Methods and related equipment for forest spatial structure regulation driven by action decoding intelligent agents
The forest spatial structure regulation method driven by action decoding agents, which uses agents and management simulators for rolling closed-loop simulation, solves the problem of low accuracy in forest spatial structure optimization, realizes dynamic multi-step intelligent decision-making, and improves the accuracy of forest spatial structure regulation.
Patent Information
- Application Number
- CN202611619278.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, the accuracy of forest spatial structure optimization is low, and static single-moment optimization makes it difficult to predict the cumulative shift of subsequent structural indicators.
A forest spatial structure regulation method driven by action decoding agents is adopted. By acquiring forest stand status information, structural index calculation and action decoding processing are performed. Evolutionary reasoning is carried out using a preset agent, and dynamic simulation is performed in combination with a preset management simulator. Rolling closed-loop deduction is performed for several management cycles to generate a target management decision sequence.
It effectively alleviates the problem of cumulative deviation of structural indicators caused by static single-moment optimization, improves the accuracy of forest spatial structure regulation, and realizes dynamic multi-step intelligent decision-making.
Smart Images

Figure CN123117871A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of combining artificial intelligence with forest multifunctional structure optimization, and in particular to a method and related equipment for forest spatial structure regulation driven by an action decoding intelligent agent. Background Technology
[0002] Forest spatial structure is a core attribute describing the spatial configuration, interspecific relationships, and competitive patterns of individual trees within a forest stand, directly impacting stand stability, productivity, and ecosystem service functions. In forests, such as secondary forests, insufficient mixed planting, clustering of individual tree species, and imbalances in diameter class differentiation are major structural obstacles hindering the evolution of forest stands towards a near-natural state during restoration. Related technologies primarily utilize static, single-moment optimization for selective logging spatial optimization, but this approach struggles to predict the cumulative shifts in subsequent structural indicators, resulting in low accuracy in forest spatial structure optimization.
[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention
[0004] The main objective of this application is to propose a forest spatial structure regulation method and related equipment driven by an action decoding agent, which can alleviate the problem of cumulative deviation of structural indicators caused by static single-moment optimization and effectively improve the accuracy of forest spatial structure regulation.
[0005] To achieve the above objectives, one aspect of this application proposes a forest spatial structure regulation method driven by an action decoding agent, the method comprising: Obtain stand status information of the secondary forest to be optimized; Based on the forest stand status information, structural indicators are calculated to obtain structural assessment index data; Based on the structural evaluation index data, a preset intelligent agent performs action decoding processing to generate operational action information; Based on the management action information, dynamic simulation is performed through a preset management simulator to update the forest stand status information, and then the step of calculating structural indicators based on the forest stand status information to obtain structural evaluation indicator data is returned, so as to carry out rolling closed-loop extrapolation for several management cycles. If the number of management cycles reaches the preset step length, the spatial structure optimization task of the secondary forest to be optimized ends, and a target management decision sequence is generated.
[0006] In some embodiments, before performing the calculation of structural indicators based on the stand status information to obtain structural assessment indicator data, the method further includes: Several structural evaluation index parameters are determined from the preset structural dimensions; wherein, the structural evaluation index parameters include mixing degree, angular scale, size ratio, competition index, and openness; A comprehensive evaluation index for forest stand spatial structure is constructed based on the structural evaluation index parameters and a preset multi-objective comprehensive decision-making algorithm.
[0007] In some embodiments, the step of generating operational action information by performing action decoding processing through a preset intelligent agent based on the structural evaluation index data includes: A state summary vector is obtained by mapping the structural evaluation index data; The preset intelligent agent performs reasoning analysis on the state summary vector to generate the business action information.
[0008] In some embodiments, the step of generating the operational action information by reasoning and analyzing the state summary vector through the preset intelligent agent includes: Based on the state summary vector, the preset intelligent agent makes decision proposals to obtain operational decision information; wherein, the operational decision information includes action type data and logging intensity data; Based on the action type data and the logging intensity data, a preset removal score is calculated according to a preset action library vector. Then, based on the preset removal score, a target tree is filtered to generate a desired logging table.
[0009] In some embodiments, before performing the dynamic simulation based on the management action information using a preset management simulator to update the forest stand status information, the method further includes: A growth simulation sub-model was constructed based on the potential-modified individual tree diameter increment model; Construct a death simulation sub-model based on the logistic regression model; Construct and update sub-models based on the forest window model; The preset management simulator is constructed by coupling the growth simulation sub-model, the death simulation sub-model, and the update sub-model according to the preset management process steps; wherein, the preset management process steps include logging, growth, death, and update.
[0010] In some embodiments, the step of dynamically simulating the forest stand status information based on the management action information using a preset management simulator, and then updating the forest stand status information, includes: The operational information is decoded, and then the corresponding operational actions are executed through the preset operational simulator to obtain operational simulation results. A forest stand snapshot is generated based on the management simulation results, and the forest stand status information is updated based on the forest stand snapshot.
[0011] To achieve the above objectives, another aspect of this application proposes a forest spatial structure regulation system driven by an action decoding agent, the system comprising: The first module is used to obtain stand status information of the secondary forest to be optimized; The second module is used to calculate structural indicators based on the forest stand status information to obtain structural assessment indicator data. The third module is used to perform action decoding processing on the structural evaluation index data through a preset intelligent agent to generate operational action information. The fourth module is used to dynamically simulate the management action information through a preset management simulator, thereby updating the forest stand status information, and returning the step of calculating structural indicators based on the forest stand status information to obtain structural evaluation indicator data, so as to carry out rolling closed-loop deduction for several management cycles. The fifth module is used to terminate the spatial structure optimization task of the secondary forest to be optimized and generate the target management decision sequence if the number of management cycles reaches the preset step length.
[0012] To achieve the above objectives, another aspect of this application provides an electronic device, the electronic device comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.
[0013] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.
[0014] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0015] The embodiments of this application include at least the following beneficial effects: This application provides a method, system, electronic device, storage medium, and program product for forest spatial structure regulation driven by an action decoding intelligent agent. This scheme obtains the stand state information of the secondary forest to be optimized, calculates structural indicators based on the stand state information to obtain structural evaluation index data, and then uses a preset intelligent agent to perform evolutionary reasoning to obtain management action information. Next, based on the management action information, a preset management simulator is used for dynamic simulation, thereby updating the corresponding stand state information and returning the steps of calculating structural indicators based on the stand state information to obtain structural evaluation index data. This process is repeated for several management cycles in a rolling closed-loop simulation, thereby achieving long-term dynamic management simulation and alleviating the problem of cumulative deviations in subsequent structural indicators that are difficult to predict in static optimization. Finally, if the number of management cycles reaches a preset step size, the spatial optimization task of the secondary forest to be optimized ends, and a target management decision sequence is generated, completing the spatial structure optimization of the secondary forest to be optimized. It is easy to understand that the rolling closed-loop decision-making mechanism of the present invention, which analyzes and reasons through a preset intelligent agent based on structural evaluation index data and combines it with a preset management simulator for dynamic simulation, can effectively alleviate the problem of cumulative deviation of structural indicators in the static single-moment optimization framework, realize dynamic multi-step intelligent decision-making, and effectively improve the accuracy of forest spatial structure regulation. Attached Figure Description
[0016] Figure 1 This is a flowchart of the steps of the forest spatial structure regulation method driven by the action decoding intelligent agent provided in the embodiments of this application; Figure 2 This is a schematic diagram of the overall process architecture of the forest spatial structure regulation method driven by the action decoding intelligent agent provided in the embodiments of this application; Figure 3 This is a schematic diagram of a five-dimensional polygon provided in an embodiment of this application; Figure 4 This is the same six-item bar chart of the five-dimensional evidence provided in the embodiments of this application; Figure 5 This is a schematic diagram of the empirical cumulative distribution of the four nearest neighbor distances provided in the embodiments of this application; Figure 6 This is a schematic diagram of the d4 histogram and kernel density curve provided in the embodiments of this application; Figure 7 This is a scatter plot of the input robustness verification provided in the embodiments of this application; Figure 8 This is the M-drift frequency histogram after randomly deleting 10% of individual trees, as provided in the embodiments of this application; Figure 9 This is a schematic diagram of the mixing degree trajectory of 594 sub-plots managed in 6 steps according to an embodiment of this application; Figure 10 This is a schematic diagram of the comprehensive index trajectory of the 6-step management of 594 sub-plots provided in the embodiments of this application; Figure 11 This is a schematic diagram of the competition index trajectory of 6-step operation of 594 sub-plots provided in the embodiments of this application; Figure 12 This is a schematic diagram of the openness trajectory of 6-step operation of 594 sub-plots provided in the embodiments of this application; Figure 13 This is a three-layer comparison diagram provided in the embodiments of this application; Figure 14 This is a schematic diagram of the forest spatial structure regulation system driven by the action decoding intelligent agent provided in the embodiments of this application; Figure 15 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of systems and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.
