A smart site selection method and system for cultivated land occupation and compensation balance

Abstract

The application discloses a smart site selection method and system for cultivated land occupation and compensation balance, belongs to the technical field of smart site selection, and comprises the following steps: S1, collecting and preprocessing multi-source spatio-temporal data to construct a basic database; S2, constructing a spatial adjacency relation graph, and inputting data in the basic database into a graph neural network model to generate a dynamic evaluation weight vector for a to-be-selected land plot; S3, calculating final utility values of each evaluation index, and combining the dynamic evaluation weight vector to adopt a multiplicative aggregation model to calculate a multi-scale ecological balance score of the land plot; and S4, adopting a long short-term memory network model to predict a future quality evolution trend of the land plot. The application can accurately capture and respond to the spatio-temporal dynamic characteristics of a cultivated land ecological system, and the evaluation result is more suitable for the physical reality.

Description

Technical Field

[0001] This invention relates to the field of intelligent site selection technology, specifically to an intelligent site selection method and system for balancing arable land occupation and compensation. Background Technology

[0002] The balance between cultivated land occupation and replenishment is a key national policy for ensuring national food security and ecological balance. Traditional site selection methods often rely on expert experience or static multi-factor weighted assessments based on the hierarchical analysis method. These methods have significant technical bottlenecks. First, cultivated land quality exhibits high spatial heterogeneity and dynamic evolution over time, making it difficult for traditional static weighted models to capture these spatial and temporal dynamics. Second, cultivated land assessment needs to consider both micro-level soil attributes and macro-level regional ecological functions, and existing methods often lead to biased assessment results due to their single scale. Third, the assessment process integrates heterogeneous data from multiple sources, such as remote sensing imagery, ground monitoring, and socio-economic data. The inherent errors of each data source accumulate and amplify in the decision-making chain, resulting in significant uncertainty in the final results. Furthermore, the health status of cultivated land ecosystems often exhibits non-linear characteristics and threshold effects. For example, when soil nutrients fall below a certain critical value, land productivity experiences a precipitous decline, and simple linear weighted models cannot accurately characterize such abrupt changes. Therefore, there is an urgent need for a smart site selection technology solution for the balance between cultivated land occupation and replenishment that can overcome the above challenges and achieve dynamic, multi-scale, highly reliable, and forward-looking decision-making.

[0003] The present invention aims to overcome the shortcomings of the prior art and provide a smart site selection method and system for balancing arable land occupation and compensation, which can comprehensively handle spatiotemporal heterogeneity, multi-scale coupling, data uncertainty and ecological threshold effect, so as to improve the accuracy of site selection, ecological protection capacity and long-term sustainability. Summary of the Invention

[0004] The purpose of this invention is to provide a smart site selection method and system for balancing arable land occupation and compensation, so as to solve the problems mentioned in the background art.

[0005] The technical solution of the present invention includes the following steps:

[0006] S1. Collect and preprocess multi-source spatiotemporal data to build a basic database;

[0007] S2. Construct a spatial adjacency graph and input the data from the basic database into the graph neural network model to generate dynamic evaluation weight vectors for the candidate plots;

[0008] S3. Calculate the final utility value of each evaluation indicator, and use a multiplicative aggregation model in conjunction with the dynamic evaluation weight vector to calculate the multi-scale ecological balance score of the plot.

[0009] S4. Use a long short-term memory network model to predict the future quality evolution trend of the land parcel;

[0010] S5. Quantify data uncertainty, model uncertainty and spatial uncertainty separately, and perform a comprehensive calculation on the above three uncertainties to generate an overall uncertainty index;

[0011] S6. Calculate the comprehensive resilience index, and linearly weight the multi-scale ecological balance score, future quality evolution trend and comprehensive resilience index according to strategic parameters to generate a forward-looking decision score;

[0012] S7. Determine the reliability of the site selection recommendations based on the overall uncertainty index, and rank the candidate sites according to the forward-looking decision score to output the final list of site selection recommendations.

