and QGIS platforms to extract plot spatial outlines and construct a three-tiered spatial hierarchy of "region-farm-plot". The collected data encompasses multi-source information including climate, topography, soil, crops, agricultural film properties, and management facilities. The collected data is categorized into four main types according to its physical attributes and decision-making function:

[0073] The first category is the baseline attribute data of land parcels, covering topographic factors at the parcel level, such as slope and elevation; benchmark values ​​of soil physicochemical properties at the farm level; and climate zone characteristics at the regional level, which serve as non-time-varying features constraining the physical background. The second category is environmental state time-series data, including high temporal resolution meteorological flow data throughout the year, such as daily average temperature, precipitation, and wind speed, as well as soil microenvironment monitoring values, such as real-time moisture content and frost depth, used to drive the dynamic updating of evaluation weights and the real-time assessment of operational suitability. The third category is agricultural phenology and rhythm data, covering key phenological nodes in the crop growth cycle, such as event-type data like sowing and harvesting periods, cropping systems, and duration of agricultural film covering, used to define the logical entry window for remediation operations. The fourth category is regional engineering resource data, covering the engineering resource capacity within the remediation area, including the number of agricultural machines, the efficiency of machinery operations, and the frequency of available labor dispatch.

[0074] In terms of data processing, spatial data are spatially registered and vectorized using a unified coordinate system (WGS84 / UTM). Continuous variables, such as temperature, precipitation, and soil moisture, are interpolated and resampled to eliminate spatial resolution differences between different data sources. Discrete variables, such as crop type, membrane material type, and phenological nodes, are classified, coded, and standardized. To eliminate the influence of dimensions, extreme value normalization is used for feature standardization, ultimately forming a comparable and modelable multi-source spatiotemporal feature vector set.

[0075] S2. Construction of Multidimensional Evaluation Index System and Parameterized Coding of Control Strategies

[0076] To achieve quantitative evaluation and strategy optimization of agricultural film management, this invention constructs a comprehensive multi-indicator evaluation system encompassing both process and outcome dimensions, based on the full life-cycle characteristics and management process of agricultural film. Process indicators characterize the use, substitution, and recycling of agricultural film, including film material properties, laying density, recycling frequency, and energy consumption intensity. Outcome indicators assess management effectiveness and environmental impact, covering key elements such as removal rate, management cost, resource recovery rate, and carbon emission reduction benefits. Each indicator is normalized based on the standardized characteristic data obtained in step S1 to ensure comparability and universality under different regional and seasonal conditions.

[0077] Based on this, this invention, combining typical governance practices with regional differences, establishes a sample set of control strategies covering source reduction, process substitution, pollution blocking, precise recycling, classified treatment, and resource utilization. To achieve the structured nature of the strategies and the computability of the algorithms, each strategy is parameterized and encoded according to its operation method, technological process, and resource requirements, forming a strategy parameter vector that can be input into the optimization model. This vector describes the technical path and resource consumption of various control schemes in a unified format, achieving standardized, computable, and comparable expression of the control strategies, providing data support for subsequent multi-objective optimization and spatiotemporal collaborative decision-making.

[0078] S3. Construction of an adaptive weighting model for evaluation metrics of regional feature-aware agents

[0079] This invention uses the land parcel baseline attribute data and environmental state time series data obtained in step S1 as joint inputs, combined with the evaluation index system constructed in step S2, and establishes a learnable mapping relationship between spatiotemporally heterogeneous environmental characteristics and evaluation index weights by constructing a regional feature perception intelligent agent, thereby achieving adaptive dynamic optimization of the governance target weights of different land parcels.

[0080] First, the system constructs an environmental feature-aware agent using a Kolmogorov-Arnold network (KAN). Leveraging the KAN network's property of configuring learnable nonlinear activation functions on the edges, its activation functions... The parameterization is represented by a linear combination of basis functions and B-splines:

[0081]

[0082] in, For trainable weights, For spline basis functions, The system uses the static environmental background feature data from step (1) as input and constructs a control coefficient based on the superposition theorem. To the influence of indicators Nonlinear mapping model, Let k be the input scalar independent variable on the network edge in the KAN network architecture, and k be the upper bound of the number index of B-spline basis functions, the value of which is determined by the preset mesh density and spline order.

[0083]

[0084] The above model automatically captures the complex nonlinear dependence between environmental factors such as soil properties, topography, and climate zones and evaluation indicators for agricultural film management. Based on this, an environment-indicator response matrix is ​​constructed to quantitatively characterize the impact of specific environmental backgrounds on indicators such as management costs, residual film removal rate, and soil disturbance.

[0085] Based on the constructed environment-indicator response matrix, the system extracts the nonlinear function features of each edge in the KAN network, quantitatively analyzes the marginal contribution of current environmental factors, such as high slope and low temperature, to various evaluation indicators, such as residual film removal rate and cost-effectiveness, and transforms the complex environmental perception and identification into specific evaluation indicator weight vectors. The normalized calculation formula for this transformation is as follows:

[0086]

[0087] in, For the first in the response matrix The response value of each indicator, To smooth the temperature coefficient, For the final generated first Weight of each indicator.

