A method and system for establishing a forestation area of a fast-growing and high-yield tree species

Abstract

The application provides a fast-growing and high-yield forest tree species afforestation area establishment method and system, and relates to the field of afforestation. The application fuses ecological environment and social economic two-dimensional information, adopts an improved maximum entropy reinforcement learning model to construct an ecological environment suitable area prediction model, determines social economic suitability based on rule modeling, and obtains a final suitable afforestation area through spatial superposition. The application realizes the intelligentization and precision of afforestation site selection, improves the scientificity and actual operability of suitability determination, and is suitable for afforestation planning and optimization under large-scale and multiple constraint conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of afforestation, in particular to a method and system for determining a suitable afforestation area for fast-growing and high-yield forest tree species. BACKGROUND

[0002] Fast-growing and high-yield forests are characterized by planting single tree species in rows and at high density, and promoting rapid growth through intensive management methods such as fertilization and irrigation. The rotation period is extremely short, and the forest will be clear-cut in a short time. Fast-growing and high-yield forest tree species are usually planted for the production of pulp, paper, wood, and wood pellets for bioenergy production, and also play an important role in reducing the pressure on natural forests. With the changes in the ecological environment, the importance of fast-growing and high-yield forests in ecological protection and carbon sequestration is also increasingly prominent. In contrast to the continuous decline in global forest coverage, the area of plantations has more than doubled in the past few decades, of which fast-growing and high-yield forests account for about half. In addition, with the growing demand for wood, the construction of fast-growing and high-yield forests has become increasingly important. It is estimated that the area of fast-growing and high-yield forests will expand by 2% per year, and the total area will reach about 90 million hectares by the middle of this century. In addition, due to the high yield of fast-growing and high-yield forests, their carbon sequestration capacity may also be higher, making them a cost-effective measure to mitigate climate change. In addition, they can increase the land value of marginal agricultural land and bring higher resource utilization benefits.

[0003] Selecting a suitable afforestation area has a decisive influence on the growth yield and implementation effectiveness of fast-growing and high-yield forest tree species. Traditional methods for determining the afforestation area of fast-growing and high-yield forest tree species rely heavily on experience-based judgments, i.e., only on the suitability of the ecological environment. The similarity between the ecological environment characteristics of the tree species distribution area and the ecological environment characteristics of the afforestation target area is compared, and then the suitable afforestation area is determined. This traditional method only considers the suitability of the ecological environment, without considering the regional construction feasibility in terms of irrigation and artificial management costs, and market accessibility. As a result, the area of suitable afforestation for fast-growing and high-yield forests is overestimated. In some areas, fast-growing and high-yield forests achieve the expected yield, but fail to achieve the expected resource utilization benefits, or some fast-growing and high-yield forest tree species grow poorly after afforestation due to human activities, which greatly affects the implementation feasibility of fast-growing and high-yield forest tree species afforestation. Therefore, there is an urgent need for a scientific, reasonable, and precise method for determining the afforestation area of fast-growing and high-yield forest tree species to optimize yield and resource utilization benefits. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a method and system for determining the afforestation area of fast-growing and high-yield forest tree species, which combines reinforcement learning and rule modeling to achieve intelligent and precise delineation of fast-growing and high-yield forest afforestation areas under the dual constraints of ecology and regional implementation conditions, significantly improving the scientificity and implementability of suitability determination.

[0005] To achieve the above object, the present application provides the following scheme:

[0006] A method for establishing a reforestation area of fast-growing and high-yield forest tree species, comprising:

[0007] Obtaining reforestation distribution data of a target fast-growing and high-yield forest tree species;

[0008] Based on the reforestation distribution data, extracting ecological environment data of a first dimension and development basis data of a second dimension, respectively;

[0009] Using an improved maximum entropy reinforcement learning model to establish an ecological environment suitable area prediction model, taking the reforestation distribution data as a response variable and the ecological environment data as an explanatory variable, and obtaining an ecological environment suitable area of a target area according to the ecological environment suitable area prediction model;

[0010] Performing index interval statistical analysis on the development basis data, constructing a rule-based construction feasible area prediction model, and obtaining a construction feasible area of the target area according to the construction feasible area prediction model;

[0011] Correcting the range of the ecological environment suitable area so that the spatial intersection of the corrected ecological environment suitable area and the construction feasible area is consistent, to obtain a final suitable reforestation area of the target fast-growing and high-yield forest tree species.

[0012] Preferably, the reforestation distribution data is obtained through remote sensing or vegetation map interpretation, forest resource inventory, published literature, online database data, and forest resource inventory data.

[0013] Preferably, based on the reforestation distribution data, the ecological environment data of the first dimension and the development basis data of the second dimension are extracted, respectively, including:

[0014] Within the coverage range of the reforestation distribution data, ecological environment spatial raster data is downloaded by searching network shared data; the ecological environment spatial raster data includes climate data, soil data, and topographic data;

[0015] Through the extraction value to point tool of the geographic information system platform, the corresponding index values in the ecological environment spatial raster data are extracted to the reforestation distribution points of the reforestation distribution data, and an ecological environment data table containing each reforestation distribution point is generated according to the field format;

[0016] The extracted ecological environment data table is subjected to field merging and standard format arrangement to form ecological environment data of the first dimension for modeling analysis;

[0017] acquire spatial vector position data of roads, settlements, water sources, cities, towns and ports of OpenStreetMap and national basic geographic information database, and population density spatial raster data, gross domestic product spatial raster data and night light index spatial raster data;

[0018] calculate spatial distance and accessibility time from each afforestation distribution point to the nearest road, settlement, water source and city according to the spatial vector position data by using distance calculation and network analysis tools of the geographic information system platform, and generate market accessibility index according to market accessibility formula;

[0019] extract corresponding index values to each afforestation distribution point from the population density spatial raster data, the gross domestic product spatial raster data and the night light index spatial raster data by using the extract value to point tool of the geographic information system platform, and combine the corresponding index values with the spatial distance and accessibility time, the market accessibility index to generate a development basis data table;

[0020] carry out field arrangement and format unification on the development basis data table to form a second dimension development basis data.

[0021] Preferably, the calculation formula of the market accessibility formula is:

[0022] ;

[0023] wherein a ij represents market accessibility of afforestation distribution point i to destination j, d ij is the distance between point i and point j; for each position i, a ij value of the nearest first group and second group of destinations j is calculated respectively; S j represents importance of destination j, wherein the first group of destinations S j is assigned a value of 1, the second group of destinations S j is assigned a value of 0.5, and v is a preset constant, and the assignment process is: ; wherein d * represents the distance from point i to the position where the accessibility decreases the fastest.