[0013] Preferably, multi-source spatiotemporal data includes: geospatial information of historical land reclamation projects, soil physicochemical properties at the plot scale, landscape pattern indices based on remote sensing image interpretation, and regional-scale ecosystem service function data and climate change data.

[0014] Preferably, S3 specifically includes:

[0015] S31. For key indicators with threshold effects, obtain the normalized measured values ​​and the corresponding ecological critical thresholds.

[0016] S32. When the measured value is lower than the ecological critical threshold, the final utility value with exponential decay is calculated using the first preset function;

[0017] S33. When the measured value is not lower than the ecological critical threshold, the final utility value with non-linear growth is calculated using the second preset function.

[0018] Preferably, S5 specifically includes:

[0019] S51. Quantify data uncertainty based on the reliability of the data source;

[0020] S52. Using the Monte Carlo dropout technique, the uncertainty of the model is estimated by repeatedly deactivating a portion of neurons in the graphical neural network model and observing the output variance.

[0021] S53. Calculate the normalized standard deviation of the key indicators of the target plot in the spatial neighborhood of all plots, as the spatial uncertainty.

[0022] S54. Perform square root calculations on the data uncertainty, model uncertainty, and spatial uncertainty to generate an overall uncertainty index.

[0023] Preferably, the comprehensive resilience index in S6 is obtained by weighted summation of the normalized values ​​of multiple resilience-related indicators, with the weights of each indicator determined by domain experts.

[0024] Preferably, the reliability of the location recommendation in S7 includes:

[0025] S71. Compare the overall uncertainty index with the preset uncertainty threshold;

[0026] S72. When the overall uncertainty index is less than the uncertainty threshold, the site selection recommendation is determined to be reliable.

[0027] S73. When the overall uncertainty index is not less than the uncertainty threshold, the site selection recommendation will be marked as high uncertainty and manual review is recommended.

[0028] A smart site selection system for balancing farmland occupation and compensation includes:

[0029] The data acquisition and preprocessing module is used to collect and preprocess multi-source spatiotemporal data to build a basic database;

[0030] The spatiotemporal weight adaptive learning module is used to construct a spatial adjacency graph and input data from the basic database into the graph neural network model to generate dynamic evaluation weight vectors for candidate plots.

[0031] The multi-scale ecological balance assessment module is used to calculate the final utility value of each assessment indicator and, in combination with the dynamic assessment weight vector, uses a multiplicative aggregation model to calculate the multi-scale ecological balance score of the plot.

[0032] The arable land quality evolution prediction module is used to predict the future quality evolution trend of land parcels using a long short-term memory network model.

[0033] The uncertainty quantification module is used to quantify data uncertainty, model uncertainty and spatial uncertainty respectively, and to perform comprehensive calculations on the above three uncertainties to generate an overall uncertainty index;

[0034] The decision support module is used to calculate the comprehensive resilience index, generate a forward-looking decision score, judge reliability based on the overall uncertainty index, and output a final list of site selection recommendations based on the forward-looking decision score.

[0035] This invention provides an improved intelligent site selection method and system for balancing arable land occupation and compensation, which has the following improvements and advantages compared with the prior art:

[0036] 1. The core breakthrough of this solution lies in the introduction of a graph neural network model to generate dynamic evaluation weight vectors. This model abstracts each plot of land as a node in a graph and the interactions between plots as edges. By learning from a large amount of historical data, it can calculate a set of the most suitable weights in real time based on the state of each plot of land at a specific point in time and its neighborhood influence. This frees the evaluation process from the constraints of subjective settings and static frameworks, and can accurately capture and respond to the spatiotemporal dynamic characteristics of the farmland ecosystem, so that the evaluation results are more consistent with physical reality.

[0037] 2. This invention synthesizes errors from three orthogonal dimensions: data uncertainty, data source quality, model uncertainty, algorithm stability, and spatial uncertainty, as well as local spatial heterogeneity. This allows each site selection suggestion to come with a clear confidence interval. When the overall uncertainty index exceeds a preset threshold, the system will suggest manual review, thus establishing a human-machine collaborative risk management mechanism. This is a fundamental advancement from simply providing an answer to providing an answer with confidence.