[0088] This weight vector can be dynamically adjusted according to changes in the environmental characteristics of the land parcel, ensuring that the same set of governance strategies can produce differentiated optimal results in different regions, serving as the input for evaluation by the subsequent decision-making agent.

[0089] S4. Construction of a dynamic constraint model for regulation strategies based on agricultural season response agents.

[0090] This invention uses the environmental state time series data and agricultural phenological rhythm data obtained in step S1 as the core input, and combines the plot background attribute data as the physical background. By constructing a time series rhythm response agent, it explores the relationship between the evolution of the working environment state and the agricultural production cycle, and endows the control strategy with the constraint features of the working window and the upper limit of the strategy execution intensity.

[0091] First, the system uses the Selective State-Space Model (Mamba) to extract the implicit operating environment state. The system uses dynamic meteorological flow and crop phenological characteristics as input vectors. The selective scanning mechanism (SSM) of Mamba is used to extract nonlinear features from long-term agricultural information. The state evolution equation is expressed as follows:

[0092]

[0093]

[0094] in, This represents the dynamic feature vector input at time t, such as time-series data like precipitation, accumulated temperature, and crop growth period. This indicates that it is based on the state of the previous moment. With the current input The hidden layer representation generated by the interaction of factors. Through recursive calculation, it can transform the cumulative effect of historical precipitation, accumulated temperature and other temporal information into a deep semantic description of the current physical environment of the plot, such as soil moisture content and crop cover density; , representing the selective system parameter matrix, whose values ​​vary with the input data. It updates in real time according to changes, and is used to filter redundant information during non-operational periods and capture key agricultural nodes; This represents the job suitability score output at time t, used to quantify the friendliness of the environmental state to governance actions at that time.

[0095] Secondly, based on the evolution logic of the environmental state, the system generates admission windows in the time dimension. Strength boundary with physical dimension The system scores based on suitability. The trajectory of change is determined by setting a threshold. Automatically define the implementable time frame for the strategy:

[0096]

[0097] This indicates the optimal management window (e.g., avoiding the planting period or locking in the post-harvest window). This is the feasibility threshold. This constraint ensures that the strategy is implemented within a period permitted by agricultural laws and climatic conditions.

[0098] parameter It is not treated as a simple temporal characteristic, but rather as a physical execution limit that dynamically changes with environmental conditions. The system incorporates static soil properties. With real-time environmental status The maximum job intensity within the current window is generated using a mapping function:

[0099]

[0100] This refers to static characteristic data of the land parcel, such as soil type and slope. This sets an upper limit for the intensity of operations, such as the maximum depth to which recycling machinery can penetrate. This parameter evolves dynamically with changes in soil conditions to prevent physical damage to the site structure caused by over-operation in sensitive conditions such as excessively wet or hard soil.

[0101] Finally, the generated and As a real-time input decision-making agent with dual constraints, the generated residual agricultural film control strategy can be accurately adapted to the dynamic operating environment under the agricultural rhythm.

[0102] S5. Resource allocation and spatial contiguous calibration based on multi-agent cooperative mechanism

[0103] This invention uses regional engineering resource data and land parcel spatial adjacency matrix obtained in S1 as input. By constructing a multi-agent collaborative calibration mechanism, it aims to quantify the resource competition conflict and spatial operation fragmentation degree of the preliminary scheme of a single land parcel at the execution level, and provide key negative feedback calibration signals for subsequent strategy optimization.

[0104] First, the system utilizes a resource allocation consistency model to achieve rational resource allocation among multiple agents. The system reads the existing upper limits of manpower, machinery, and consumable resources within the region. Real-time conflict detection is performed on the governance plans initially proposed by each decision-making agent. Among these, a function quantitatively describes the resource overload penalty. The calculation formula is as follows:

[0105]

[0106] The parameters have the following meanings: : indicates the first The intelligent agent of each plot of land The parameters of the preliminary treatment plan to be formulated at any time include the selected machinery model, the length of the working window and the intensity; : Represents the resource demand function, used to calculate the amount of specific resources such as manpower and machine shifts required for a particular strategy; : indicates the upper limit of the operational resource capacity defined in the regional engineering resource definition in step (1), such as the total number of schedulable recycling machines in the region; : Indicates the total number of plots of land within the collaborative operation area; This indicates that the overflow is calculated only when the total demand exceeds the limit; otherwise, the penalty is 0.

[0107] when At this time, the collaborative calibration module uses a message passing mechanism among multiple agents to send the global resource occupancy penalty signal back to the relevant agents, driving them to stagger their operations or switch to non-scarce resource models.