[0024] Preferably, the afforestation distribution data is taken as a response variable, and the ecological environment data is taken as an explanatory variable, an ecological environment suitability area prediction model is established by using an improved maximum entropy reinforcement learning model, and an ecological environment suitability area of a target region is obtained according to the ecological environment suitability area prediction model, including:

[0025] a state vector is constructed by using the ecological environment data of each afforestation distribution point, and existence or nonexistence of afforestation distribution record is taken as a label ;

[0026] Will The dataset is randomly divided into a training set and a validation set.

[0027] Within the improved maximum entropy reinforcement learning framework, the state vector A hierarchical spatiotemporal state embedding network coupled with a graph convolution-transformer is used to obtain multi-scale embeddings. ;

[0028] Setting the motion space ,in This indicates that the grid is determined to be an ecologically suitable area. This indicates that the area has been determined to be unsuitable.

[0029] Constructing a reward function The reward function This includes ecological prediction rewards and resource allocation costs; the ecological prediction reward is the reward for the current action. Give positive rewards at that time. Otherwise, the reward is 0; the resource allocation cost item is: based on the development foundation data of the second dimension, it is determined whether the regional construction feasibility standard is met according to the rule model. If it is met, then the cost item... Otherwise, cost item The reward function The formula is: ;in, To develop constraint weight coefficients;

[0030] Construct a maximum entropy reinforcement learning model, the maximum entropy reinforcement learning model including: a policy network. and value network The policy network Output in state Take action below probability distribution; value network Used for estimating state-action pairs Expected returns;

[0031] The Soft Actor-Critic strategy is used for optimization to maximize the objective function during model training; the objective function is: ;in, For the first The state at each time step; The action selected by the strategy in this state; This is the entropy weighting coefficient, used to adjust the exploratory and stable nature of the strategy;

[0032] After each update, the status-action-reward triplet will be updated. The experience replay buffer is stored and the entropy weight coefficient is adaptively adjusted according to the validation set performance ;

[0033] When the policy network converges, the model parameters are fixed to obtain the ecological environment suitable area prediction model ;

[0034] Based on the ecological environment suitable area prediction model, all to-be-judged positions in the target area are predicted, and a probability distribution is calculated ; ;

[0035] All positions with probability distribution exceeding a set threshold are aggregated, and the output is the ecological environment suitable area of the target area ;

[0036] Preferably, the development basis data is subjected to index interval statistical analysis, a rule-based construction feasible area prediction model is constructed, and a construction feasible area of the target area is obtained according to the construction feasible area prediction model, including:

[0037] The minimum value, maximum value, mean value and standard deviation of each index in the development basis data are calculated on the afforestation distribution points of the afforestation distribution data to form a statistical analysis table

[0038] According to the statistical analysis table and in combination with a preset suitable threshold reference table, a lower limit and an upper limit are determined for each index to obtain an index interval set ; k is an index, ; ;

[0039] A construction feasible area prediction model is constructed; the determination rule of the construction feasible area prediction model is: if are all true, the regional construction feasibility determination result is ; otherwise ; is the value of the i-th index of the to-be-evaluated grid, indicates the construction feasible area, and indicates the unsuitable area ; The determination rule is called for all candidate grids of the target area one by one to generate a binary grid layer ; wherein the value 1 of the binary grid layer indicates a construction feasible area pixel, and the value 0 of the binary grid layer indicates an unsuitable area pixel

[0040] ; ;

[0041] ​​​The connected domain analysis is performed, the connected patches with an area not less than a first minimum threshold value is reserved, and a morphological opening operation is performed to smooth the boundary, so as to obtain the construction feasible area.

[0042] Preferably, the indicators include population density, night light index, gross domestic product, road distance, residential distance, water source distance, urban accessibility time and market accessibility index.

[0043] Preferably, the scope of the ecological environment suitable area is corrected so that the spatial intersection of the corrected ecological environment suitable area and the construction feasible area is consistent, so as to obtain the final suitable afforestation area of the target fast-growing and high-yield forest tree species, including:

[0044] In the geographic information system platform, the ecological environment suitable area and the construction feasible area are uniformly set as the same plane coordinate system;

[0045] The spatial overlay-intersection tool of the geographic information system platform is called to calculate the spatial intersection of the layers of the ecological environment suitable area and the construction feasible area, and an intersection vector layer is generated ;

[0046] The A fusion operation is performed to remove the internal boundary and form a single spatial feature set;

[0047] The "area filtering" tool is used to delete isolated patches with an area less than a second minimum threshold value set in the single spatial feature set;

[0048] The morphological opening operation is performed on the reserved patches to smooth the boundary, and the corrected ecological environment suitable area vector is output;

[0049] The corrected ecological environment suitable area is determined as the final suitable afforestation area.

[0050] A fast-growing and high-yield forest tree species afforestation area establishment system, comprising:

[0051] Afforestation distribution data acquisition unit for acquiring afforestation distribution data of target fast-growing and high-yield forest tree species;

[0052] Multi-dimensional environment-development data extraction unit for extracting first-dimensional ecological environment data and second-dimensional development basis data based on the afforestation distribution data;

[0053] Ecological environment suitable area prediction unit for taking the afforestation distribution data as a response variable and the ecological environment data as an explanatory variable, establishing an ecological environment suitable area prediction model by using an improved maximum entropy reinforcement learning model, and obtaining an ecological environment suitable area of a target area according to the ecological environment suitable area prediction model;

[0054] The construction feasible area prediction unit is configured to perform index interval statistical analysis on the development basis data, construct a rule-based construction feasible area prediction model, and obtain a construction feasible area of a target region according to the construction feasible area prediction model;

[0055] The range correction and result fusion unit is configured to perform range correction on the ecological environment suitability area, so that the spatial intersection of the corrected ecological environment suitability area and the construction feasible area is consistent, and finally obtain a final suitable afforestation area of the target fast-growing high-yield forest tree species.