[0038] 3. This invention balances and optimizes multiple strategic objectives, such as ensuring current output, pursuing future production potential, and ensuring long-term ecological stability, making site selection decisions not only scientific but also precisely aligned with macroeconomic policy objectives. Attached Figure Description

[0039] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0040] Figure 1 This is a flowchart of a smart site selection method for balancing arable land occupation and compensation according to the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0042] Example 1:

[0043] Please see Figure 1 This invention provides a smart site selection method for balancing arable land occupation and compensation, comprising the following steps:

[0044] S1. Collect and preprocess multi-source spatiotemporal data to build a basic database;

[0045] S2. Construct a spatial adjacency graph and input the data from the basic database into the graph neural network model to generate dynamic evaluation weight vectors for the candidate plots;

[0046] S3. Calculate the final utility value of each evaluation indicator, and use a multiplicative aggregation model in conjunction with the dynamic evaluation weight vector to calculate the multi-scale ecological balance score of the plot.

[0047] S4. Use a long short-term memory network model to predict the future quality evolution trend of the land parcel;

[0048] S5. Quantify data uncertainty, model uncertainty and spatial uncertainty separately, and perform a comprehensive calculation on the above three uncertainties to generate an overall uncertainty index;

[0049] S6. Calculate the comprehensive resilience index, and linearly weight the multi-scale ecological balance score, future quality evolution trend and comprehensive resilience index according to strategic parameters to generate a forward-looking decision score;

[0050] S7. Determine the reliability of the site selection recommendations based on the overall uncertainty index, and rank the candidate sites according to the forward-looking decision score to output the final list of site selection recommendations.

[0051] To implement the above method, a preferred embodiment is provided; the method aims to overcome the limitations of traditional static assessment and achieve dynamic, multi-scale, highly reliable and forward-looking decision-making for the site selection of farmland occupation and compensation balance projects;

[0052] In step S1, multi-source spatiotemporal data are collected and preprocessed to construct a basic database. The purpose of this step is to provide a standardized, high-quality data foundation for subsequent analysis and modeling. In this embodiment, data collection and preprocessing are performed by a dedicated data acquisition and preprocessing module. This module is responsible for integrating data from multiple sources and formats, and performing operations such as data cleaning, spatial coordinate registration, time series alignment, and numerical normalization to ensure that all data are analyzed under a unified spatiotemporal reference.

[0053] In step S2, a spatial adjacency graph is constructed, and data from the basic database is input into the graph neural network model to generate dynamic evaluation weight vectors for the candidate land parcels. This step aims to solve the core problem of traditional methods where evaluation weights are fixed and cannot reflect spatiotemporal heterogeneity. Spatial adjacency graph It is constructed and defined as a graph structure:

[0054]

[0055] The set of nodes Representing each candidate plot of farmland, the edges are assembled. This represents the spatial or ecological relationships between plots, which can be determined based on whether the Euclidean distance between plots is less than a preset distance threshold or whether there is geographical adjacency. The preset distance threshold is determined objectively and reproducibly by performing spatial autocorrelation analysis on plots within the study area, such as calculating the Morans I index. By calculating the Morans I at different distance scales, the distance value with the most significant spatial positive correlation that begins to decay is found. This distance value is considered the characteristic distance where there is significant interaction between plots and is thus set as the preset distance threshold. For example, if the analysis shows that the soil nutrient content of plots within a 500-meter range exhibits the strongest spatial aggregation, then the threshold can be set to 500 meters. A pre-trained graph neural network model is used to process this graph.

[0056] To enable implementation by those skilled in the art, the graph neural network model can specifically adopt an architecture containing a two-layer graph attention network (GNN). The node features input to the GNN are time-series flattened one-dimensional vectors of soil physicochemical properties for each plot over the past five years, such as organic matter, total nitrogen content, and vegetation cover index. The first layer of the GNN uses eight attention heads and outputs a 64-dimensional feature vector. The second layer uses one attention head and outputs the final dynamic evaluation weight vector, with the dimension equal to the number of evaluation indicators. This design ensures that the model can generate an accurate and adaptive weight value for each indicator to be evaluated. During training, mean squared error is used as the loss function, Adam is selected as the optimizer, the initial learning rate is set to 0.005, and 200 iterations of training are performed to ensure model convergence. The choice of loss function aims to minimize the difference between the model's predicted weights and the ideal weights calculated based on historical successful project cases.