[0108] Secondly, to reduce agricultural machinery losses and manpower waste during relocation, this invention utilizes the NetworkX algorithm library to construct a regional plot topology map based on the plot's geographic vector boundaries and generates a weighted adjacency matrix. The system does not simply strive for adjacent plots to have identical strategies, but rather calculates a spatial consistency penalty term. :

[0109]

[0110] The meanings of the parameters are as follows: : Represents a plot of land determined by the NetworkX topology map. The first-order neighborhood set; : Indicates the topological adjacency weight. If the plots are physically adjacent and unobstructed, the weight is 1; otherwise, it is 0. : Represents the weighted compatibility distance of strategy vectors between adjacent plots. This distance increases the weight of "operational process conflicts," such as mechanical vs. manual.

[0111] The resource overload penalty calculated in this step Space incompatibility penalty This will be used as a dynamic calibration signal and input into the comprehensive scoring function of the decision-making agent in step S6 in real time.

[0112] S6. Multi-agent driven regulation strategy integrated decision-making and optimal output

[0113] The residual agricultural film regulation system described in this invention adopts a distributed architecture. Each plot-level decision-making agent integrates the environment adaptive weight module (step S3) and the temporal physical constraint module (step S4), and achieves lateral interaction between agents through the collaborative calibration mechanism (step S5). This step aims to describe how each decision-making agent integrates internal and external information, converges through multiple rounds of iteration, and finally outputs a plot-level regulation strategy set.

[0114] First, the agent extracts candidate solutions from the policy vector library in step S2 and instantiates them as the execution plan for the current plot. When evaluating this plan, the agent invokes the internally integrated KAN network to obtain the value judgment criteria (weights). ), and call the internally integrated Mamba model to obtain the physical execution boundary (window) With strength (and combine the collaborative penalty signals from neighboring agents and the resource manager to calculate a comprehensive score.) :

[0115]

[0116] The meanings of each parameter are as follows: : indicates that the candidate strategy is in the th... Standardized performance values ​​on each evaluation index; : indicates the first generation generated by the KAN network in step S3 The dynamic weight vector elements of each indicator; : Indicates an internal physical constraint penalty term, which occurs when the strategy's preset job time deviates from the window. Or the execution intensity exceeds the boundary At that time, this item increased; : Indicates the collaborative calibration penalty term generated in step S5. Resource competition penalty: When the agricultural machinery or human resources required by the strategy exceed the upper limit of the area's inventory defined in step S1, this item reduces the score through exponential growth, forcing the agent to operate off-peak.

[0117] Secondly, during the decision-making process, each agent independently runs the aforementioned evaluation function to score and correct candidate strategies. As the interaction information in step S5 is continuously updated, i.e. The dynamic changes automatically adjust the optimization direction, so that the control strategies of all intelligent agents ultimately converge to achieve the goal of optimal comprehensive governance and engineering resource compatibility.

[0118] Finally, the decision-making agent identifies the strategy combination with the highest final score and decodes it into a standardized work instruction output, obtaining the optimal work time window for that specific plot, such as start and end dates, selection of membrane material and machinery, and rated work intensity parameters, such as tillage depth and traction force, etc.

[0119] Example 2:

[0120] This embodiment takes Farm A in Northeast China as the research object to verify the policy adaptive matching accuracy, spatiotemporal resource collaborative optimization capability, and dynamic convergence performance of the algorithm based on multi-agent cooperation of the present invention.

[0121] (1) Data acquisition and regional feature extraction

[0122] This embodiment takes Farm A in Northeast China as a typical sample area, constructs a basic dataset for the treatment of residual agricultural film with multiple types and time series, conducts comprehensive feature extraction for 10,000 standard plots in the sample area, and performs standardization processing.

[0123] First, baseline attribute data of the land parcels were collected as time-invariant physical background constraints. A 30m resolution digital elevation model (DEM) was used to extract topographic elevation and slope information at the parcel level. Combined with GF-2 and Sentinel-2 satellite imagery, the vector boundaries of the parcels were constructed and spatial contours extracted using the ArcGIS platform. Simultaneously, soil databases were accessed to obtain baseline values ​​for soil texture, organic matter content, and pH at the parcel level, and the climatic zone characteristics of the sample area were determined.

[0124] Secondly, the system integrates high-temporal-resolution environmental condition time-series data. This data is obtained by real-time access to meteorological flow data from meteorological stations in the sample area, acquiring time-series indicators such as daily average temperature, precipitation, wind speed, and light intensity. Simultaneously, it utilizes collected soil microenvironment monitoring values, including real-time soil moisture content and permafrost depth evolution data. These dynamic data are updated daily to accurately capture sudden weather changes and fluctuations in soil physical conditions.

[0125] To define the logical entry window for management operations, agricultural phenological and rhythmic data were further integrated. Through satellite imagery remote sensing interpretation and ground phenological history, the system recorded in detail the key growth nodes of major crops in the sample area, including the sowing period in early May, the harvest period in early October, and the cropping system. In addition, the data also included event-based data such as the duration of agricultural film coverage, the film coverage ratio, and the amount of residual film at different stages.