[0056] The present application discloses the following technical effects:

[0057] The present application introduces an improved maximum entropy reinforcement learning model and a rule-based construction feasible area prediction model, dynamically coupling ecological environment suitability and development feasibility at the same spatial scale for the first time: on the one hand, the maximum entropy reinforcement learning uses multi-scale ecological environment characteristics to adaptively optimize the decision-making strategy, which can significantly improve the accuracy of ecological suitability area determination while maintaining exploration; on the other hand, the regional implementation rule model takes eight indexes such as population density, traffic accessibility and resource utilization return output as hard constraints to ensure that the selected area has realistic operability in market, infrastructure and cost investment; finally, through spatial intersection and spot removal post-processing, the final suitable afforestation area is output, which is continuous, implementable and takes into account the dual needs of ecology and implementation conditions. This method not only avoids the ecological risks caused by pure experience threshold, but also overcomes the waste of afforestation cost caused by ignoring implementation constraints, realizes the precise layout of fast-growing high-yield forest, efficient use of resources and long-term sustainable management, and significantly improves the afforestation survival rate and resource utilization return rate. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0059] Figure 1 The method flowchart provided for the embodiments of the present application;

[0060] Figure 2 The technical route schematic diagram provided for the embodiments of the present application;

[0061] Figure 3 The system structure schematic diagram provided for the embodiments of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.

[0063] The present application aims to provide a fast-growing and high-yield forest tree species afforestation area establishment method and system, which combines reinforcement learning and rule modeling, and realizes intelligent and accurate delineation of fast-growing and high-yield forest afforestation areas under the dual-dimensional constraints of ecology and regional implementation conditions, significantly improving the scientificity and implementability of suitability determination.

[0064] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0065] Figure 1 The method flowchart provided by the embodiments of the present application is shown in Figure 1 The present application provides a fast-growing and high-yield forest tree species afforestation area establishment method, characterized in that it comprises:

[0066] Step 100: Obtain afforestation distribution data of a target fast-growing and high-yield forest tree species;

[0067] Step 200: Based on the afforestation distribution data, extract ecological environment data of the first dimension and development basis data of the second dimension, respectively;

[0068] Step 300: Take the afforestation distribution data as the response variable, take the ecological environment data as the explanatory variable, use an improved maximum entropy reinforcement learning model to establish an ecological environment suitable area prediction model, and obtain the ecological environment suitable area of the target area according to the ecological environment suitable area prediction model;

[0069] Step 400: Perform index interval statistical analysis on the development basis data, construct a rule-based construction feasible area prediction model, and obtain the construction feasible area of the target area according to the construction feasible area prediction model;

[0070] Step 500: Correct the range of the ecological environment suitable area so that the spatial intersection of the corrected ecological environment suitable area and the construction feasible area is consistent, and obtain the final suitable afforestation area of the target fast-growing and high-yield forest tree species.

[0071] Preferably, the afforestation distribution data is obtained through remote sensing or vegetation map interpretation, forest resource inventory, published literature, online database data and forest resource inventory data.

[0072] As shown in Figure 2As shown, the afforestation distribution data of the fast-growing and high-yield forest tree species in the embodiment must come from commercial artificial afforestation sites, including: (1) remote sensing interpretation map; (2) vegetation map or forest resource map; (3) published literature (monograph, paper) data; (4) online database data; (5) forest resource inventory data.

[0073] Exemplarily, the online database data includes: Global Biodiversity Platform GBIF, China Digital Plant Herbarium CVH, etc.

[0074] Preferably, based on the afforestation distribution data, the first dimension ecological environment data and the second dimension development basis data are extracted respectively, including:

[0075] Within the coverage range of the afforestation distribution data, ecological environment spatial raster data is downloaded by searching network shared data; the ecological environment spatial raster data includes: climate data, soil data and topographic data;

[0076] The corresponding index values in the ecological environment spatial raster data are extracted to the afforestation distribution points of the afforestation distribution data by the extract value to point tool of the geographic information system platform, and an ecological environment data table containing each afforestation distribution point is generated in field format;

[0077] The extracted ecological environment data table is subjected to field merging and standard format arrangement to form the first dimension ecological environment data for modeling analysis;

[0078] The spatial vector position data of roads, residential areas, water sources, cities, towns and ports of OpenStreetMap and national basic geographic information database, and the population density spatial raster data, gross domestic product spatial raster data and night light index spatial raster data are obtained;

[0079] The spatial distance and access time of each afforestation distribution point to the nearest road, residential area, water source and city are calculated according to the spatial vector position data by using the distance calculation and network analysis tool of the geographic information system platform, and the market accessibility index is generated according to the market accessibility formula;

[0080] The corresponding index values are extracted from the population density spatial raster data, the gross domestic product spatial raster data and the night light index spatial raster data to each afforestation distribution point by using the extract value to point tool of the geographic information system platform, and the corresponding index values are combined with the spatial distance and access time, the market accessibility index to generate a development basis data table;

[0081] The development basis data table is subjected to field arrangement and format unification to form the second dimension development basis data.

[0082] Specifically, the ecological environment data acquisition step of the fast-growing and high-yield forest tree species distribution site of the embodiment is as follows:

[0083] The ecological environment data is acquired by searching network shared data, and the ecological environment spatial grid data includes climate, soil, and topographic data. Then, the "extract value to point" function module in the geographic information system or R language and the like is used to extract the ecological environment data values at the corresponding positions according to the geographic coordinate positions of the tree species distribution sites.

[0084] The climate data includes annual average temperature, annual temperature difference, annual average precipitation, summer precipitation, and the like. The data can be acquired in the following manner: (1) spatial interpolation of weather station data is performed by using Kriging interpolation and the like to acquire climate spatial grid data, (2) literature or network free shared spatial grid data. The literature data includes spatial grid climate data published in data papers, and the network data includes climate surface spatial grid data provided by the WorldClim and CHELSA websites.

[0085] The soil data includes soil organic carbon content, total nitrogen, total phosphorus, and total potassium element content, soil pH value, and the like. The data can be acquired from literature or network free contribution data. The network data includes spatial grid data from the global soil spatial grid data set SoilGRID250m, the world soil database HWSD, and the national Qinghai-Tibet Plateau scientific data center, and the like.

[0086] The global topographic data sources include elevation, slope, and slope direction, and the like. The spatial grid topographic data can be generated from topographic maps or digital elevation models (DEM) by using a geographic information system (GIS). The digital elevation model can be acquired from network shared data, such as the ASTER DEM, SRTM DEM, ALOS DEM, and the like.

[0087] Further, the development condition index spatial data of the embodiment includes but is not limited to market accessibility, distance from water source, population density, traffic time to town, distance from road, distance from residential area, nighttime light index, gross domestic product, and the like.