[0018] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if” or “when” as used herein may be interpreted as “when…” or “in response to determination.”
[0019] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.
[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.
[0021] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.
[0022] Forest spatial structure: refers to the process of selectively adjusting an unreasonable spatial structure through human intervention, guiding it towards an ideal structural state that can achieve management goals.
[0023] Large Language Models (LLMs) are deep learning models trained on large amounts of text data. These models are capable of generating natural language text or understanding the meaning of language text. They can perform not only simple language tasks such as spell checking and grammar correction, but also complex tasks such as text summarization, machine translation, and content recommendation. Correspondingly, action decoding agents combine large language models and action decoders to transform language understanding into executable action information.
[0024] Forest spatial structure is a core attribute describing the spatial configuration, interspecific relationships, and competitive patterns of individual trees within a forest stand, directly impacting stand stability, productivity, and ecosystem service functions. In forests, such as secondary forests, insufficient mixed planting, clustering of individual tree species, and imbalances in diameter class differentiation are major structural obstacles hindering the evolution of forest stands towards a near-natural state during restoration. Related technologies primarily utilize static, single-moment optimization for selective logging spatial optimization, but this approach struggles to predict the cumulative shifts in subsequent structural indicators, resulting in low accuracy in forest spatial structure optimization.
[0025] In view of this, this application provides a method, system, electronic device, storage medium, and program product for forest spatial structure regulation driven by an action decoding intelligent agent. This scheme obtains the stand state information of the secondary forest to be optimized, calculates structural indicators based on the stand state information to obtain structural evaluation index data, and then uses a preset intelligent agent to perform evolutionary reasoning to obtain management action information. Next, based on the management action information, a preset management simulator is used for dynamic simulation to update the corresponding stand state information and return the steps of calculating structural indicators based on the stand state information to obtain structural evaluation index data. This process is repeated for several management cycles in a rolling closed-loop simulation, thereby achieving long-term dynamic management simulation and alleviating the problem of cumulative deviations in subsequent structural indicators that are difficult to predict in static moment-to-moment optimization. Finally, if the number of management cycles reaches a preset step size, the spatial optimization task of the secondary forest to be optimized ends, and a target management decision sequence is generated, completing the spatial structure optimization of the secondary forest to be optimized. This effectively alleviates the problem of cumulative deviations in structural indicators in a static single-moment optimization framework, achieves dynamic multi-step intelligent decision-making, and effectively improves the accuracy of forest spatial structure regulation.
[0026] The forest spatial structure regulation method driven by action-decoding intelligent agents provided in this application relates to the technical field of combining artificial intelligence with forest multifunctional structure optimization. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the forest spatial structure regulation method driven by action-decoding intelligent agents, but is not limited to the above forms.
[0027] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.
[0028] Figure 1 This is an optional flowchart of the forest spatial structure regulation method driven by the action decoding agent provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S110 to S150.
[0029] Step S110: Obtain the stand status information of the secondary forest to be optimized.
[0030] Step S120: Calculate structural indicators based on forest stand status information to obtain structural assessment index data.
[0031] Step S130: Based on the structural evaluation index data, the action decoding process is performed by a preset intelligent agent to generate operational action information.
[0032] Step S140: Based on the management action information, a dynamic simulation is performed using a preset management simulator to update the stand status information, and then the structural indicators are calculated based on the stand status information to obtain structural assessment indicator data, so as to carry out rolling closed-loop simulations for several management cycles.
[0033] Step S150: If the number of management cycles reaches the preset step length, the spatial structure optimization task of the secondary forest to be optimized ends, and the target management decision sequence is generated.
[0034] In this specific embodiment, the present invention first acquires the stand status information of the secondary forest to be optimized, and then calculates structural indicators based on the stand status information to obtain structural assessment index data. Specifically, the secondary forest to be optimized in this embodiment refers to the target secondary forest that needs spatial structure optimization, i.e., the forest that needs spatial structure regulation. Correspondingly, the stand status information refers to the comprehensive performance data of the forest (stand) at a certain moment, such as data on tree structure, composition, growth, health, and regeneration capacity. In this embodiment, the stand status information of the secondary forest to be optimized is determined by initializing a stand snapshot, including tree coordinate information, tree species information, diameter at breast height (DBH) information, tree height information, and vitality value. Furthermore, the structural assessment index data in this embodiment refers to indicators used to evaluate the spatial structure of the stand, which can comprehensively evaluate the spatial structure from multiple different structural dimensions. Accordingly, this embodiment quantifies the current structural state of the secondary forest to be optimized by calculating the corresponding structural assessment index based on the stand status information, so as to facilitate subsequent spatial structure optimization. Next, this embodiment uses a preset intelligent agent to perform action decoding processing based on the structural assessment index data to generate management action information. Specifically, in this embodiment of the invention, the pre-defined intelligent agent refers to an interface abstraction layer used for action decoding, which can be implemented by any decision-maker as the backend and does not represent autonomous artificial intelligence. For example, in this embodiment of the invention, the pre-defined intelligent agent can be connected to one of the following models: Genetic Algorithm (GA), Simulated Annealing (SA), Particle Swarm Optimization (PSO), Reward Increment Algorithm (REINFORCE), and Large Language Model (LLM). Accordingly, this embodiment of the invention, through a unified interface and a unified evaluation budget, achieves for the first time a fair juxtaposition of classical numerical optimizers (GA / SA / PSO) and emerging artificial intelligence decision-makers (REINFORCE / LLM) on the same forestry task, thereby objectively revealing the inherent capability boundaries of each decision-making paradigm, rather than claiming the performance advantage of any one decision-maker.
[0035] Furthermore, in this embodiment of the invention, dynamic simulation is performed using a preset management simulator based on management action information to update the stand status information. The process then returns to the step of calculating structural indicators based on the stand status information to obtain structural assessment indicator data, thereby performing a rolling closed-loop simulation across several management cycles. Specifically, in this embodiment, the preset management simulator refers to a multi-step dynamic management simulator used for long-term dynamic management simulation of secondary forests. Correspondingly, this embodiment constructs a preset management simulator to input management action information into it and simulate the management decisions generated by a preset agent, such as simulating stand spatial structure, logging disturbance, and regeneration processes at the individual tree scale. After simulating and executing the decision actions output by the preset agent, this embodiment updates the stand status information of the secondary forest to be optimized based on the simulation results and recalculates the corresponding structural assessment indicator data to rerun the entire closed-loop simulation process, achieving rolling closed-loop simulation.
[0036] It is easy to understand that, by employing a rolling closed-loop deduction method, this embodiment of the invention can read the forest stand state updated in the previous simulation step at the beginning of each simulation step, such as each simulation year, and recalculate the corresponding structural evaluation indicators, generating a structured text summary. Then, the agent analyzes and reasons based on the latest state, outputting the corresponding management judgment and actions. In other words, the optimization of each step starts from the latest forest stand state. Therefore, the final optimization result obtained by this embodiment is not obtained through a single static optimization, but rather through a dynamic management decision sequence formed by successive state updates, agent reasoning, and simulation feedback. This effectively alleviates the problem of cumulative structural indicator shifts caused by static single-moment optimization, effectively improving the accuracy of forest spatial structure regulation, such as the accuracy of secondary forest spatial structure optimization. Finally, if the number of management rounds reaches the preset number of steps, the spatial structure optimization task of the secondary forest to be optimized ends, and the target management decision sequence is generated. Specifically, in this embodiment, the preset number of steps refers to the expected number of management steps, which can be set according to optimization needs, such as each management step corresponding to 5 years, for a total of 6 management steps (30 years). Correspondingly, when the number of management cycles reaches the expected number, such as 6 steps, the entire rolling closed-loop decision-making process is considered to have been completed, the structural optimization task is terminated, and the corresponding target management decision sequence is output, thus completing the spatial structure optimization of the secondary forest to be optimized.