[0126] To address constraints at the engineering execution level, this embodiment collected regional engineering resource data. By collecting data on Farm A and its surrounding area, the existing stock of residual film treatment machinery in the region was quantified, and the upper limits of operational efficiency for different equipment under specific soil physical conditions were collected. Simultaneously, data including the frequency of dispatchable labor at each stage of the farm, the upper limit of the treatment budget, and the spatial distribution and relocation routes of existing residual film recycling stations were integrated.

[0127] Finally, all collected multi-source heterogeneous data were uniformly fed into the preprocessing engine for standardization. Spatially, the WGS84 / UTM coordinate system was used to perform projection transformation and spatial overlay registration on all vector and raster data. In terms of attributes, interpolation was used to eliminate spatial resolution differences for continuous meteorological and soil variables, while classification coding was performed for discrete variables such as crop type and farming methods. Ultimately, extreme value normalization was used to eliminate the influence of different physical dimensions, integrating static environmental features, dynamic agricultural flows, and engineering resource parameters into a multi-source spatiotemporal feature vector set that can be used by the model, providing robust data support for subsequent strategy optimization based on regional feature adaptive matching.

[0128] (2) Construction of farmland evaluation index and regulation strategy system

[0129] Based on the multi-source farmland characteristic data obtained in step (1), this embodiment constructs a multi-dimensional evaluation index system for the whole life cycle management of agricultural film covering, which serves as the evaluation benchmark for the multi-objective optimization model.

[0130] This system, based on two core dimensions—"process" and "outcome"—transforms collected static background data and dynamic agricultural time-series data into quantifiable evaluation inputs. Process indicators primarily characterize the implementation features of governance actions, including the physical properties of the membrane material, such as thickness, tensile strength, laying density, and expected recycling frequency. Outcome indicators focus on governance effectiveness and environmental response, covering elements such as residual film removal rate, governance cost, and resource utilization rate. The system utilizes pandas and numpy modules in Python to clean and standardize various monitoring, statistical, and policy document data, ensuring consistency across different sources in terms of time and space. Building upon this, an initial correlation matrix and baseline weights between indicators were established using the Analytic Hierarchy Process (AHP) combined with industry expert experience, forming a three-level matrix structure of "stage-indicator-weight," providing a basic evaluation framework for the subsequent adaptive weight evolution of the KAN network.

[0131] To address the specific needs of residual agricultural film regulation in the sample areas, this embodiment simultaneously established a regulation strategy library covering six main aspects: source reduction, process substitution, pollution blocking, precise recycling, classification treatment, and resource utilization. The system utilizes the geopandas module to map strategy parameters to plot spatial information and performs parameterized coding based on operational methods, technical processes, and resource requirements. After consistency verification, the strategy system comprises 6 main aspects, 20 core strategies, and 48 callable sub-strategies, capable of covering the entire process of governance from source control to end-of-pipe resource utilization in typical areas, enabling strategy migration and dynamic adaptation under multiple crop and regional conditions.

[0132] (3) Adaptive weighting of evaluation indicators based on the perception of environmental features of land parcel intelligent agents

[0133] After constructing a multidimensional evaluation system, this embodiment uses an environmental feature perception agent constructed through a KAN network to realize the mapping from the heterogeneity of the farmland environment to the adaptive weighting of evaluation index weights.

[0134] The system first uses the baseline attribute data of the land parcel collected in step (1), such as topographic slope, elevation, and soil physicochemical properties, and the time series data of environmental conditions, such as meteorological flow, real-time ground temperature, and moisture content monitoring values, as a joint input vector. Utilizing the characteristic of the KAN network to configure learnable nonlinear activation functions on the edges, the system uses a linear combination of basis functions and B-splines to parametrically characterize the complex relationship between environmental factors and governance indicators, thereby constructing an environment-indicator response matrix to quantitatively characterize the contribution of specific environmental backgrounds to indicators such as governance costs, residual film removal rate, and soil disturbance intensity.

[0135] The output of the self-learning perception and recognition mechanism is as follows: Figure 2 As shown. Figure 2 A year-round heatmap showing the dynamic evolution of evaluation index weights throughout the agricultural cycle is presented. The graph reveals that the system accurately captures the unique agricultural rhythms of the cold Northeast region: during the spring windy season (March-April) and the autumn-winter monsoon season (October-November), the increased wind speed in the environmental data leads to "wind-induced membrane rupture risk" and "implementation feasibility" exhibiting deep red high-weight zones; while during the severe winter season from December to February of the following year, the low temperatures cause a sharp increase in the weights of "frozen soil workability index" and "energy consumption index," prompting the system to forcibly reduce the priority of strategy recommendations for unconventional operation periods.