[0088] The road, water source (river, lake, marsh, and the like), residential area, city, town, and port spatial vector position data are from network shared data, including the world open street OpenStreetMap, 1:100 million public version basic geographic information data, 1:250,000 national basic geographic database, and the like.

[0089] The population density spatial grid data is obtained from free network sharing data, including population density data from the National Earth System Science Data Center, population density data GWPv4 from the National Aeronautics and Space Administration, population density data Landscan from the Oak Ridge National Laboratory, global population density data WorldPop, global human settlement grid data GHS POP, global population grid data GPW, and the like.

[0090] The gross domestic product spatial grid data is obtained from free network sharing data, including gross domestic product data from the National Earth System Science Data Center, and gross domestic product grid paper data in China.

[0091] The night light index spatial grid data is obtained from free network sharing data, including night light data from the National Earth System Science Data Center, VIIRS data from the National Oceanic and Atmospheric Administration, Suomi NPP satellite data, DMSP satellite data, and the like.

[0092] Specifically, the market accessibility calculation method is as follows:

[0093] Market accessibility refers to the distance, time and monetary cost that occurs when going to a series of market destinations. Due to the lack of market location data, two groups of locations are used as destinations, the first group uses the locations of cities and ports as market destinations instead, and the second group uses the locations of towns as market destinations instead. The market accessibility of each spatial location is determined by the maximum value of the market accessibility to the first group and the second group of destinations, respectively, according to the following calculation formula:

[0094]

[0095] wherein a ij represents the market accessibility of point i to destination j, d ij is the distance between point i and point j. For each location i, the a ij value to the nearest first group and second group of destinations j needs to be calculated respectively.

[0096] S j represents the importance of destination j, wherein the first group of destinations S j is assigned a value of 1, and the second group of destinations S j is assigned a value of 0.5.

[0097] v is a preset constant, and the assignment process is as follows:

[0098]

[0099] wherein d *This represents the distance from point i to the location where accessibility decreases most rapidly. Different constants were calculated for the two groups of destinations: the inflection point corresponding to the fastest decrease in accessibility for the first group of destinations was set to 2 hours, and for the second group of destinations, the inflection point was set to 45 minutes. The distance d corresponding to the inflection point was calculated based on the assumed speeds for calculating travel time provided in Table 1 below. * .

[0100] Table 1. Assumed speeds for calculating travel time

[0101]

[0102] The d-axis is achieved through the "Network Analysis" module of the Geographic Information System platform. ij and d * The calculations are performed in the "Raster Calculator" module of the Geographic Information System (GIS) to create spatial raster data for market accessibility based on the formulas mentioned above. Distances to water sources, roads, and settlements are calculated using the "Distance Analysis" module of the GIS platform, generating spatial raster surface data. Travel time to towns is calculated using the "Network Analysis" module of the GIS platform.

[0103] Preferably, the afforestation distribution data is used as the response variable, the ecological environment data is used as the explanatory variable, an improved maximum entropy reinforcement learning model is used to establish an ecological environment suitability zone prediction model, and the ecological environment suitability zone of the target area is obtained based on the ecological environment suitability zone prediction model, including:

[0104] Construct a state vector using the ecological environment data of each of the afforestation distribution points. The presence or absence of afforestation distribution records is used as a label. ;

[0105] Will The dataset is randomly divided into a training set and a validation set.

[0106] Within the improved maximum entropy reinforcement learning framework, the state vector A hierarchical spatiotemporal state embedding network coupled with a graph convolution-transformer is used to obtain multi-scale embeddings. ;

[0107] Setting the motion space ,in This indicates that the grid is determined to be an ecologically suitable area. This indicates that the area has been determined to be unsuitable.

[0108] Constructing a reward function The reward function This includes ecological prediction rewards and resource allocation costs; the ecological prediction reward is the reward for the current action. a positive reward is given , otherwise the reward is 0; the resource configuration cost term is: from the development basis data in the second dimension, whether the regional construction feasibility standard is met is determined according to the rule model, if met, the cost term , otherwise the cost term ; the formula of the reward function is: ; wherein, is a development constraint weight coefficient;

[0109] A maximum entropy reinforcement learning model is constructed, and the maximum entropy reinforcement learning model comprises: a policy network and a value network ; the policy network outputs a probability distribution of taking an action in a state ; the value network is used to estimate the expected return of a state-action pair ;

[0110] A Soft Actor-Critic policy optimization method is used to train the model to maximize the objective function; the objective function is: ; wherein, is the state of the i th time step; is the action selected by the policy in the state; is an entropy weight coefficient for adjusting the exploration and stability of the policy; After each update, a state-action-reward triple

[0111] is stored in an experience replay buffer, and the entropy weight coefficient is adaptively adjusted according to the performance of the validation set ;

[0112] When the policy network converges, the model parameters are fixed to obtain the ecological environment suitable area prediction model;

[0113] Based on the ecological environment suitable area prediction model, all to-be-judged positions in the target region are predicted, and a probability distribution is calculated;

[0114] All positions with a probability distribution exceeding a set threshold are aggregated, and the output is the ecological environment suitable area of the target region.

[0115] Specifically, the core approach of the present embodiment is to collect and splice temperature, precipitation, soil nutrients, terrain, and other ecological indicators for each afforestation sample point, and then input them into a "hierarchical spatiotemporal embedding network". The network first reads the spatial correlation between adjacent samples using graph convolution, then captures the climate-topography coupling relationship at different scales using multi-head attention, and finally outputs an ecological feature vector that can express both local continuity and macro gradient. Through this deep embedding, the subsequent model no longer relies on manual feature selection and avoids the defects of traditional pixel-by-pixel methods, which are fragile and lack context.

[0116] In the reinforcement learning phase, the model only needs to judge "whether this location is suitable for afforestation" between two actions. The reward design follows the dual goals of ecological accuracy and development feasibility: positive reinforcement is given when the prediction is consistent with historical afforestation records; if the location's market accessibility, transportation distance, and other development condition indicators fall outside the pre-set acceptable interval, a portion of the score will be deducted as a penalty. The policy network and value network use the Soft Actor-Critic structure, accelerating convergence through experience replay and early stopping mechanisms; the exploration intensity is dynamically adjusted during training to ensure that new areas are expanded while adhering to prior knowledge.