[0037] To improve the accuracy of secondary forest spatial structure assessment, in some embodiments of the present invention, before performing structural index calculations based on stand status information to obtain structural assessment index data, the action-decoding agent-driven forest spatial structure regulation method provided in the embodiments of the present invention further includes, but is not limited to, the following steps: Several structural evaluation index parameters are determined from the preset structural dimensions. These parameters include mixing degree, angular scale, size ratio, competition index, and openness.
[0038] A comprehensive evaluation index for forest stand spatial structure is constructed based on structural assessment index parameters and a pre-set multi-objective integrated decision-making algorithm.
[0039] In this specific embodiment, the present invention first determines several structural evaluation index parameters from a preset structural dimension, and then constructs a comprehensive evaluation index of forest stand spatial structure based on the structural evaluation index parameters and a preset multi-objective comprehensive decision-making algorithm. Specifically, the preset structural dimensions in this embodiment include five core structural dimensions: forest stand mixing characteristics, spatial distribution pattern, degree of forest differentiation, intraspecific and interspecific competition intensity, and forest stand spatial openness. Correspondingly, the structural evaluation index parameters in this embodiment include five indicators: mixing degree, angular scale, size ratio, competition index, and openness. For example, this embodiment uses... The nearest neighbor and 1-meter edge buffer are used to calculate various structural evaluation parameters. For example, consider the target tree. Neighboring wood , For target tree The set of 4 nearest neighbors (i.e., the set of 4 nearest adjacent trees sorted by Euclidean distance) has an azimuth difference of . The distance is Adjacent wood base area In this embodiment of the invention, the degree of mixing, used to quantify the degree of tree species isolation, represents the proportion of dissimilar trees in a neighborhood, and its calculation formula is as follows: , Where, in the formula Target Tree The degree of mixing; , Target trees With the A neighboring tree species; The nearest neighbor number ( =4); This is represented as an indicator function, which takes the value 1 when the condition within the parentheses is true, and 0 otherwise; that is, when the neighboring tree... With target tree Different tree species ( When the number is 1, take 1; when the number is the same, take 0.
[0040] Meanwhile, the angular scale is used to determine the randomness and clustering of the horizontal distribution pattern, and its calculation formula is as follows: , Where, in the formula Representing the target tree angular scale, This represents the azimuth distance between adjacent trees. Accordingly, Corresponding to random distribution, Tend to cluster It tends to be uniform.
[0041] In addition, the size ratio characterizes the path-order competition position of the target tree in the neighborhood, and its calculation formula is as follows: , Where, in the formula Representing the target tree The size ratio, , These represent the target number and the diameter at breast height (DBH) of neighboring trees, respectively. Correspondingly, the size ratio... The larger the value, the more competitive the target tree is in the neighborhood.
[0042] Meanwhile, the competition index, referencing relevant competition index concepts, estimates local competitive pressure by combining the cross-sectional area and distance of adjacent timbers. Its calculation formula is shown below: , Where, in the formula Representing the target tree Competition index; Indicates the first The basalarea (m²) of the adjacent tree at breast height. 2 ); Represents the target tree With the The horizontal distance (m) between neighboring trees; in the model implementation, the competing terms are ( +0.5) m is used as the effective distance to avoid numerical amplification caused by extremely close distances; The expression represents the competition normalization constant (in this embodiment of the invention, the value is taken as...). = 0.1), The smaller the value, the lower the competitive pressure.
[0043] In addition, openness reflects the light penetration conditions and growth space of the target tree's neighborhood, and its calculation formula is shown in the following formula: , Where, in the formula , These are the tree heights of the target tree and neighboring trees, respectively. Representing the target tree The degree of openness. Accordingly, The higher the value, the less shading there is around the target tree, and the better the light transmission conditions.
[0044] Next, in this embodiment of the invention, the comprehensive evaluation of forest stand spatial structure adopts the Forest Stand Spatial Structure Comprehensive Evaluation Index (FSS) to achieve multi-indicator synergistic quantitative assessment. Accordingly, this embodiment of the invention constructs the Forest Stand Spatial Structure Comprehensive Evaluation Index (FSS), the calculation formula of which is shown below: , Where, in the formula Expressed as angular scale balance, and , ( Corresponding to an ideal random distribution, ); Indicates the degree of mixing. Indicates the ratio of large to small. Indicates openness, This indicates the competition index.
[0045] It should be noted that the embodiments of the present invention employ a weighted summation method to construct the Forest Stand Spatial Structure Comprehensive Evaluation Index (FSS), which remains monotonic even when the weights are unbalanced, exhibiting boundary robustness and interpretability of forestry prior weights, and its magnitude is consistent with relevant published data. Furthermore, the calculation formula for the Forest Stand Spatial Structure Comprehensive Evaluation Index (FSS) in the embodiments of the present invention uses... Replace the original , so that it is with , Same direction (the larger the better). Correspondingly, the ratio of large to small. Similarly, the principle of "centering is best" applies, but in this embodiment of the invention, the size ratio of the 30 large sample plots in area 30 is different. The mean is concentrated in the range of 0.71-0.75 (far from 0.5), and the ratio of large to small values is... The variance between sample plots is much smaller than If the ratio of large to small Also perform balance conversion It will be artificially compressed. This information alters the comparability of the FSS with related comprehensive scores. Therefore, embodiments of the present invention only address the diagonal scale. Perform balance conversion; size ratio The original value is retained, and its directionality is interpreted through the differentiation of diameter levels within the sample plot. It is not converted within the formula of the forest stand spatial structure comprehensive evaluation index (FSS).
[0046] To improve the accuracy of large-scale business decision-making, in some embodiments of the present invention, action decoding processing is performed by a preset intelligent agent based on structural evaluation index data to generate business action information, including but not limited to the following steps: The state summary vector is obtained by mapping the structural assessment index data.
[0047] By using a pre-defined intelligent agent to perform reasoning analysis on the state summary vector, operational action information is generated.
[0048] In this specific embodiment, the present invention first obtains a state summary vector based on structural assessment index data. Specifically, the present invention maps a collection of individual trees to a plot-scale compact state vector, i.e., a state summary vector, as shown in the following equation: , Where, each component in the formula These represent individual woods respectively. The mixed degree, angular scale, size ratio, competition index, openness, and forest stand spatial structure comprehensive evaluation index (FSS) are all considered. Represents a state summary vector. This represents a 6-dimensional real vector space (i.e., a vector space consisting of 6 real components). This represents a state summary mapping function, whose function is to map a set of single units. Mapped to compact state vectors at the plot scale .
[0049] Next, in this embodiment of the invention, the state summary vector is analyzed and reasoned by a preset intelligent agent to obtain production and operation action information. Specifically, in this embodiment of the invention, the state summary vector is input into a preset intelligent agent, such as an arbitrary decision-maker. (GA / SA / PSO / REINFORCE / LLM) proposes business actions based on the state summary vector and generates corresponding business action information.