[0136] To further explain the environmental driving mechanism of weights Figure 3 The first week of July was selected as a typical micro-slice, and a mapping diagram of environmental background and indicator weights was drawn. The inner ring represents the environmental state input for the current period, and the outer ring represents the indicator weight output calculated by the KAN network. As shown in the figure, this week coincides with the rainy season and high temperature period in Northeast China. In the inner ring, "precipitation," "average temperature," and "light intensity" occupy the main sectors (environmental state values ​​of 2.17, 1.04, and 0.96, respectively). Driven by this environmental condition, the system automatically identifies the "photolysis intensity index" (weighted 0.96) in the outer ring as a strong response indicator of light intensity, and maps the "soil adhesion coefficient" (weighted 0.89) as a key feedback of precipitation. At the same time, the higher light intensity in summer significantly increases the "photolysis intensity index" (weighted 0.96); while the "permafrost workability index" is compressed to an extremely low value (weighted 0.05) due to the inconsistency caused by the high temperature mapping in summer.

[0137] Finally, the system normalizes the response values ​​of each indicator in the KAN network response matrix and adjusts them using a smoothing temperature coefficient, ultimately generating a dynamic indicator weight vector for land parcel-level governance. This mechanism not only enables intelligent mapping of farmland environmental characteristics to key evaluation indicators, but also provides an evaluation basis with environmental awareness for the subsequent decision engine to search for the optimal control strategy in a multi-objective space.

[0138] (4) Modeling of regulation strategies based on the rhythmic response of land parcel intelligent agents and dynamic environmental constraints

[0139] like Figure 4 As shown, after determining the weights of the environmental perception indicators, this embodiment constructs a temporal rhythm response agent and uses the selective state-space model (Mamba) to achieve precise constraints on the execution time window and the upper limit of the workload of the control strategy.

[0140] The system first integrates the agricultural phenological and rhythmic data from step (1), such as sowing and harvesting nodes, cover duration, and environmental state time-series data, such as meteorological flow, soil temperature, and moisture content monitoring values, as a joint input vector. It then deploys a Mamba pre-trained model with a selective scan mechanism (SSM) using Python's PyTorch framework. The agent recursively calculates the cumulative effect of historical precipitation, accumulated temperature, and other time-series information, transforming it into a deep semantic description of the current plot's physical environment, such as soil moisture content and crop cover density, and generates an operational suitability score in real time. Furthermore, the selective parameter matrix can be updated in real time as the input data changes, effectively filtering out redundant interference during non-operational periods in long-term agricultural information and accurately capturing key agricultural nodes.

[0141] Based on the fitness score evolution logic output by Mamba, this embodiment automatically generates the time-dimensional admission window by calling the threshold determination functions of the NumPy and SciPy modules. Strength boundary with physical dimension The system sets a feasibility threshold. It automatically defines the feasible range of strategies, ensuring that governance actions avoid sensitive periods such as sowing seeds. Meanwhile, parameters... Instead of being set as a fixed constant, it serves as a dynamic physical execution upper limit that changes with environmental conditions, combining static characteristics of the land parcel, such as soil viscosity and real-time semantic state. The maximum job intensity within the current window is generated through a mapping function.

[0142] The output of the time-rhythm response agent is as follows Figure 5 As shown. Figure 5 The graph shows the suitability score scan curve of Northeast Farm A during the entire autumn operating cycle using the Mamba model. As can be seen from the graph, the system accurately filters out interference signals from non-operating periods using a selective scanning mechanism: during the crop maturity period in late September, the suitability score consistently remained below the low threshold; while as crop harvesting ended in early October and soil temperature conditions stabilized, the suitability score... Rapidly climbed and broke through the threshold The model automatically identified the period from October 5th to October 25th as the optimal operating window. The definition of this window fully considers the climate zone characteristics in the background attributes of the land parcel, accurately avoiding early obstacles to mechanical entry and later risks of freezing.

[0143] To further elucidate the dynamic constraint mechanism of the strength boundary, Figure 6 A typical period of sudden weather events within the operational window (October 12th to 15th) was selected as a micro-slice, and a dynamic response map of soil condition and operational intensity was plotted. As shown in the figure, during this period, a sudden short-duration heavy rainfall occurred, and the soil volumetric water content surged from 18% to 35% within 12 hours (corresponding to the peak value of the blue curve in the figure). Driven by this state evolution, the state equations within the Mamba model captured the decreasing trend of soil bearing capacity in real time and rapidly lowered the physical intensity boundary. (Corresponding to the red stepped lines in the diagram). The system forcibly reduces the rated depth of heavy-duty recycling equipment from the conventional 15cm to below 5cm, and slowly restores operational intensity during the soil moisture receding phase after rainfall. This real-time negative feedback mechanism of "state-intensity" ensures that the remediation efforts can maintain continuity while minimizing physical damage to the topsoil structure under extreme weather disturbances.

[0144] Finally, the generated and As a real-time input decision engine with dual hard constraints, the final residual agricultural film control strategy has strong operability.

[0145] (5) Resource allocation and space operation calibration optimization based on multi-agent collaborative interaction

[0146] This embodiment takes the fourth type of regional engineering resource data and the spatial topology relationship of the plot integrated in step (1) as input, and aims to solve the resource competition conflict and work space fragmentation problem that may occur at the execution level of the preliminary plan for a single plot by constructing a multi-agent collaborative calibration mechanism.