[0117] After the policy network is fixed, the "ecological suitability probability map" is obtained by reasoning each grid in the study area. For locations with a probability higher than the threshold, connected component merging, boundary smoothing, and removal of small patches with an area below the minimum final patch threshold are performed to obtain the ecological environment suitable area. Subsequently, the "construction feasibility area" grid is generated by calling the implementation rule model, and the two are combined using intersection operation to retain only the contiguous areas that meet both ecological and feasibility conditions, finally forming the fast-growing high-yield forest afforestation suitability area result map that can be directly used for planning.

[0118] Demonstratively, in the reinforcement learning training process of the ecological environment suitability area prediction model, to enable the model not only to have ecological judgment ability but also to actively avoid areas with high resource allocation cost, a "rule model" module is introduced in the construction of the reward function. The rule model does not directly participate in spatial zoning, but serves as a basis for punishment in the reward function. Its operation mode is as follows: in each training iteration, the system reads the development condition index values of the current training sample and compares them with the predefined intervals of various regional construction feasibility indicators; if any indicator does not meet the pre-set interval conditions, the sample is marked as "not meeting the implementation constraint conditions", and this result is input as a deduction signal into the reward function of the reinforcement learning algorithm. The rule model, as an auxiliary module embedded in the maximum entropy reinforcement learning framework, is an important mechanism to achieve the "ecological-implementation dual constraint" goal.

[0119] Optionally, in addition to using reinforcement learning, the embodiment can also use statistical analysis methods or other machine learning / artificial intelligence techniques to realize the correlation of tree species distribution data and the obtained ecological environment data of distribution points, that is, to establish a prediction model with tree species distribution data as the response variable (tree species existing is assigned as 1, and tree species not existing is assigned as 0) and the environmental data values corresponding to the distribution points as the explanatory variable, and then use the established model to predict the ecological environment suitable for afforestation in the target afforestation area.

[0120] Statistical methods include but are not limited to the following types: generalized linear models, generalized additive models, multivariate adaptive regression splines, flexible discriminant analysis, etc.

[0121] Machine learning / artificial intelligence algorithms include but are not limited to the following types: various algorithms based on decision trees and ensemble learning, various algorithms based on artificial neural networks, support vector machines, Bayesian, maximum entropy model MAXENT, etc.

[0122] In the R language or Python software platform, multiple statistical methods or machine learning / artificial intelligence algorithms are used to establish models, and then the prediction accuracy of the multiple established models is compared and analyzed to screen the optimal model, and then the optimal model is used to predict the ecological environment suitable for afforestation of tree species in the target afforestation area. Or use one method to establish a model and evaluate the model, if its prediction accuracy is high, directly use this model to predict the ecological environment suitable for afforestation of tree species in the target afforestation area.

[0123] Feature selection in model training process includes but is not limited to the following methods: variance inflation factor method, correlation coefficient method, information gain method based on information entropy theory, feature importance method, forward recursive addition, backward recursive deletion, genetic algorithm, etc.

[0124] Cross-validation or random data segmentation method is used to distinguish model construction data and model evaluation data.

[0125] Model evaluation indicators include but are not limited to the following indicators: root mean square error (RMSE), mean absolute prediction error (MAE), coefficient of determination (R 2 ) and mean cross entropy (MXE), etc.

[0126]

[0127]

[0128]

[0129]

[0130] where p i and o i are the predicted and observed values (1, tree species present; 0, tree species absent) at location i, observed mean; n dataset size, p model validation data tree species occurrence rate.

[0131] Preferably, the development basis data is subjected to index interval statistical analysis, a rule-based construction feasible area prediction model is constructed, and a construction feasible area of a target region is obtained according to the construction feasible area prediction model, comprising:

[0132] The minimum value, maximum value, mean value and standard deviation are calculated for each index in the development basis data on the afforestation distribution points of the afforestation distribution data to form a statistical analysis table;

[0133] According to the statistical analysis table and in combination with a preset suitable threshold reference table, the lower limit and the upper limit are determined for each index, to obtain an index interval set ; k is an index index, ;

[0134] A construction feasible area prediction model is constructed; the determination rule of the construction feasible area prediction model is: if are all true, then the regional construction feasibility determination result is ; otherwise ; is the value of the kth index of the to-be-evaluated grid, indicates a construction feasible area, indicates a non-suitable area; The determination rule is called for all candidate grids of the target region one by one to generate a binary grid layer

[0135] ; wherein the value 1 of the binary grid layer indicates a construction feasible area pixel, and the value 0 of the binary grid layer indicates a non-suitable area pixel;

[0136] ​​The connected domain analysis is performed, the connected patches with an area not less than a first minimum threshold are reserved, and a morphological opening operation is performed to smooth the boundary, to obtain the construction feasible area.

[0137] Preferably, the indicators include population density, night light index, gross domestic product, road distance, settlement distance, water source distance, urban accessibility time, and market accessibility index.

[0138] Specifically, the embodiment first calculates the minimum value, the maximum value, the mean value, and the standard deviation of the eight indicators in the second dimension development foundation data, i.e., population density, night light index, regional gross domestic product, road distance, settlement distance, water source distance, urban accessibility time, and market accessibility, on the afforestation distribution points, to form a statistical analysis table. On this basis, in combination with the preset suitable threshold reference standard, the upper and lower limit intervals of each indicator are determined, to form a complete indicator interval set. The interval not only reflects the regional support condition distribution range of the existing afforestation activities, but also provides an adjustable parameter basis, to provide clear boundary conditions for the subsequent suitability determination rule.

[0139] Optionally, the embodiment adopts a series type Boolean determination rule based on the indicator interval to construct the construction feasible area prediction model. Specifically, for each to-be-evaluated grid in the target region, it is checked whether each development condition indicator is within the corresponding interval range; if all indicators meet the preset upper and lower limits, it is determined as a construction feasible area, otherwise it is determined as an unsuitable area. The rule model has clear logic and transparent parameters, does not need model training, is convenient for quickly adjusting the determination standard under different regional conditions, and has good adaptability and interpretability.

[0140] Demonstratively, the embodiment adopts a rule-based determination method to independently judge each candidate position of the target region when constructing the construction feasible area. Specifically, the system compares the eight development condition indicators of each position with the interval upper and lower limits obtained from the afforestation sample statistics in sequence, and only when all indicators are within the corresponding interval range, the position is determined as a construction feasible area; if any indicator is out of range, it is determined as an unsuitable area. The rule takes the explicit “full satisfaction” logic as the determination standard, to form a Boolean logic model with strong interpretability and flexible adjustment. The model is directly used for grid determination and spatial zoning, and is the core method for constructing the construction feasible area layer.