[0057] The network parameters of the GNN model were determined through supervised learning training on a dataset containing a large number of historical land reclamation projects and their subsequent evolution results. The feature data of the corresponding plots in the basic database, such as the time-series variation vector composed of multiple periods of soil physicochemical property monitoring data, were input into the model as node features. The model learns the plot's own attributes and its interaction with neighboring plots, and outputs a dynamic evaluation weight vector. This vector is a multi-dimensional vector, with each dimension corresponding to the weight of an evaluation index, and can adaptively adjust with changes in time and spatial context.

[0058] In step S3, the final utility value of each evaluation indicator is calculated, and a multiplicative aggregation model is used in conjunction with the dynamic evaluation weight vector to calculate the multi-scale ecological balance score of the plot. The purpose of this step is to achieve a comprehensive evaluation across scales and to reflect the bottleneck effect where the overall value is limited if there are serious deficiencies in key dimensions. The underlying logic is that for each evaluation indicator, its final utility value is calculated. This value is a dimensionless numerical value that reflects the contribution of the indicator after standardization. For indicators with threshold effects, a nonlinear transformation is performed. Based on the above utility value, a multiplicative aggregation model is used to aggregate the dynamic evaluation weight vector from step S2 with the final utility value of each indicator.

[0059] The mathematical expression of this multiplicative aggregation model is a weighted geometric mean model; for those possessing The assessment indicators for each plot of land, and the multi-scale ecological balance score. The calculation formula is:

[0060]

[0061] in, It is the first The final utility value of each evaluation indicator comes from steps S31-S33, where n represents the total number of evaluation indicators owned by the land parcel; and i represents the index of the i-th evaluation indicator. The measured value represents the i-th evaluation index; It is generated by the graph neural network model in step S2, and is related to the first... Each indicator corresponds to a dynamic evaluation weight; this formula ensures the utility value of any indicator. When the total score approaches 0, the total score is... It will also approach 0 sharply, thus accurately reflecting the weakest link effect;

[0062] The model is designed so that severe deficiencies in any dimension will significantly lower the overall score, thus calculating a multi-scale ecological balance score; this score comprehensively reflects the overall ecological health status of the plot at three scales: micro-plot, meso-landscape, and macro-region.

[0063] In step S4, a long short-term memory network model is used to predict the future quality evolution trend of the land parcel; in order to make the decision forward-looking, this step aims to assess the long-term sustainability of the land parcel; the long short-term memory network model is constructed, which is particularly good at processing and predicting time series data;

[0064] The LSTM model consists of two stacked LSTM layers, each containing 100 neurons, with a fully connected layer following the last layer to output the prediction result. To prevent overfitting, a dropout layer with a dropout rate of 0.2 is added after each LSTM layer. The input data format of the model is a three-dimensional tensor with the following dimensions: number of plot samples, time step, and number of features. The time step is set to the past 5 years, i.e., 5 time points. The number of features is 5 key indicators collected at each time point, namely soil organic matter content, vegetation cover index, soil total nitrogen content, soil available phosphorus content, and soil available potassium content. The goal of model training is to predict a single quantitative quality indicator in the next 5 years. The mean squared error is used as the loss function and the Adam optimizer is used to optimize the model parameters.

[0065] The model takes multi-source monitoring time-series data from the past few years, such as soil organic matter content and vegetation cover index, as input to predict the quality evolution trend of a plot in the future over a predetermined period, such as 5-10 years. The output is a quantitative indicator representing the rate of quality change, which is normalized and used for the final decision score.