[0050] It should be noted that when inputting to the Large Language Model (LLM) backend, this embodiment of the invention renders the state summary vector into structured state summary text. In this embodiment, the text summary data refers to multi-dimensional forest stand state data described by structured text, which can be used for contextual reasoning in the Large Language Model. In this embodiment, the large model agent refers to an agent model built based on the Large Language Model (LLM), which can receive forest stand state described by structured text, integrate multi-dimensional information for comprehensive analysis and judgment, and output management suggestions and decision-making reasons in natural language, without the need to construct a differentiable objective function or conduct extensive iterative searches. Accordingly, this embodiment of the invention performs structured processing based on forest stand state information to construct a structured text summary, which is input to a preset agent, i.e., an action decoding agent. The preset agent perceives forest stand state through natural language, deeply infers structural defects, autonomously generates management strategies, and iteratively optimizes spatial patterns, providing a new decision-making paradigm for multi-step dynamic management and determining corresponding management action information. For example, this embodiment of the invention first renders text summary data based on the state summary vector, and then uses the preset agent to analyze and reason about the text summary data to generate management action information. Furthermore, in this embodiment of the invention, text summary data is input into a preset intelligent agent, enabling the preset intelligent agent to output corresponding management actions through natural language reasoning. Additionally, the management action information output by the preset intelligent agent in this embodiment may include management judgment information, plan update information, action type information, and action intensity information corresponding to the current round. Since this embodiment uses a rolling closed-loop deduction method for dynamic management decision-making, in each management round, the preset intelligent agent analyzes the current structural assessment index data and outputs the management judgment, plan update, action type, and action intensity corresponding to the current stage (e.g., the current year or the current forecast stage). It should be noted that the decision-making rules of the preset intelligent agent in this embodiment embed the core concept of near-natural management: when the mixing degree is low (e.g., mixing degree <0.55), priority is given to promoting interspecific isolation; when local competition is too strong, priority is given to alleviating crowding; in other cases, comprehensive balance or protection of dominant trees is chosen. This tone of "moderate intervention and gradual optimization" is consistent with the principle that "management intervention should be guided by improving the spatial structure of the forest stand." Unlike numerical optimization methods that encode management objectives as mathematical functions, the agent in this embodiment of the invention understands management intentions through natural language, and its decision-making process is closer to the judgment mode of forestry workers based on experience and forestry knowledge.
[0051] To further improve the reliability of operational action selection, in some embodiments of the present invention, a preset intelligent agent performs inference analysis on the state summary vector to generate operational action information, including but not limited to the following steps: Based on the state summary vector, a pre-defined agent makes decision proposals to obtain operational decision information. This operational decision information includes action type data and logging intensity data.
[0052] Based on action type data and logging intensity data, a preset removal score is calculated according to a preset action library vector. Then, based on the preset removal score, a target tree is selected to generate a desired logging table.
[0053] In this specific embodiment, the present invention first proposes a decision based on the state summary vector through a preset intelligent agent to obtain operational decision information. Specifically, the operational decision information in this embodiment includes action type data and logging intensity data. Wherein, any decision-maker in this embodiment... (GA / SA / PSO / REINFORCE / LLM) proposes an action type based solely on the state vector. With an original strength Two low-dimensional scalars, without directly specifying a single tree. Among them, the action type... , This represents a finite set (action space) of preset operational action types. It should be noted that in this embodiment, the decision-maker is the only replaceable component, and the rest of the framework does not make any differentiable or numerical assumptions about its form. Next, this embodiment calculates a preset removal score based on the action type data and logging intensity data according to a preset action library vector, and performs target tree filtering based on this preset removal score to generate a desired logging table. Specifically, before calculating the removal score, this embodiment first performs constraint processing on the logging intensity data of the operational decision information. For example, this embodiment uses... The controlled single-step harvesting intensity is obtained by normalizing and capping the process. Next, in this embodiment of the invention, a preset removal score is calculated for each target tree according to the action-related library vector, i.e., the preset action library vector, as shown in the following formula: , Where, in the formula Representing the target tree Remove ratings This represents the crossover weight corresponding to action a. This represents the angular scale uniformity weight corresponding to action a. This indicates the appropriate weight of the size ratio corresponding to action 'a'. This represents the competition index weight corresponding to action 'a'. From action library vectors It is obtained using a scaling factor of 2.5, which can be positive or negative; , The degree of angular scale uniformity and the appropriateness of size differentiation are measured separately. For congestion items, correspondingly, since the library vectors have signs, the direction in which each indicator contributes to the decongestion score is determined by its sign; for example, "alleviating congestion" takes... , and in the formula The superposition of negative signs results in highly crowded trees receiving higher removal scores and being prioritized for felling. Additionally, openness... For monitoring / output indicators, logging selection scoring is not included. Furthermore, in this embodiment of the invention, the library vectors are manually set (based on forestry experience and near-natural management priors, non-adaptive learning). Accordingly, in this embodiment of the invention, scoring is removed according to a preset method. Take first in descending order The plant is added to the candidate harvesting set, and then a constraint layer (DBH) is superimposed. P85 is reserved by default; single-step logging intensity The final harvest table, i.e. the expected harvest table, is obtained by buffering the margins by 1 m. In addition, the minimum intervention control (OBSERVE) action is retained, with its library vector at zero and its intensity at a lower bound.
[0054] Accordingly, the action library and decoding full weights in this embodiment of the invention are shown in Table 1 below: Table 1 , in, This represents the base weight of the corresponding action on the mixing degree dimension. The basic weights of the corresponding actions in the angular scale dimension. The basic weights of the corresponding actions in the size ratio dimension. This represents the base weight of the corresponding action on the competition index dimension. In this embodiment of the invention, the library vector... through Enter the process of removing the scoring method.
[0055] It should be noted that, in some embodiments of the present invention, after receiving text summary data, the preset intelligent agent outputs a structured response containing current problem analysis, business plan updates and action selection, and selects the corresponding action type and corresponding intervening intensity from a pre-set business measure library (preset business measure library), that is, determines the corresponding expected action information.
[0056] To achieve multi-step dynamic management simulation and improve the accuracy of the simulation, in some embodiments of the present invention, before performing dynamic simulation based on management action information through a preset management simulator and updating the forest stand status information, the action decoding agent-driven forest spatial structure regulation method provided in the embodiments of the present invention further includes, but is not limited to, the following steps: A growth simulation sub-model was constructed based on the potential-modified individual tree diameter increment model; Construct a death simulation sub-model based on the logistic regression model; Construct and update sub-models based on the forest window model; The growth simulation sub-model, death simulation sub-model, and update sub-model are coupled according to the preset operation process steps to construct a preset operation simulator. The preset operation process steps include logging, growth, death, and update.
[0057] In this specific implementation, the embodiments of the present invention first construct a growth simulation sub-model based on the potential-corrected individual tree diameter-at-breast-width increment model. Specifically, the potential-corrected individual tree diameter-at-breast-width increment model in the embodiments of the present invention refers to a model used to predict the growth of the diameter-at-breast-width (DBH) of a single tree. Accordingly, the embodiments of the present invention construct a growth simulation sub-model based on this potential-corrected individual tree diameter-at-breast-width increment model to simulate the growth process of individual trees. The potential growth is based on the differential form of the von Bertalanffy growth equation, as shown in the following equation: , Where, in the formula Indicates potential growth in diameter at breast height; The intrinsic growth rate of the tree species; This represents the maximum diameter at breast height (DBH) of the tree species. Indicates the current diameter at breast height; This represents the shape parameter indicating allometric growth; in this embodiment of the invention, empirical values are used. Actual growth was adjusted using the Hegyi competitive correction factor. ( (This refers to the competition sensitivity coefficient, with a larger value for positive species) and the viability correction factor. ( Adjustments to tree vitality indicators, such as final actual diameter at breast height (DBH) growth. The calculation formula is: Among them, the tree height increment is based on the derivative form of the relevant tree height-diameter-thickness allometric equation ( , Indicates the increment of tree height. The derivative of the tree height-diameter-thickness allometric equation is given. Represents the tree height variable The differential, Represents the diameter at breast height (DBH) variable The derivatives are gradually accumulated to preserve the measured initial tree height baseline.
[0058] Simultaneously, this embodiment of the invention constructs a mortality simulation sub-model based on a logistic regression model and a regeneration sub-model based on a forest gap model. Specifically, the mortality probability in this embodiment adopts a logistic regression model, while the regeneration process references the JABOWA model framework. In this embodiment, the regeneration pool is divided into light-tolerant and shade-tolerant types; the former accumulates rapidly when the forest gap ratio is high, while the latter grows slowly under a closed canopy. Further, this embodiment couples the growth simulation sub-model, mortality simulation sub-model, and regeneration sub-model according to preset management process steps to construct a multi-step management simulator, i.e., a preset management simulator. Specifically, the preset management process steps in this embodiment include harvesting, growth, mortality, and regeneration, thus forming a "harvesting-growth-mortality-regeneration" closed loop, which is then coupled to construct a multi-step dynamic management simulator, thereby simulating the stand spatial structure, harvesting disturbance, and regeneration process at the individual tree scale.