[0141] Further, the embodiment expresses the result of the above determination in a binary raster form, where a value of one indicates that the region satisfies the implementation condition, and a value of zero indicates that it does not. To enhance the spatial operability of the region, further connected component analysis is performed to remove scattered patches with an area less than a first minimum threshold. Subsequently, the boundaries of the remaining patches are smoothed by a morphological opening operation to remove burrs and holes, ensuring that the structure of the spatial patches is regular and the boundaries are clear. The final output of the construction feasible region layer can be used as the basis for subsequent overlay analysis with the ecological suitability zone, providing an accurate implementation constraint boundary for the final suitable afforestation zone.

[0142] Further, the embodiment extracts tree species distribution point data (i.e., the above afforestation distribution data), then, statistically analyzes the interval range corresponding to these distribution points in each development condition index spatial raster data, and establishes a rule-based construction feasible region prediction model according to the interval range; finally, the prediction model is used to predict the afforestation area with implementation conditions in the target afforestation zone.

[0143]

[0144] In the formula, y is the output result of the rule-based regional construction feasibility model, 1 represents the afforestation area with implementation conditions, and 0 represents the unsuitable afforestation area. The indices a, b,..., and c are all the development condition indices used, is the interval range of index a. represents logical AND, i.e., all conditions must be met simultaneously.

[0145] Preferably, the range of the ecological environment suitability zone is corrected so that the spatial intersection of the corrected ecological environment suitability zone and the construction feasible region is consistent, obtaining the final suitable afforestation zone of the target fast-growing and high-yield forest tree species, including:

[0146] In the geographic information system platform, the ecological environment suitability zone and the construction feasible region are uniformly set as the same plane coordinate system;

[0147] The spatial overlay-intersection tool of the geographic information system platform is called to calculate the spatial intersection of the layers of the ecological environment suitability zone and the construction feasible region, generating an intersection vector layer ;

[0148] The fusion operation is performed to remove internal boundaries and form a single spatial feature set;

[0149] The "area filtering" tool is used to delete isolated patches with an area less than a second minimum threshold set in advance in the single spatial feature set;

[0150] Perform morphological opening operation on the reserved patches to smooth the boundaries, and output the corrected ecological environment suitability zone vector;

[0151] Determine the corrected ecological environment suitability zone as the final suitable afforestation area.

[0152] Specifically, first, in the geographic information system platform, the ecological environment suitability zone layer and the construction feasibility zone layer are set to the same projection coordinate system to eliminate scale distortion and offset; then, the "spatial overlay-intersection" function provided by the platform is run to perform per-pixel intersection on the two layers and generate new intersection vector data. The intersection result only retains spatial units that meet the requirements of ecological and implementation conditions, laying a foundation for subsequent patch optimization.

[0153] To avoid repeated boundaries in the intersection layer, the embodiment calls the "fusion" tool to integrate fragmented surface features into a single surface collection, and then uses the "area screening" function to delete isolated patches smaller than the second minimum threshold value. The threshold is customized for the final afforestation area, which is larger than the first threshold value in the aforementioned implementation condition screening stage, ensuring that the remaining patches have practical operability in terms of ecological connectivity, management and mechanical operation area.

[0154] For the continuous area remaining after area screening, the embodiment uses morphological opening operation to smooth the boundaries, remove jagged edges and narrow angles, and fill local holes, making the final patch shape regular and the edges smooth. The corrected layer obtained after the above processing is confirmed as the final suitable afforestation area for the target fast-growing high-yield forest tree species, which can be directly used for afforestation planning and subsequent tender drawing preparation, significantly improving the spatial accuracy and implementability of afforestation site selection decisions.

[0155] Corresponding to the above method, the embodiment also provides a fast-growing high-yield forest tree species afforestation area establishment system, comprising:

[0156] Afforestation distribution data acquisition unit, configured to acquire afforestation distribution data of a target fast-growing high-yield forest tree species;

[0157] Multi-dimensional environment-development data extraction unit, configured to extract first-dimensional ecological environment data and second-dimensional development basis data based on the afforestation distribution data;

[0158] Ecological environment suitability zone prediction unit, configured to use the afforestation distribution data as the response variable and the ecological environment data as the explanatory variable, establish an ecological environment suitability zone prediction model using an improved maximum entropy reinforcement learning model, and obtain the ecological environment suitability zone of the target area according to the ecological environment suitability zone prediction model;

[0159] The construction feasible area prediction unit is configured to perform index interval statistical analysis on the development basis data, construct a rule-based construction feasible area prediction model, and obtain a construction feasible area of a target region according to the construction feasible area prediction model.

[0160] The range correction and result fusion unit is configured to perform range correction on the ecological environment suitable area, so that the spatial intersection of the corrected ecological environment suitable area and the construction feasible area is consistent, and finally obtain a final suitable afforestation area of the target fast-growing and high-yield forest tree species.

[0161] The present application has the following advantages:

[0162] The present application first introduces two dimensions of ecological environment data and development basis data in the afforestation area demarcation of fast-growing and high-yield forest, and takes the spatial intersection of the ecological environment suitable area and the construction feasible area as the final afforestation suggestion area, effectively overcoming the one-sidedness problem of site selection caused by only using a single ecological factor such as climate and soil as a criterion in the prior art, and significantly improving the scientificity and practicability of the afforestation area demarcation.

[0163] The present application innovatively uses an improved maximum entropy reinforcement learning model for ecological suitability modeling, and by constructing a reward function composed of ecological prediction rewards and implementation constraint costs, the model can automatically avoid high implementation constraint cost areas while learning the historical afforestation distribution pattern; combined with an adaptive entropy adjustment strategy, the model can realize migration generalization and stable decision-making in different regions, and enhance the flexibility and accuracy of suitability prediction.

[0164] The present application constructs a rule determination model based on the index interval of development conditions as an adjustable external constraint logic, to ensure that the demarcation result meets the actual afforestation conditions such as actual accessibility, rationality of land use, and supportability of infrastructure. The model has good interpretability and flexibility, and can be quickly adjusted according to different regions and planning policies, significantly improving the policy fit and operation landing of the afforestation scheme.