[0066] In step S5, data uncertainty, model uncertainty, and spatial uncertainty are quantified separately, and the three uncertainties are comprehensively calculated to generate an overall uncertainty index. To ensure the reliability of the decision, this step quantifies the main sources of uncertainty in the site selection assessment process. Data uncertainty is assessed based on the reliability, accuracy, and timeliness of the data source. Model uncertainty is assessed using technical means, such as the Monte Carlo dropout method, to evaluate the stability of the predictions made by models such as GNN and LSTM. Spatial uncertainty is quantified by calculating the degree of variation of key indicators of the target site within its spatial neighborhood. These three uncertainties are then combined to generate an overall uncertainty index, which is used to measure the credibility of the final site selection recommendation.

[0067] In step S6, the comprehensive resilience index is calculated, and the multi-scale ecological balance score, future quality evolution trend, and comprehensive resilience index are linearly weighted according to strategic parameters to generate a forward-looking decision score. This step integrates all assessment dimensions to form the final ranking basis. Specifically, the comprehensive resilience index is calculated, quantifying the land's ability to resist and recover from external disturbances, such as extreme climate events. It is typically derived by weighting multiple indicators such as biodiversity and soil aggregate stability. This index is then linearly weighted with the current quality calculated in step S3 (i.e., the multi-scale ecological balance score) and the future trend predicted in step S4. The strategic parameters used for weighting are set by decision-makers according to their strategic preferences, such as prioritizing current yields or emphasizing long-term sustainability. To assist decision-makers in making scientific settings, the system can provide multiple preset standardized decision scenario models, each corresponding to a fixed set of strategic parameters to ensure... ,in, It is the weight of the multi-scale ecological balance score that reflects the current comprehensive quality of the land parcel; It is the weighting of the future quality evolution trend that predicts the future development potential of the land parcel; It is the weight of the comprehensive resilience index used to assess the long-term stability of the land parcel;

[0068] Production Priority Mode: Suitable for urgent food production tasks, focusing on the quality of the current plot. Parameter settings can be: ;

[0069] Sustainable development model: suitable for ecologically sensitive areas or long-term planning, focusing on future potential and stability. Parameter settings can be: ;

[0070] Balanced development model: taking into account the present, the future, and risks. Parameter settings can be: ;

[0071] Decision-makers can directly choose a model or fine-tune it based on these models, thereby quantifying and realizing their preferences; and calculating a forward-looking decision score.

[0072] In step S7, the reliability of the site selection recommendations is judged based on the overall uncertainty index, and the candidate sites are sorted according to the forward-looking decision score to output the final site selection recommendation list; this is the decision output step; the system compares the overall uncertainty index generated in step S5 with a preset threshold; if it is lower than the threshold, the evaluation result of the site is considered reliable; otherwise, it is marked as high uncertainty and manual review is recommended; all candidate sites judged as reliable are sorted in descending order according to the forward-looking decision score generated in step S6 to form a clear and well-founded final site selection recommendation list for decision-makers to use;

[0073] This invention constructs a complete, closed-loop intelligent site selection method through the above steps. Based on multi-source spatiotemporal data, it utilizes graph neural networks to capture spatiotemporal dynamics, achieves comprehensive evaluation through a multi-scale multiplicative aggregation model, leverages long short-term memory networks to provide forward-looking decision-making, and introduces uncertainty quantification to ensure the reliability of the results. Compared with traditional methods that rely on static weights and expert experience, this invention can significantly improve the scientific nature, accuracy, and long-term sustainability of farmland occupation-compensation balance site selection, providing advanced technical support for ensuring national food security and ecological balance.

[0074] Multi-source spatiotemporal data include: geospatial information of historical land occupation compensation projects, soil physicochemical properties at the plot scale, landscape pattern index based on remote sensing image interpretation, and regional-scale ecosystem service function data and climate change data.

[0075] In this embodiment, the multi-source spatiotemporal data in step S1 is specifically defined; the selection of data directly determines the depth and breadth of the model analysis; in this embodiment, the construction of the basic database covers four key categories of data:

[0076] Geospatial information of historical land occupation and compensation projects: refers to the location, scope, implementation time, project type, and post-project benefit evaluation data of completed farmland occupation and compensation projects in the past; its function is to provide samples with real result labels for training machine learning models such as GNN and LSTM, so that the models can learn the relationship between site selection quality and the long-term evolution of the land plot.