[0059] It should be noted that in some embodiments of the present invention, state observation includes attributes such as tree location, tree species, diameter at breast height (DBH), tree height, and vitality. The action space is a continuous vector a ∈ [-1,1]. 5 The decoding is performed using four structural weights and thinning intensity. Simultaneously, canopy retention constraints are set for large-diameter individuals (individuals with a diameter at breast height ≥ 0.85 quantile are not included in the felling) to prevent canopy fragmentation caused by logging. For example, the key parameters of the simulated environment in this embodiment of the invention are shown in Table 2 below: Table 2
[0060] Meanwhile, examples of simulator sub-model parameter values in this embodiment of the invention are shown in Table 3 below: Table 3
[0061] In some embodiments of the present invention, dynamic simulation is performed using a preset management simulator based on management action information, thereby updating the stand status information, including but not limited to the following steps: The business operation information is decoded, and then the corresponding business operations are executed through a preset business simulator to obtain the business simulation results.
[0062] Forest stand snapshots are generated based on the management simulation results, and forest stand status information is updated based on these snapshots.
[0063] In this specific embodiment, the present invention first decodes the management action information, then executes the decoded management actions through a preset management simulator, and generates a stand snapshot based on the obtained management simulation results. The stand status information is then updated based on the stand snapshot. Specifically, in this embodiment, a stand snapshot refers to a three-dimensional vector data containing time and spatial dimensions, which records in detail and completely the spatial location, attribute information, and status data of each tree within the entire stand (or standard plot) at a specific moment. Correspondingly, when the agent outputs the current management action information, the present invention decodes the output management action information and inputs it into the preset management simulator for multi-step dynamic simulation to obtain the management simulation results. This results in the stand simulation results obtained after executing the management simulation according to the output management actions. The present invention also records various data information of the entire stand and generates a stand snapshot. Based on this stand snapshot, the stand status of the secondary forest to be optimized is further updated, facilitating the subsequent closed-loop extrapolation process.
[0064] The following section provides a detailed description and explanation of the embodiments of the present invention, using specific examples of forest spatial structure regulation driven by action decoding agents: For example, such as Figure 2 As shown, this embodiment of the invention constructs a general framework for action decoding agents for secondary forest management scenario simulation. Its core is a four-stage abstract interface, "State Summary," that is decision-independent, interpretable, and auditable. -Motion Classification-Intensity Constraints - Deterministic target tree filtering The arbitrary decision-maker proposes only two low-dimensional scalars: action type and intensity. Individual timber harvesting is uniquely generated from structural index scores using deterministic rules, and output as an independently verifiable decision log. This decoding layer does not depend on a differentiable objective function, allowing Genetic Algorithm (GA), Simulated Annealing (SA), Particle Swarm Optimization (PSO), Reinforcement, and Lightweight Large Language Model (LLM) to be equally integrated into the same pipeline. Correspondingly, the framework synthesizes a comprehensive index at the index layer. In the simulation layer at 5 yr The six-step explicit iterative harvesting-growth-death-update-edge correction method differs from existing single-moment optimal methods.
[0065] Meanwhile, addressing the issue of a lack of true values from long-term field retests, this invention constructs a five-dimensional indirect validation framework (recalculation consistency / internal consistency and robustness / ecological rationality / control discriminant validity / cross-method consistency) and supplements it with independent external validation of the growth sub-model through three phases of retesting in the National Forest Inventory (NFI) within the region; the main significance test is based on 30 large sample plots (each 100 m²). Paired design of 100 m aggregate mean was used to avoid spurious subplot replication. Results: Initial mixing degree The Pearson correlation coefficient between the recalculated and published values. ; Deterministic rule-based strategy operation, final state of FSS in 30 large sample plots Significantly higher than the minimum intervention control of OBSERVE (p-value of paired Wilcoxon signed-rank test) Cohen's d value Furthermore, the embodiments of this invention, through component ablation experiments, confirmed that the improvement was mainly driven by the action decoding index-driven scoring bearing—after replacement with randomly selected wood of the same intensity, approximately 80% of the improvement disappeared (8-seed robustness). This is not simply a byproduct of "natural growth after half of them are cut down". Furthermore, to verify the existence of interface-decision-independence, this embodiment of the invention allows six types of heterogeneous decision-makers (GA / SA / PSO / REINFORCE / LLM / Random) to be executed equally under the same interface and the same 144 evaluation budget (dataset 22); their performance ranking (GA / SA outperforms lightweight LLM stand-in, ...) This is a descriptive report only and does not represent the claims of any embodiment of the invention. The embodiments of the invention do not claim any performance advantage of any decision-maker; their core contribution lies in the interpretable and auditable action decoding interface and its coupling with multi-step closed-loop simulation. The embodiments of the invention position this model as an auxiliary tool for business scenario simulation and candidate solution generation, rather than a field prescription; its long-term predictive validity is limited by the lack of true values from same-site verification and cannot yet be empirically verified. Independent NFI validation within the region further indicates that the uncalibrated growth sub-model systematically underestimates the magnitude of diaphragm growth by approximately one order of magnitude (its mean bias can be corrected by one-step local calibration). Therefore, the absolute values of the 30-year (yr) trajectory in the embodiments of the invention only reflect the relative direction of structural selection under a given simulator and cannot be extrapolated to predict business output.
[0066] The lack of long-term, independently retested true values is a common challenge in secondary forest simulation modeling. However, this invention does not regard it as an insurmountable limitation, but rather elevates it to a methodological challenge, constructing a five-dimensional indirect verification evidence chain based on the forest management paradigm. Its logic is the progressive construction of "cumulative plausibility": V1 (recalculation) ensures the error-free implementation layer. V2 (robustness) guarantees that the conclusions are not random or dependent on specific parameters. V3 (Ecological Rationality) ensures that the process conforms to domain knowledge. V4 (Comparison Differentiation) ensures the model is sensitive to business operations. V5 (cross-method) guarantees that the interface is non-ad hoc (the conclusion does not depend on a single decision maker). Finally, the applicable boundaries are defined by external NFI verification. It is easy to understand that no single piece of evidence is decisive, but this progressively building and mutually independent chain of evidence can construct a defensible argument for the model's credibility. The correspondence between the five dimensions and the evidence in the embodiments of this invention is as follows: (V1) Recalculation consistency: the initial indicators can be independently reproduced from the original single-tree data ( The recalculated Pearson correlation coefficient is (V2) Internal consistency and robustness: FSS weight perturbation robust, input missing robust, and buffer scale insensitive; (V3) Ecological rationality: plant number dynamics, / The simulation process conforms to known ecological laws; (V4) the differential validity of the control, the operating scenario is significantly different from the OBSERVE minimum intervention control (Cohen d value is 1). (V5) Cross-method consistency: multiple decision-makers behave comparablely under the same interface. Beyond the five dimensions, this embodiment of the invention further uses three-phase retesting of the NFI within the region to perform independent external calibration tests on the growth sub-model, quantifying the cross-site bias of uncalibrated literature parameters and demonstrating the calibrability of their means. Accordingly, the five-dimensional indirect verification overview provided by this embodiment of the invention is as follows: Figure 3 and Figure 4 As shown. Among them, Figure 3 The original statistic (V1) is directly labeled for each axis. V2 V3 falls within the literature range, V4 V5 Ext 10 (Underestimated); the radar's radial position is merely a "transparent, monotonically normalized display" that plots these heterogeneous statistics together on a single graph, and is not an achievement / quality score or a truth statement. Meanwhile, Figure 3 The five quantitative dimensions V1 to V5 constitute a polygon; the external calibration Ext is a semi-quantitative judgment, marked separately with orange hollow dots and not included in the polygon to avoid confusion. Additionally, Figure 6 The bar chart for the same six items shows the raw statistics directly labeled in the right column. Normalization mapping rules (pre-defined, monotonic, reproducible, for display purposes only): For correlation / rank correlation classes, directly take their values (V1). V2 takes weighted perturbation ); Effect size classification Saturation mapping (V4) V3 is assigned a value of 0.85 for "falling within the literature range but not locally calibrated"; V5 is assigned a value of 0.80 for "significant main effect and method separable"; and Ext is assigned a value of 0.50 for "systematic underestimation by about one order of magnitude but mean calibrable". The validation conclusions are supported by the original statistics, and the normalized position does not carry evidence weight.