[0165] In the post-processing stage of the suitable area determination, the present application introduces spatial optimization steps such as connected domain analysis, minimum patch area screening, and morphological opening operation, effectively eliminating isolated patches and patch interference, and improving the coherence, regularity and management operability of the patch. The final output afforestation suitable area layer has smooth boundaries and reasonable structure, and can be directly used for forestry management practical applications such as planting planning, investment evaluation and land allocation.

[0166] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0167] The principles and implementations of the present application are described in the specific examples, and the above examples are only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation and application range will be changed. In view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for establishing a forestation area of a fast-growing and high-yield tree species, characterized by, The application comprises the following steps: obtaining afforestation distribution data of target fast-growing high-yield forest tree species; extracting first-dimension ecological environment data and second-dimension development basis data based on the afforestation distribution data; using an improved maximum entropy reinforcement learning model to establish an ecological environment suitable area prediction model by taking the afforestation distribution data as a response variable and the ecological environment data as an explanatory variable, and obtaining the ecological environment suitable area of the target area according to the ecological environment suitable area prediction model; performing index interval statistical analysis on the development basis data to construct a rule-based construction feasible area prediction model, and obtaining the construction feasible area of the target area according to the construction feasible area prediction model; correcting the ecological environment suitable area to make the spatial intersection of the corrected ecological environment suitable area and the construction feasible area consistent, so as to obtain the final suitable afforestation area of the target fast-growing high-yield forest tree species; extracting first-dimension ecological environment data and second-dimension development basis data based on the afforestation distribution data, which comprises the following steps: downloading ecological environment spatial raster data by searching network shared data within the coverage range of the afforestation distribution data; the ecological environment spatial raster data comprises climate data, soil data and topography data; extracting corresponding index values in the ecological environment spatial raster data to afforestation distribution points of the afforestation distribution data by using the extract value to point tool of a geographic information system platform, and generating an ecological environment data table containing each afforestation distribution point according to a field format; performing field merging and standard format arrangement on the extracted ecological environment data table to form first-dimension ecological environment data for modeling analysis; obtaining spatial vector position data of roads, residential areas, water sources, cities, towns and ports of OpenStreetMap and national basic geographic information database, as well as population density spatial raster data, gross domestic product spatial raster data and night light index spatial raster data; calculating spatial distances and access times of each afforestation distribution point to the nearest road, residential area, water source and city according to the spatial vector position data by using the distance calculation and network analysis tool of the geographic information system platform, and generating a market accessibility index according to a market accessibility formula; extracting corresponding index values from the population density spatial raster data, the gross domestic product spatial raster data and the night light index spatial raster data to each afforestation distribution point by using the extract value to point tool of the geographic information system platform, and combining the corresponding index values with the spatial distances and access times and the market accessibility index to generate a development basis data table; performing field arrangement and format unification on the development basis data table to form second-dimension development basis data; using an improved maximum entropy reinforcement learning model to establish an ecological environment suitable area prediction model by taking the afforestation distribution data as a response variable and the ecological environment data as an explanatory variable, and obtaining the ecological environment suitable area of the target area according to the ecological environment suitable area prediction model, which comprises the following steps: constructing a state vector with the ecological environment data of each afforestation distribution point and taking the existence of afforestation distribution record as a label ; Will Randomly divided into training set and validation set; Within the improved maximum entropy reinforcement learning framework, a state vector Input graph convolution-transformer coupled hierarchical spatio-temporal state embedding network to obtain multi-scale embedding ; Setting the action space wherein denotes that the grid cell is determined to be an ecologically suitable area, denotes that the grid cell is determined to be a non-suitable area; Constructing a reward function , the reward function includes an ecological prediction reward and a resource allocation cost term; the ecological prediction reward gives a positive reward when the action , otherwise the reward is 0; the resource allocation cost term is: from the development base data in the second dimension, it is determined whether the regional construction feasibility standard is met according to the rule model, if met, the cost term , otherwise the cost term ; the formula of the reward function is: ; wherein, is a development constraint weight coefficient; A maximum entropy reinforcement learning model is constructed, which includes a policy network and a value network ; the policy network outputs a probability distribution of taking an action in a state ; the value network is used to estimate the expected return of a state-action pair ; The Soft Actor-Critic policy optimization method is adopted to train the model to maximize a target function; the target function is: ; wherein, is a state at a th time step; is an action selected by the policy at the state; is an entropy weight coefficient for adjusting the exploratory and stability of the policy; After each update, the state-action-reward triplets are stored in an experience replay buffer and the entropy weight coefficient is adaptively adjusted according to the validation set performance ; When the policy network After convergence, the model parameters are fixed to obtain the ecological environment suitability zone prediction model. Based on the aforementioned ecological suitability zone prediction model, all locations to be determined in the target area... Make predictions and calculate probability distributions. ; Aggregating all locations whose probability distribution exceeds a set threshold The output is the eco-environmental suitability zone of the target area.

2. The method for establishing a plantation area of fast-growing and high-yielding tree species according to claim 1, characterized by, The afforestation distribution data is obtained through remote sensing or vegetation interpretation, forest resource inventory, published literature, online database data and forest resource inventory data.

3. The method for establishing a plantation area of fast-growing and high-yielding tree species according to claim 1, characterized by, The calculation formula of the market accessibility formula is: ； wherein a ij represents the market accessibility of afforestation distribution point i to destination j, d ij is the distance between point i and point j; for each location i, a ij value is respectively determined for its to the nearest first group and second group of destinations j; S j represents the importance of destination j, wherein the first group of destinations S j is assigned a value of 1, and the second group of destinations S j is assigned a value of 0.5, and v is a preset constant, and the assignment process is as follows: ; wherein d * represents the distance from point i to the position at which the accessibility drops the fastest.

4. The method for establishing a plantation area of fast-growing and high-yielding tree species according to Claim 1, characterized by, The development basis data is subjected to index interval statistical analysis, a rule-based construction feasible area prediction model is constructed, and a construction feasible area of the target region is obtained according to the construction feasible area prediction model, comprising: The minimum value, maximum value, mean value and standard deviation of each index in the development basis data are calculated on the afforestation distribution points of the afforestation distribution data to form a statistical analysis table. According to the statistical analysis table and in combination with a preset suitable threshold reference table, a lower limit is determined for each index and the upper limit , to obtain an index interval set ; k is an index of the index, ; Construct a construction feasibility zone prediction model; the determination rule for the construction feasibility zone prediction model is: if If all conditions are met, then the feasibility assessment result for regional construction is... ;otherwise ; For the grid to be evaluated Item index value, Indicates a feasible area for construction. Indicates an unsuitable area; The decision rule is called for all candidate grids in the target area to generate a binary grid layer ; wherein the value 1 of the binary grid layer represents a construction feasible area pixel, and the value 0 of the binary grid layer represents a non-suitable area pixel right Perform connected component analysis, retain connected patches with an area not less than a pre-set first minimum threshold, and perform morphological opening operations to smooth the boundaries to obtain the feasible construction region.