[0147] The system utilizes Python's Pyomo modeling framework to construct a resource allocation consistency model, performing real-time conflict detection on the initial governance plans proposed by agents across 10,000 plots within Farm A. The system reads the existing resource capacity limits within the area, including manpower, machinery (such as film rolling machines and residual film recycling machines), and governance budget. Using constraint equations, each agent is subjected to constraints within the same time window. The resource occupancy within the area is cumulatively verified. When the demand for agricultural machinery during a specific period exceeds the cooperative's existing inventory, the collaborative calibration module will send a resource occupancy penalty signal to the relevant plot's intelligent agent, driving it to stagger peak times or replace machinery within a feasible time window, thereby achieving the optimal balance of human and material resources at the regional level.

[0148] Meanwhile, in order to reduce the losses of agricultural machinery during relocation and the waste of manpower due to the different management methods, this embodiment uses the NetworkX library to construct a topological adjacency matrix between plots. The system does not simply strive for identical strategies, but rather uses a space consistency penalty term. The agent guides neighboring plots to select a combination of strategies compatible with the technology path. During the interaction process, the agent will evaluate the additional costs caused by technological conflicts with neighboring plots, such as the time loss from frequently changing mounted equipment, and tend to calibrate to a compatible solution that can be executed under the same work cluster.

[0149] Spatial coordination effects before and after calibration, such as Figure 7 , Figure 8 As shown. Figure 7 The original scheme, without collaborative calibration, resulted in a highly discrete and random distribution of governance strategies across different plots. This meant that agricultural machinery needed to frequently move between different plots and constantly switch operating modes. After collaborative calibration, as... Figure 8 As shown, the governance strategy spatially resulted in a clear trend of concentrated and contiguous zones, integrating previously fragmented operational units into continuous "reduction" or "recycling" operational blocks. Table 1 further quantifies this scale effect: after collaborative calibration, the spatial clustering of each strategy significantly increased, with the "reduction (A)" and "substitution (B)" strategies showing the largest increases, reaching 40.37% and 23.52%, respectively. This indicates that the system successfully guided the transformation of dispersed individual decisions into orderly regional collaboration.

[0150] Table 1: Summary of Concentration Comparison and Improvement of Various Strategies

[0151] Strategy Before constraints After constraint Improvement rate A. Reduction 0.191 0.518 40.37% B Substitution 0.184 0.376 23.52% C blocking 0.19 0.37 22.21% D Recycling 0.19 0.333 17.69% E-sorting 0.192 0.341 18.53% F Resource Utilization 0.188 0.294 13.09%

[0152] (6) Intelligent optimization and result output of residual agricultural film regulation strategy

[0153] During the decision engine operation phase, this embodiment uses the environment adaptive weight vector generated in step (3). The timing admission window generated in step (4) With dynamic physical boundary The system uses the deap evolutionary computation library in Python to build a comprehensive evaluation model that integrates "environmental requirements - physical constraints - global collaboration". Through parallel search and multiple rounds of iterative evolution among multiple agents, under the premise of satisfying the upper limit of resource capacity and the boundary of work intensity, the optimal combination of control strategies that meet the heterogeneity characteristics of each plot of environment is searched.

[0154] The system's output strategy schemes exhibit significant intensive zoning characteristics: in the central production area with flat terrain and high spatial connectivity, the system recommends a large-scale "reduction" and "mechanical recycling" compatible strategy to reduce the cost of agricultural machinery relocation by utilizing contiguous operations; while in the peripheral areas with large terrain undulations, the system adaptively matches a manual recycling scheme with lower work intensity.

[0155] Finally, the decision engine generated a standardized set of precise governance instructions for Farm A's 10,000 plots. Each instruction was parameterized and coded to specify the exact start time of the operation, the execution method (such as the type of machinery), the rated workload, and the resource consumption parameters, forming an intelligent closed loop from multi-objective optimization to plot-level implementation suggestions.