[0067] Accordingly, in this embodiment of the invention, the data from 594 sample plots are independently recalculated using a single tree table. (398 were successfully recalculated; the rest were skipped due to insufficient internal trees or incomplete fields), compared with the results document. (Mixing degree value before intervention) (Comparison with the competition index value after intervention) etc. Results showed... and The recalculation results were almost identical (Pearson correlation coefficient). Root Mean Square Error (RMSE) Competition Index Pearson correlation coefficient Root Mean Square Error (RMSE) ; Due to different dimensional mappings, there is a systematic bias ( / (This is not a direct comparison.) This evidence proves that the initial state at the plot scale used in this invention can be reproduced from the original single-tree data, and the simulation model does not introduce hidden initial disturbances. Meanwhile, as... Figure 5 and Figure 6 As shown, the distances to the first to fourth nearest neighbors were independently calculated from 87,772 target trees in 568 subplots. .result: The median is 1.416 m, and the 75th percentile is P75. 1.78 m, 95th percentile (P95) 2.50 m, meaning that 75% of the trees have their fourth nearest neighbor distance less than 1.78 meters, 95% of the trees have a distance less than 2.50 meters, and the maximum is 6.65 m. The target tree at m accounted for 83.8%. This indicates that at 20 m With a 20 m subplot and a 1 m buffer setting, the fourth nearest neighbor distance of most target trees exceeds 1 m, but is still much smaller than the plot side length (20 m). The nearest neighbor search can be completed entirely within the plot. Simultaneously, due to the "edge trees still participate in nearest neighbor search" protocol used in this embodiment, the 1 m buffer only filters a small number of samples where the target tree is "too close to the boundary," without limiting the search. Upper limit. This evidence suggests a 1 m buffer is reasonable; however, in sparse plots or later stages of intensive logging... It may still exceed 2.5 m, and future verification should be carried out in a 1 ha standard plot with a wider buffer zone.
[0068] Furthermore, in verifying internal consistency, ecological rationality, and robustness, this embodiment of the invention first performs dynamic decomposition of the plant count. During the 30-year simulation period, each subplot experienced an average of 92.3 plants harvested, approximately 7.6 simulated deaths, and approximately 3.5 simulated regenerations, ultimately retaining 86.5 plants. (Post-management plant count) Compared with the number of plants before management The ratio is This corresponds to an average operating intensity of 0.551. Meanwhile, the competition index in the simulated final state of operation... The mean increased from 0.035 to 0.060; this was not due to deliberate management failure, but rather to the thinning of the forest stand after logging and the continued increase in the diameter at breast height of the remaining trees, thus affecting the Hegyi competition index formula. With the target wood itself Simultaneously increase For the first The cross-sectional area at breast height of the competing trees. For target tree Competing wood The distance between them. This direction is compared with the minimum intervention of OBSERVE coding. Consistent, indicating that proactive management releases some competition, but still falls short of the "no intervention" control.
[0069] Correspondingly, the sensitivity to the FSS parameter is... Relative change of FSS under perturbation Largest component Secondly, this result reflects the weight design in the FSS formula. Dominant (e.g., mixed degree weight value) Regarding the FSS weight sensitivity on real data (based on 568 complete sample plots), this embodiment of the invention further performs a sensitivity test on each FSS weight across the 568 complete before-after sample plots. The disturbance was assessed, and the overall change in forest structure score of the disturbed plots was recorded. Sorting and Benchmark The Spearman correlation test was performed on the sorted data. Specifically, the Spearman rank correlation coefficients of the 10 single-weighted perturbation groups were analyzed. all ( : 0.9987 / 0.9994; Angular scale weight value 0.9999; Size ratio weighting value : 0.9999; Openness weight value : 0.9994; Non-competitive index weight value (0.9999). Further perform 100 full-weighted joint operations. Uniform random perturbation: Spearman rank correlation coefficient The mean is 0.9995, the minimum is 0.9969, and the 5th percentile (P5) is 0.9987. Conclusion: Even with weights... Any joint disturbance within the range, sample plot The order is always related to the rank of the benchmark. FSS sorting is highly robust to weight selection. This result is consistent with the results under different FSS constructions (weighted / productive / geometric). - "The formation of distinctions, i.e., significant differences between the structural forms, but robustness to weight perturbations under the same weighted summation form, indicates that the FSS formula has good robustness in its weight design. Furthermore, at the decision-making level: for the Managed vs. OBSERVE aggregated from 30 large sample plots, paired tests were performed on 1000 fully weighted joints..." Significant under random perturbation ( sampling (and all 30 large sample plots had positive pairing differences), Cohen d-value (mean) -- That is, the main conclusion that "proactive management is significantly better than minimal intervention" does not depend on any specific weight value.
[0070] At the same time, such as Figure 7 and Figure 8 As shown, in this embodiment of the invention, 10% of the individual tree weights were randomly removed from 30 sample plots during the weight calculation. Repeat 5 times; The median was 0.005, and the 95th percentile (P95) was 0.013; the difference in sample plots was 86.7%. 0.01. Among them, The value of M is calculated based on the original complete data. This is the value of M calculated after randomly deleting 10% of the individual tree data. This result supports... The indicators remain stable under moderate input loss, demonstrating input robustness. Furthermore, regarding the sensitivity of 1 m and 2 m buffers (all 6 indicators), the embodiments of this invention, based on actual recalculation of 60 sample plots, show that changing the 1 m buffer to 2 m reduces the average number of target trees per subplot by 22.4% (38 → 29 trees), but there are no significant differences in all 6 indicators at the plot scale, as shown in Table 4 below: Table 4
[0071] Among them, the paired Wilcoxon test: , , , , FSS (all Pearson correlation coefficient Accordingly, The Wilcoxon p-value represents the degree of mixing. The Wilcoxon p-value for the size ratio. The Wilcoxon p-value is the angular scale. The Wilcoxon p-value is the competition index. The Wilcoxon p-value for openness, FSS This represents the Wilcoxon p-value for the Forest Stand Spatial Structure (FSS) index. It's easy to understand that this is the p-value of the Wilcoxon signed-rank test (a non-parametric test) for all six indicators. )all , / / Pearson of FSS ; of Low but There were still no significant differences, indicating that the 1 m buffer did not produce systematic bias for any of the indicators, and expanding the buffer to 2 m did not significantly change any indicators. Therefore, setting 1 m in this study area is reasonable.
[0072] Furthermore, the embodiments of the present invention specify the growth parameters for this region. Inversion (30-year simulation based on minimal intervention using 594 OBSERVE codes): The simulated annual increase in chest diameter in this region is... mm / yr (range 0.80–2.01), falling within the 1–3 mm / yr range for subtropical evergreen broad-leaved forests. Potentially growing from Wykoff species. ( cm) is used to deduce the area (95% CI (95% confidence interval) is [0.014, 0.027], sample size) (lower than the default value) Approximately 58%. Of which, It indicates the annual growth of diameter at breast height. It represents the intrinsic growth rate of a tree species (that is, the maximum growth potential of the tree species under ideal conditions). Indicates the current chest diameter. This represents the maximum diameter at breast height (DBH) of the tree species. It should be noted that this inversion is based on OBSERVE-encoded simulation output rather than long-term field measurements and verification; therefore, it represents indirect evidence of "simulation-documentary consistency" rather than true value calibration. If future verification records for this region (5 years or more) are obtained, the data should be directly fitted. Replace the default value.