5. The method for establishing a plantation area of fast-growing and high-yielding tree species according to claim 4, characterized by, The indexes include population density, night light index, gross domestic product, road distance, residential distance, water source distance, urban accessibility time and market accessibility index.

6. The method for establishing a plantation area of fast-growing and high-yielding tree species according to Claim 1, characterized by, The ecological environment suitable area is subjected to range correction so that the spatial intersection of the corrected ecological environment suitable area and the construction feasible area is consistent, and a final suitable afforestation area of the target fast-growing high-yield forest tree species is obtained, comprising: In a geographic information system platform, the ecological environment suitable area and the construction feasible area are uniformly set as the same plane coordinate system; calling a space superposition-intersection tool of the geographic information system platform, calculating a space intersection of layers of the ecological environment suitable area and the construction feasible area, and generating an intersection vector layer ; To perform a fusion operation to remove internal boundaries and form a single set of space elements; Using the "area screening" tool, isolated patches with an area less than a second minimum threshold value set in advance are deleted from the single spatial element set; A morphological opening operation is performed on the retained patches to smooth the boundaries, and a corrected ecological environment suitable area vector is output; The corrected ecological environment suitable area is determined as the final suitable afforestation area.

7. A fast-growing and high-yield forest tree species afforestation zone establishment system characterized by, Comprising: An afforestation distribution data acquisition unit configured to acquire afforestation distribution data of a target fast-growing high-yield forest tree species; A multi-dimensional environment-development data extraction unit configured to extract, based on the afforestation distribution data, first-dimensional ecological environment data and second-dimensional development basis data; An ecological environment suitable area prediction unit configured to use an improved maximum entropy reinforcement learning model to establish an ecological environment suitable area prediction model by taking the afforestation distribution data as a response variable and the ecological environment data as an explanatory variable, and to obtain an ecological environment suitable area of a target region according to the ecological environment suitable area prediction model; A construction feasible area prediction unit configured to perform index interval statistical analysis on the development basis data, construct a rule-based construction feasible area prediction model, and obtain a construction feasible area of a target region according to the construction feasible area prediction model; A range correction and result fusion unit configured to perform range correction on the ecological environment suitable area so that the spatial intersection of the corrected ecological environment suitable area and the construction feasible area is consistent, and to obtain a final suitable afforestation area of a target fast-growing high-yield forest tree species; Based on the afforestation distribution data, first-dimensional ecological environment data and second-dimensional development basis data are extracted, comprising: Within the coverage range of the afforestation distribution data, ecological environment spatial raster data is downloaded by searching network shared data; the ecological environment spatial raster data includes climate data, soil data and topography data; extracting corresponding index values in the ecological environment spatial grid data to afforestation distribution points of the afforestation distribution data by a value-to-point extraction tool of a geographic information system platform, and generating an ecological environment data table containing each afforestation distribution point according to a field format; performing field merging and standard format arrangement on the extracted ecological environment data table to form ecological environment data of a first dimension for modeling analysis; obtaining spatial vector position data of roads, residential areas, water sources, cities, towns and ports of OpenStreetMap and a national basic geographic information database, and population density spatial grid data, gross domestic product spatial grid data and night light index spatial grid data; calculating spatial distances and access times of each afforestation distribution point to the nearest road, residential area, water source and city according to the spatial vector position data by a distance calculation and network analysis tool of the geographic information system platform, and generating a market accessibility index according to a market accessibility formula; extracting corresponding index values from the population density spatial grid data, the gross domestic product spatial grid data and the night light index spatial grid data to each afforestation distribution point by a value-to-point extraction tool of the geographic information system platform, and merging the corresponding index values with the spatial distances and access times and the market accessibility index to generate a development basis data table; performing field arrangement and format unification on the development basis data table to form development basis data of a second dimension; using the afforestation distribution data as a response variable and the ecological environment data as an explanatory variable, establishing an ecological environment suitability zone prediction model by using an improved maximum entropy reinforcement learning model, and obtaining an ecological environment suitability zone of a target region according to the ecological environment suitability zone prediction model, including: constructing a state vector with the ecological environment data of each afforestation distribution point and taking the existence of afforestation distribution record as a label ; Will Randomly divided into training set and validation set; Within the improved maximum entropy reinforcement learning framework, a state vector Input graph convolution-transformer coupled hierarchical spatio-temporal state embedding network to obtain multi-scale embedding ; Setting an action space wherein denotes that the grid cell is determined to be an ecologically suitable area, denotes that it is determined to be a non-suitable area; Constructing a reward function , the reward function includes an ecological prediction reward and a resource allocation cost term; the ecological prediction reward gives a positive reward when the action , otherwise the reward is 0; the resource allocation cost term is: from the development base data in the second dimension, it is determined whether the regional construction feasibility standard is met according to the rule model, if met, the cost term , otherwise the cost term ; the formula of the reward function is: ; wherein, is a development constraint weight coefficient; A maximum entropy reinforcement learning model is constructed, which includes a policy network and a value network ; the policy network outputs a probability distribution of taking an action in a state ; the value network is used to estimate the expected return of a state-action pair ; The Soft Actor-Critic policy optimization method is used to train the model to maximize a target function, wherein the target function is: ; wherein, is a state at a th time step; is an action selected by the policy at the state; is an entropy weight coefficient for adjusting the exploration and stability of the policy. After each update, the state-action-reward triplets are stored in an experience replay buffer and the entropy weight coefficient is adaptively adjusted according to the validation set performance ; When the policy network After convergence, the model parameters are fixed to obtain the ecological environment suitability zone prediction model. Based on the aforementioned ecological suitability zone prediction model, all locations to be determined in the target area... Make predictions and calculate probability distributions. ; Aggregating all locations whose probability distribution exceeds a set threshold The output is the eco-environmental suitability zone of the target area.

Images