Claims

1. A method for optimizing an adaptive control strategy for residual agricultural film driven by multi-agent collaboration, characterized in that, A decision-making agent for each farmland plot is constructed as the smallest decision-making unit. A distributed multi-agent collaborative architecture is adopted, in which each plot decision-making agent integrates an environmental feature perception agent and a temporal rhythm response agent. Global policy optimization is achieved through a collaborative calibration mechanism among multiple agents, including the following steps: S1. Integration and Standardization of Multi-Source Spatiotemporal Data of Farmland: Construct a three-level spatial hierarchy system of region, farm, and plot; use a geographic information system to extract the geographic vector boundary of the plot and construct the spatial adjacency matrix of the plot; collect the background attribute data of the plot, the time series data of environmental status, the phenology and rhythm data of agricultural crops, and the regional engineering resource data including the number of agricultural machines and the frequency of labor; define the upper limit of regional operation resource capacity based on this; after spatial registration and vectorization processing, interpolate and resample continuous variables, classify and encode discrete variables and perform extreme value normalization processing to form a multi-source spatiotemporal feature vector set; S2. Construction of Multidimensional Evaluation Index System and Parametric Coding of Control Strategies: Construct a comprehensive evaluation system with multiple indicators covering process indicators and result indicators, establish a sample set of control strategies including source reduction, process substitution, pollution blocking, precise recycling, classification treatment and resource utilization, and parametrically encode the strategies to form strategy parameter vectors. S3. Construction of an adaptive weighting model for evaluation indicators based on an environmental feature-sensing agent: Using the base attribute data of the land parcel and the time series data of the environmental state as input, an environmental feature-sensing agent is constructed using a Kolmogorov-Arnold KAN network. A learnable mapping relationship is established between spatiotemporally heterogeneous environmental features and evaluation indicator weights, and a dynamic indicator weight vector is generated. The dynamic indicator weight vector serves as the evaluation basis for land parcel regulation strategies. S4. Construction of a dynamic constraint model for regulation strategy based on time-series rhythmic response agent: Using environmental state time-series data and agricultural phenological rhythm data as input, a time-series rhythmic response agent is constructed using the selective state space model Mamba, the implicit working environment state is extracted, and the working admission window in the time dimension and the working intensity boundary in the physical dimension are generated as the time-series and physical dual dynamic constraints of the regulation strategy. S5. Resource allocation and spatial contiguous calibration based on multi-agent collaborative calibration mechanism: Using regional engineering resource data and land parcel spatial adjacency matrix as input, a multi-agent collaborative calibration mechanism is constructed, including a resource allocation consistency model and a spatial consistency penalty term. Through the message passing mechanism between multiple agents, staggered operation is achieved, and the resource overload penalty term and spatial consistency penalty term are output as dynamic calibration signals, which are input into the comprehensive scoring function of the land parcel decision-making agent in real time. S6. Multi-agent-driven comprehensive decision-making and optimal output of regulation strategy: Each plot decision agent integrates dynamic index weight vector, temporal and physical dual dynamic constraints and multi-agent collaborative calibration mechanism, calculates comprehensive score and iteratively converges, and outputs the optimal residual agricultural film adaptive regulation strategy at the plot level.

2. The method for optimizing the adaptive control strategy of residual agricultural film driven by multi-agent cooperative action according to claim 1, characterized in that, In step S1, The baseline attribute data of the land parcel includes slope, elevation, baseline values ​​of soil physical and chemical properties, and climate zone characteristics, which are non-time-varying features; The environmental status time series data includes daily average temperature, precipitation, wind speed, real-time soil moisture content, and frost depth; The agricultural phenological and rhythmic data include crop sowing period, harvesting period, cropping system and duration of agricultural film mulching; The regional engineering resource data includes the amount of agricultural machinery in stock, the efficiency of machinery operation, and the frequency of labor dispatch. Spatial registration adopts the WGS84 / UTM unified coordinate system.

3. The method for optimizing the adaptive control strategy of residual agricultural film driven by multi-agent cooperative action according to claim 1, characterized in that, In step S2, process indicators include membrane material properties, laying density, recycling frequency and energy consumption intensity; result indicators include removal rate, treatment cost, resource utilization rate and carbon emission reduction benefits. The process indicators and result indicators are normalized based on the multi-source spatiotemporal feature vector set obtained in step S1.

4. The optimization method for adaptive regulation strategy of residual agricultural film driven by multi-agent cooperative action according to claim 1, characterized in that, In step S3, the Kolmogorov-Arnold KAN network constructs an environmental feature-aware agent as follows: S31. Utilizing the characteristic of KAN networks to configure learnable nonlinear activation functions on the edges, the activation functions are parameterized by using a linear combination of basis functions and B-splines: ； in, For trainable weights, For B-spline basis functions, Let be the input scalar independent variables on the network edges in the KAN network architecture. The control coefficient is k, which is the upper limit of the number index of B-spline basis functions. Its value is determined by the preset mesh density and spline order. Using the land parcel baseline attribute data and environmental state time series data from step S1 as input, a nonlinear mapping model from input environmental characteristics to indicator influence is constructed based on the superposition theorem: ； in, For input features, To represent the influence of the indicator, q is the index of the output layer neuron, p is the index of the input layer neuron, and N is the index of the input layer neuron. in The environmental feature input dimension, i.e., the total number of neurons in the input layer, N. out The output dimension for the indicator's influence is the total number of neurons in the output layer, x. p Input environment feature vector The p-th feature component in (Xenv) is the overall nonlinear mapping function constructed by the KAN network from the environmental feature space to the influence space of evaluation indicators; Automatically capture the nonlinear dependency relationship between environmental factors and agricultural film management evaluation indicators, and construct an environmental and indicator response matrix; S32. Extract the nonlinear function features on each edge of the KAN network, quantitatively analyze the marginal contribution of the current environmental factors to each evaluation index, and convert the environmental perception and identification results into an evaluation index weight vector through a normalization calculation formula. The evaluation index weight vector is used as the evaluation basis input for the subsequent decision-making agent. The normalized calculation formula for the transformation is as follows: ； in, For the first in the response matrix The response value of each indicator, To smooth the temperature coefficient, For the final generated first The weight of each indicator is denoted by n, which represents the total number of indicators participating in the regulation and evaluation.