[0073] Furthermore, relevant documents provide plot structure indicators under the OBSERVE-coded minimal intervention implementation over 30 years. Here, the OBSERVE action corresponds to "no logging, only simulating growth-death-renewal," not the "true value of natural succession" in the field; in this embodiment, it is collectively referred to as the "OBSERVE-coded minimal intervention control." Accordingly, the main significance test in this embodiment uses 30 large plots (each 100m²). The aggregate mean of 100m plots is used as the unit of analysis, and the paired difference is the difference in the means of subplots within the large plot, to avoid spurious duplication caused by nested subplots. The 594 subplots are used only for descriptive statistics (final state distribution, mean). SD), not involved in significance Value calculation. Under this design, the paired Wilcoxon rank-sum test and paired Cohen's test for the main indicators are performed. The summary is shown in Table 5 below: Table 5
[0074] Note: (a) Wilcoxon for Precision under one-sided all-positive / all-negative difference Lower limit ( Multiple indicators have bottomed out due to a high degree of consistency in their dominant direction; The values should not be interpreted as exactly the same test statistic. (b) Cohen's (Pairing) Press ( The mean of the pairing differences. (Calculated as standard deviation); After aggregating 30 large sample plots, the sample size decreased, but the variance within each plot was significantly compressed, therefore... The value is too high. (c) Threshold reference: This is a small effect. A value of 0.2-0.5 falls under the category of medium effect. for - This is a large effect. This belongs to the maximum effect; (Rosenthal correlation coefficient) No longer applicable to this table). (d) Weighted balanced size ratio A negative pairing difference reflects the originality of the study area. (Preferring clustered distribution), after management A further deviation of 0.5, therefore decline.
[0075] Furthermore, the embodiments of the present invention operate under the same simulator, the same deterministic strategy, and the same single-step strength. Under the same seed (seed=42), component ablation is performed on the Action Decoding Interface (ADI): only the tested component is replaced / removed, while all other growth-depletion-update-metric-FSS calculations remain unchanged. Accordingly, the ablation yields two causal conclusions: (i) the action decoding metrics drive the scoring. Instead of "cutting down half and letting it grow naturally," it carries the responsibility of structural improvement, that is, to... After replacing it with random felling of the same intensity from Collapsed to (Approximately 80% of the improvement disappeared) Furthermore, this randomization strategy is robustly close to zero on 8 seeds (the standard deviation between seeds is only...). (i) The difference is much smaller than the full-random difference, proving that the index-driven choice of the interface is the main cause of the structural gain; (ii) The constraint layer (85th percentile (P85) of large-diameter timber is retained) is not the driving term for maximizing FSS, but an explicit ecological trade-off. Removing it slightly increases FSS ( Compared to only the difference However, this caused the initial elite wood retention rate to decrease from... Plunging to ( percentage points Elite cross-sectional area, That is, removing the constraint layer is achieved by cutting down the large-diameter dominant trees that should have been protected in exchange for that... The FSS (Free Surface Space) precisely quantifies the value of the constraint layer in "trading a small exponential cost for significant preservation of ecological elites," rather than a performance defect. This ablation transforms "easy-to-use interface" into hard evidence of "which component causally contributes to structural gains and ecological protection."
[0076] Furthermore, to verify that all six types of heterogeneous decision-makers (including traditional optimizers and lightweight LLMs) can execute equally under a unified interface and unified evaluation budget, this embodiment of the invention performs an existence proof of interface universality. Accordingly, this embodiment of the invention provides... The ranking (GA / SA outperforms lightweight LLM stand-in) is merely a descriptive observation used to illustrate the interface's ability to objectively distinguish the boundaries of different decision-making paradigms, thus constituting a fair benchmark platform. In this embodiment, the comparative experiments of the six methods were conducted on 22 datasets, with the following sampling rules, which are independently reproducible: (i) Population: The entire set of 594 subplots; (ii) Sampling method: 22 subplots were randomly selected uniformly from the entire set, and a random number seed sample_seed was set, such as a fixed random seed sample_seed=42 (determined by the hash input to the experimental program, not based on result filtering); (iii) The number of random seeds eval_seeds for evaluation was set, such as setting the number of evaluation seeds eval_seeds=2 (two random initializations were performed for each dataset); (iv) The context window truncation value max_trees for large language model prompt words was set, such as setting max_trees=64 (keeping the first 64 trees in descending DBH order), without data filtering; GA, SA, PSO, and REINFORCE used the entire subplot data. Correspondingly, the distribution of tree_count (the total number of trees in each dataset) in dataset 22 is as follows: minimum 56 trees, median 170 trees, maximum 404 trees (mean 184.7), and the median species_count (the total number of species in each dataset) is 36, which is close to the distribution of the entire dataset (Kolmogorov-Smirnov test (two-sided)). (Unbiased sampling). The original sampling parameters are recorded in a preset field and can be completely reproduced using the same sample seed. Furthermore, this embodiment of the invention does not use any post-hoc screening based on method performance; the 22 datasets serve as an independent test set for method comparison. Further, on the aforementioned 22 datasets, six decision-makers—Random, REINFORCE, PSO, LLM stand-in (qwen3:0.6b), Genetic Algorithm (GA), and Simulated Annealing (SA)—are compared. The performance on the screen. Here. The Forest Structure Index (FSS) is a reward alignment metric used internally by the simulator to drive the search of each decision-maker. It has a different name and scale than the FSS, but both improve in the same direction. The methods in this invention embodiment are used uniformly for comparison. To align with the target caliber of each optimizer, the structural evaluation elsewhere still uses FSS as the standard. GA / SA and PSO share the same action-intensity search space. With the same assessment budget (144 assessments, compared to 12 particles of PSO) (12 iters are equal), and all results are obtained from actual operation in this study area. The comparison results of the six real-world decision-making methods are shown in Table 6 below: Table 6
[0077] What is easy to understand is that the decision-maker (Incremental forest structure index) sorted in descending order; / This is a paired comparison with LLM stand-in. Conclusion: The interface successfully distinguishes the capability boundaries of "search-based optimizers (GA / SA)" and "lightweight LLMs based on zero-shot inference," which is evidence of the interface's fair benchmark capability, rather than a performance claim of either decision-maker.
[0078] In addition, in this embodiment of the invention, it is based on 5 yr 6 steps to aggregate 594 sample plots FSS , Time series (mean) 95% CI) respectively Figure 9 , Figure 10 , Figure 11 as well as Figure 12 As shown. Among them, Figures 9 to 12 All 594 sub-plots represent 6-step operational trajectories (mean). 95% confidence interval; FSS independent ordinate), horizontal axis from 0 to 30 years, gray dashed line is the baseline reference for year 0. The 95% confidence interval here is a descriptive interval at the subplot level (uncorrected for spatial autocorrelation within the plot, so it is narrow and only used to show the trend); the main significance inference is always performed using the aggregation of 30 large plots + cluster bootstrapping method.
[0079] At the same time, refer to Figure 13 , Figure 13 Compare "what the decision-maker outputs and how individuals are selected" under three paradigms: Original Numerical Optimizer (GA / SA) output Action vectors are reproducible but unreadable; the original black-box LLM directly specifies harvesting in natural language, readable but unconstrained, uncertain, and risky of illusion; the framework provided by this invention allows any decision-maker to propose only (action, intensity) + reason, which is then decoded by the deterministic rule layer into an auditable list of individual trees to harvest – thus natively possessing the four attributes of human readability, controllable constraints, reproducibility, and auditability of individual trees. It should be reasonably stated that while GA can obtain some of these attributes through external decoders and LLM through constraint post-processing, both require additional engineering assumptions and cannot inherently guarantee all four attributes simultaneously (e.g., GA decoding still struggles to guarantee auditability of individual tree semantics, and LLM verification still cannot guarantee determinism); the contribution of this framework lies in treating these four attributes as built-in guarantees of the interface rather than external patches. "Original output" specifically refers to the original output form without any domain decoders / constraint guardrails: the original numerical optimizer (GA / SA) directly outputs action vectors, and the original black-box LLM directly outputs natural language harvesting instructions. In the bottom attribute matrix: checkmark = native interface guarantee, triangle = partially obtainable only through additional engineering (non-native), cross = not provided. This framework (marked with ★) is the only paradigm that natively satisfies all four requirements simultaneously: human readability, controllable constraints, deterministic reproducibility, and auditable single-unit structure. GA / LLM can supplement some attributes through engineering patches (marked with triangles in the diagram), but cannot intrinsically guarantee all four requirements simultaneously.