[0077] Soil physicochemical properties at the plot scale refer to high-precision data collected for each candidate plot, such as soil organic matter, total nitrogen, available phosphorus, available potassium content, pH value, soil texture, and bulk density. Their role is to provide the most direct basis for assessing the current basic soil fertility and production potential of the plot, and they are the core data input at the micro scale in multi-scale assessment.

[0078] Landscape pattern indices based on remote sensing image interpretation refer to quantitative indicators that describe the landscape structure of a region where a plot is located, extracted by analyzing and processing multiple high-resolution satellite remote sensing images. These indicators include patch density, edge index, aggregation index, and Shannon diversity index. Their function is to assess the ecological connectivity and habitat quality of a plot at the meso-landscape scale, reflecting its role in a broader ecological network.

[0079] Regional-scale ecosystem service function data and climate change data refer to macro-level data covering the entire study area; ecosystem service function data include spatial distribution assessment results of functions such as water conservation, soil and water conservation, and biodiversity maintenance; climate change data include historical and future estimated precipitation, temperature and other meteorological factors; their role is to evaluate the macro-ecological background of the plot and the long-term climate risks it faces at the macro-regional scale, and to provide support for the comprehensiveness and forward-looking nature of the assessment.

[0080] By clearly defining the specific composition of the aforementioned multi-source spatiotemporal data, this invention ensures the comprehensiveness and multi-scale characteristics of the input information. This data combination enables the model to comprehensively consider the suitability of a site from multiple levels, including soil micro-properties, surrounding landscape structure, regional macro-ecological functions, and climate background. The resulting technical benefits are: significantly improving the dimensionality and quality of the input information of the evaluation model, making the basis for site selection decisions more sufficient and comprehensive, and the final evaluation results more accurate and reliable.

[0081] This solution achieves spatiotemporal adaptive dynamics of evaluation weights. Traditional techniques generally employ methods such as the hierarchical analysis method to assign static weights to each evaluation indicator once determined, which remain unchanged. This approach fails to reflect the high spatial heterogeneity and dynamic evolution of arable land quality over time. The core breakthrough of this solution lies in introducing a graph neural network-based model to generate dynamic evaluation weight vectors. This model abstracts each plot as a node in a graph and the interactions between plots as edges. By learning from a large amount of historical data, it can calculate a set of optimal weights in real time based on the state of each plot at a specific point in time and its neighborhood influence. This frees the evaluation process from the constraints of subjective settings and static frameworks, enabling it to accurately capture and respond to the spatiotemporal dynamic characteristics of the arable land ecosystem, thus making the evaluation results more consistent with physical reality.

[0082] S3 specifically includes:

[0083] S31. For key indicators with threshold effects, obtain the normalized measured values ​​and the corresponding ecological critical thresholds.

[0084] S32. When the measured value is lower than the ecological critical threshold, the final utility value with exponential decay is calculated using the first preset function;

[0085] S33. When the measured value is not lower than the ecological critical threshold, the final utility value with non-linear growth is calculated using the second preset function;

[0086] In this embodiment, the process of calculating the final utility value in step S3 is specified, especially for key indicators with ecological threshold effects. The threshold effect refers to the fact that the impact of changes in certain ecological indicators, such as soil nutrient content, on ecosystem function is not linear. Once the value is below or above a certain critical point, the state of the system may change drastically. In order to accurately simulate this nonlinear response, this embodiment introduces an ecological threshold-aware nonlinear utility function.

[0087] In step S31, for key indicators with threshold effects, normalized measured values ​​and corresponding ecological critical thresholds are obtained; key indicators include soil organic matter content or concentration of a certain type of pollutant; normalized measured values ​​refer to dimensionless values ​​obtained by scaling the original observation data to the [0,1] interval through mathematical transformations, such as minimum-maximum normalization; ecological critical thresholds are key reference points, which are determined based on relevant ecological research, national or local agricultural production technical regulations, or expert knowledge, and are also normalized in the same way as the measured values.