5. The method for optimizing the adaptive control strategy of residual agricultural film driven by multi-agent cooperative action according to claim 1, characterized in that, Step S4 specifically includes: S41. Using the environmental state time-series data and agricultural phenological rhythm data from step S1 as input vectors, the selective scanning mechanism (SSM) of Mamba is used to extract nonlinear features from long-term agricultural information. Recursive calculations are performed using the state evolution equation to transform the cumulative effect of historical time-series information into a deep semantic description of the current plot's physical environment, outputting the work suitability score at time t in real time. The state evolution equation is: ； ； in, This represents the dynamic feature vector input at time t. Indicates the state from the previous moment. With current input The hidden layer representation generated by the interaction is as follows: At, Bt, and C are the selective system parameter matrices, and yt is the job suitability score output at time t. S42. Based on the changing trajectory of the job suitability score, automatically determine the feasible time interval of the strategy by setting a judgment threshold: ； in, This indicates the job access window. As a feasibility threshold; combining static soil properties and real-time environmental conditions, the maximum operational intensity within the current window is generated through a mapping function: ； in, This refers to the static characteristic data of the land parcel. This serves as the boundary for work intensity. S43. The generated operation admission window and operation intensity boundary are used as dual constraints and input into the land parcel decision-making agent in real time.

6. The optimization method for adaptive regulation strategy of residual agricultural film driven by multi-agent cooperative action according to claim 1, characterized in that, In step S5, the system utilizes a resource allocation consistency model to achieve rational resource allocation among multiple agents. By reading the existing resource capacity limit within the region, it performs real-time conflict detection on the governance plans proposed by the decision-making agents of each block. This includes a function that quantitatively describes the resource overload penalty. The calculation formula is as follows: ； in, Indicates the first The intelligent agent of each plot of land Parameters of the preliminary treatment plan to be formulated at any time; Represents the resource demand function; This indicates the upper limit of the operation resource capacity defined in the area engineering resource definition in step S1; This indicates the total number of plots of land within the collaborative operation area; This means that the overflow is calculated only when the total demand exceeds the limit; otherwise, the penalty is 0. when At that time, the collaborative calibration module uses the message passing mechanism between multiple agents to feed back the global resource occupancy penalty signal to the decision-making agent of a specific plot that has resource competition or workspace conflict, driving it to stagger its operation or switch to a non-scarce resource model.

7. The method for optimizing the adaptive control strategy of residual agricultural film driven by multi-agent cooperative action according to claim 1, characterized in that, The formula for calculating the spatial consistency penalty term in step S5 is as follows: ； in, Represents a plot of land as determined by the NetworkX topology map. The first-order neighborhood set; This represents the topological adjacency weight. If the parcels are physically adjacent and unobstructed, the weight is 1; otherwise, it is 0. This represents the weighted compatibility distance of the policy vectors between adjacent land parcels.

8. The optimization method for adaptive regulation strategy of residual agricultural film driven by multi-agent cooperative action according to claim 1, characterized in that, Step S6 specifically includes: S61. The land parcel decision-making agent extracts candidate solutions from the strategy parameter vector in step S2 and instantiates them as the execution plan for the current land parcel; it calls the internally integrated regional feature perception agent from step S3 to obtain the evaluation index weight vector, and calls the internally integrated temporal rhythm response agent from step S4 to obtain the operation admission window and operation intensity boundary. Combined with the dynamic calibration signal from step S5, a comprehensive scoring function is constructed: ; in, Indicates the candidate strategy solution in the th... Standardized performance values ​​on each evaluation index; This indicates the generation of the first generation by the KAN network in step S3. The dynamic weight vector of each indicator; This indicates an internal physical constraint penalty when the preset job time deviates from the job admission window. Or the intensity of the work exceeds the boundary of the work intensity. At that time, this item increased; Indicates a collaborative calibration penalty term. This represents a resource competition penalty. When the required agricultural machinery or human resources exceed the upper limit of the area's inventory defined in step S1, this item reduces the score through exponential growth, causing the agents to operate at off-peak times. S62: Each plot's decision-making agent independently runs the comprehensive scoring function to score and correct candidate strategies. Through information interaction among multiple agents, the optimization direction is dynamically adjusted until the control strategies of all agents converge to the optimal solution that achieves the best comprehensive governance effect and is compatible with engineering resources. S63. The site decision-making agent locks the strategy combination with the highest comprehensive score, decodes it into standardized operation instruction output, and obtains the specific execution plan of the optimal operation time window, membrane material and equipment selection, and rated operation intensity parameters for the corresponding site.

Images