[0080] It should be noted that existing spatial structure optimization studies mostly use multi-objective genetic algorithms to find the optimal state at a single moment. This invention extends this to a multi-step dynamic closed loop, explicitly characterizing the intermediate states of "execution → growth → death → update," thus transforming from static single-step optimization to dynamic multi-step feedback. Simultaneously, natural language reasoning is applied only to action type and intensity, and individual timber harvesting selection is returned to structural indicators to drive scoring, avoiding direct outputs of difficult-to-audit instructions such as "harvest timber #1xx." This mechanism allows the decision-maker to only undertake "operation-oriented" level decisions, significantly improving interpretability. Furthermore, this invention transforms from single-indicator optimization to multi-indicator comprehensive response. The FSS comprehensive index integrates mixing degree, angular scale, size ratio, competition index, and openness into a single evaluation objective, allowing the simulated trajectory to simultaneously reflect changes in multiple structural dimensions, avoiding "local optima, overall degradation" of a single indicator. Additionally, this invention converts the black-box LLM results into corresponding decision logs (action type, intensity, text summary) and saves them step-by-step in corresponding documents, enabling independent auditing of the decision-maker output afterward. For example, the audit of decision logs completed on 56 sample plot subsets constitutes interpretable specific evidence.
[0081] It is readily understood that this invention proposes a unified abstract interface for the four stages of "state summary—action classification—intensity assignment—target tree screening" adapted to forestry time-series regulation. Through this unified interface and unified evaluation budget, it achieves for the first time a fair juxtaposition of classical numerical optimizers (GA / SA / PSO) and emerging AI decision-makers on the same forestry task, objectively revealing the inherent capability boundaries of each decision paradigm rather than claiming the performance advantage of any single decision-maker. Furthermore, the action decoding process in this invention does not rely on a differentiable objective function or perform gradient backpropagation; instead, it uses spatial structure index scoring as the deterministic decision-making basis, achieving gradient-free and non-objective function-dependent dynamic regulation. In addition, this invention implements a multi-step dynamic closed loop, a six-period, five-year step-size closed-loop iteration of "harvesting—growth—death—renewal—edge correction," with explicit modeling of time-series feedback, unlike the single-moment optimality in related technologies. Moreover, each decision step (action, intensity, text summary, single-tree harvesting list) in this invention is output as an independently verifiable decision log, and individual tree harvesting is deterministically completed by structural index scoring, enabling full-process auditability.
[0082] Please see Figure 14 This application also provides a forest spatial structure regulation system driven by an action decoding agent, which can implement the above-mentioned method. The system includes: The first module 210 is used to obtain stand status information of the secondary forest to be optimized.
[0083] The second module 220 is used to calculate structural indicators based on forest stand status information to obtain structural assessment indicator data.
[0084] The third module 230 is used to perform action decoding processing on the structural evaluation index data through a preset intelligent agent to generate operational action information.
[0085] The fourth module 240 is used to dynamically simulate the forest stand status information through a preset management simulator based on the management action information, and then update the forest stand status information. It also returns the steps of calculating the structural indicators based on the forest stand status information to obtain the structural evaluation indicator data, so as to carry out rolling closed-loop simulation for several management cycles.
[0086] The fifth module 250 is used to terminate the spatial structure optimization task of the secondary forest to be optimized and generate the target management decision sequence if the number of management cycles reaches the preset step length.
[0087] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.
[0088] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.
[0089] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0090] Please see Figure 15 , Figure 15 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 310 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 320 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 320 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 320 and is called and executed by the processor 310 using the methods described in the embodiments of this application. Input / output interface 330 is used to realize information input and output; The communication interface 340 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 350 transmits information between various components of the device (e.g., processor 310, memory 320, input / output interface 330, and communication interface 340); The processor 310, memory 320, input / output interface 330 and communication interface 340 are connected to each other within the device via bus 350.
[0091] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.
[0092] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.
[0093] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0094] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0095] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
[0096] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.
[0097] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.
Claims
1. A method for regulating forest spatial structure driven by an action-decoding intelligent agent, characterized in that, The method includes the following steps: Obtain stand status information of the secondary forest to be optimized; Based on the forest stand status information, structural indicators are calculated to obtain structural assessment index data; Based on the structural evaluation index data, a preset intelligent agent performs action decoding processing to generate operational action information; Based on the management action information, dynamic simulation is performed through a preset management simulator to update the forest stand status information, and then the step of calculating structural indicators based on the forest stand status information to obtain structural evaluation indicator data is returned, so as to carry out rolling closed-loop extrapolation for several management cycles. If the number of management cycles reaches the preset step size, the spatial structure optimization task of the secondary forest to be optimized ends, and a target management decision sequence is generated; The step of generating operational action information by performing action decoding processing through a preset intelligent agent based on the structural evaluation index data includes: A state summary vector is obtained by mapping the structural evaluation index data; Based on the state summary vector, the preset intelligent agent makes decision proposals to obtain operational decision information; wherein, the operational decision information includes action type data and logging intensity data; Based on the action type data and the logging intensity data, a preset removal score is calculated according to a preset action library vector. Then, based on the preset removal score, a target tree is filtered to generate a desired logging table.
2. The method according to claim 1, characterized in that, Before performing the calculation of structural indicators based on the forest stand status information to obtain structural assessment indicator data, the method further includes: Several structural evaluation index parameters are determined from the preset structural dimensions; wherein, the structural evaluation index parameters include mixing degree, angular scale, size ratio, competition index, and openness; A comprehensive evaluation index for forest stand spatial structure is constructed based on the structural evaluation index parameters and a preset multi-objective comprehensive decision-making algorithm.
3. The method according to claim 1, characterized in that, Before performing the dynamic simulation based on the management action information using a preset management simulator to update the forest stand status information, the method further includes: A growth simulation sub-model was constructed based on the potential-modified individual tree diameter increment model; Construct a death simulation sub-model based on the logistic regression model; Construct and update sub-models based on the forest window model; The preset management simulator is constructed by coupling the growth simulation sub-model, the death simulation sub-model, and the update sub-model according to the preset management process steps; wherein, the preset management process steps include logging, growth, death, and update.
4. The method according to claim 1, characterized in that, The step of dynamically simulating the forest stand status information using a preset management simulator based on the management action information includes: The operational information is decoded, and then the corresponding operational actions are executed through the preset operational simulator to obtain operational simulation results. A forest stand snapshot is generated based on the management simulation results, and the forest stand status information is updated based on the forest stand snapshot.
5. A forest spatial structure regulation system driven by an action decoding intelligent agent, characterized in that, The system includes: The first module is used to obtain stand status information of the secondary forest to be optimized; The second module is used to calculate structural indicators based on the forest stand status information to obtain structural assessment indicator data. The third module is used to perform action decoding processing on the structural evaluation index data through a preset intelligent agent to generate operational action information. The fourth module is used to dynamically simulate the management action information through a preset management simulator, thereby updating the forest stand status information, and returning the step of calculating structural indicators based on the forest stand status information to obtain structural evaluation indicator data, so as to carry out rolling closed-loop deduction for several management cycles. The fifth module is used to terminate the spatial structure optimization task of the secondary forest to be optimized if the number of management cycles reaches a preset step size, and to generate the target management decision sequence. The step of generating operational action information by performing action decoding processing through a preset intelligent agent based on the structural evaluation index data includes: A state summary vector is obtained by mapping the structural evaluation index data; Based on the state summary vector, the preset intelligent agent makes decision proposals to obtain operational decision information; wherein, the operational decision information includes action type data and logging intensity data; Based on the action type data and the logging intensity data, a preset removal score is calculated according to a preset action library vector. Then, based on the preset removal score, a target tree is filtered to generate a desired logging table.
6. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1 to 4.
7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.
8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.