[0088] In steps S32 and S33, the final utility value is calculated using a piecewise function. To clarify this calculation logic, when the measured value Below the ecological critical threshold When the first preset function is used:

[0089]

[0090] Calculate utility value; when measured value Not lower than the ecological critical threshold In this case, the second preset function will be used:

[0091]

[0092] Calculate the utility value; where exp: exponential function; i: index of the i-th key indicator; It is the first Normalized measurements of key indicators; Is with The corresponding normalized ecological critical threshold; and These are preset dimensionless coefficients that control the rates of utility decay and growth, respectively. The values ​​are determined based on the severity level of ecological protection policies applicable to specific scenarios. For example, a larger value can be set for basic farmland requiring strict protection. The value is set to enable strong penalties for any behavior below the threshold;

[0093] To objectively map policy levels to coefficients, a calibration table can be predefined. For example, three protection levels can be defined: Level I, within the ecological red line area (most stringent); Level II, permanent basic farmland; and Level III, general arable land. Specific coefficient ranges can be set for each level through expert consultation and historical data fitting. An example is shown below.

[0094] Level I Area: , use high The value is used to severely penalize any situation that falls below the threshold.

[0095] Level II area: ;

[0096] Level III Area: This allows for relatively mild changes in utility.

[0097] In practical applications, the median value within this range is taken based on the protection level of the land parcel, or it is finely adjusted according to more detailed policy documents;

[0098] When measured value Below the ecological critical threshold When, the first preset function is used; this function is an exponentially decaying function, characterized in that when The further away and below Utility value It will drop sharply, approaching zero; this simulates a scenario where arable land productivity or ecological health will plummet when key indicators fall below critical levels.

[0099] When measured value Not less than That is, greater than or equal to the ecological critical threshold. When the second preset function is used, this function is a non-linear growth function, characterized by the fact that when... Exceed Afterward, the utility value will increase rapidly, but the growth rate will gradually slow down and approach 1; this reflects the positive benefits brought about by the indicator exceeding the threshold, and also reflects the general law of diminishing marginal utility.

[0100] By introducing this nonlinear utility function, we can more realistically simulate and quantify the nonlinear response and threshold effect in the farmland ecosystem, overcoming the shortcomings of traditional linear weighted models that cannot accurately characterize such abrupt changes. This makes the calculation results of multi-scale ecological balance scores more sensitive and accurate to changes in key indicators, thereby improving the scientific nature and risk avoidance capabilities of site selection decisions.

[0101] A nonlinear assessment model capable of reflecting the intrinsic laws of ecosystems has been constructed. Existing technologies mostly use linear weighted summation to calculate the total score, which implicitly assumes that the contributions of each indicator to the overall quality are independent and linear. This contradicts real ecosystem behavior, especially failing to characterize threshold effects. This scheme addresses this problem on two levels: in calculating multi-scale ecological balance scores... At that time, a multiplicative aggregation model was adopted. The physical implication is that the overall health of the system depends on the synergistic performance of all key dimensions. A severe weakness in any dimension, i.e., extremely low utility value, will lead to a sharp drop in the overall score. This precisely corresponds to the bottleneck effect or limiting factor law in ecosystems. For key indicators with threshold effects, the final utility value... It is calculated using piecewise nonlinear functions; at the ecological critical threshold The following shows an exponential decline in utility:

[0102]

[0103] in, : Index measurement value, Critical threshold : Attenuation coefficient; above the threshold, it exhibits non-linear growth:

[0104]

[0105] in, : Index measurement value, Critical threshold : Growth coefficient; This design enables the model to sensitively capture abrupt changes in ecological state near the critical point, avoiding misjudgment of potential ecological risks due to linear approximation.

[0106] S5 specifically includes:

[0107] S51. Quantify data uncertainty based on the reliability of the data source;

[0108] S52. Using the Monte Carlo dropout technique, the uncertainty of the model is estimated by repeatedly deactivating a portion of neurons in the graphical neural network model and observing the output variance.

[0109] S53. Calculate the normalized standard deviation of the key indicators of the target plot in the spatial neighborhood of all plots, as the spatial uncertainty.

[0110] S54. Perform square root calculations on data uncertainty, model uncertainty, and spatial uncertainty to generate an overall uncertainty index;

[0111] In this embodiment, the process of generating the overall uncertainty index in step S5 is described in detail; this process is executed by the uncertainty quantification module, which aims to comprehensively evaluate the reliability of the decision from three dimensions: data, model, and space.

[0112] In step S51, the data uncertainty is quantified based on the reliability of the data source. The reliability of a data source is determined by a comprehensive score based on the data source's authority, measurement accuracy, spatial resolution, and the freshness of the timestamp, according to a pre-defined scoring rule. For example, authoritative ground station data from a national soil survey are assigned high reliability and low uncertainty, while data indirectly retrieved from low-resolution remote sensing imagery is assigned lower reliability and high uncertainty. By assigning uncertainty weights to data from different sources and combining them, a quantitative data uncertainty index is obtained. ;

[0113] The formula for comprehensive calculation can be the one used for the assessed land parcel. The formula for the weighted sum of squares and square root of the uncertainties of each data source is as follows:

[0114]

[0115] in, It is the first Normalized weights of each data source's contribution to the final evaluation. It is the uncertainty score of the data source itself, a value between [0,1], obtained by looking up its metadata such as source, accuracy, and timeliness, for example, national-level measured data. Low-resolution remote sensing interpretation data This formula can reasonably summarize the uncertainties of different data sources; k: the index of the kth data source;

[0116] In step S52, the Monte Carlo dropout technique is used to estimate the model uncertainty. Monte Carlo dropout is a widely used method in deep learning for estimating model uncertainty. In this embodiment, for the graph neural network model used in step S2, during prediction, neurons are randomly dropped with a preset probability, such as p=0.1, i.e., temporarily rendered inactive. This process is repeated multiple times, for example, N=50 times, each time yielding a slightly different dynamic evaluation weight vector. By calculating the variance of these N output results, the stability of the model prediction can be quantified, and the square root of this variance is defined as the model uncertainty. Therefore, the larger the standard deviation of the output, the more unstable the model's prediction for that plot of land, and the greater the model's uncertainty. The higher the level, the better;

[0117] In step S53, the normalized standard deviation of the key indicators of the target plot among all plots in the spatial neighborhood is calculated as the spatial uncertainty. This step aims to quantify the uncertainty caused by spatial heterogeneity. The spatial neighborhood is defined as all other plots within a certain geographical radius surrounding the target plot, such as 500 meters. One or more key indicators sensitive to spatial variability, such as soil organic matter, are selected, and the standard deviation of these indicators across the target plot and all its neighboring plots is calculated. A larger standard deviation indicates stronger spatial heterogeneity in the area and lower representativeness of measurements from individual plots, thus increasing spatial uncertainty. The higher the standard deviation, the more it needs to be normalized to eliminate the influence of dimensions.

[0118] In step S54, the sum of squares and square roots of the data uncertainty, model uncertainty, and spatial uncertainty are calculated to generate an overall uncertainty index. These three uncertainties are considered to be independent sources of error; they are combined using the sum of squares and square root method commonly used in error propagation theory, i.e., the Euclidean distance calculation method, as shown in the following formula:

[0119]

[0120] in, Due to overall uncertainty, Due to data uncertainty, Due to model uncertainty, Due to spatial uncertainty;

[0121] Through the specific steps described above, this invention can unify uncertainties from different sources and of different natures into a measurable framework. The resulting technical benefits are: it enables a comprehensive and quantitative assessment of the reliability of site selection decisions; it not only considers the quality of input data and the stability of the model itself, but also innovatively introduces consideration of spatial heterogeneity, so that the final overall uncertainty index can more accurately reflect the decision risk, providing a clear direction for the subsequent manual review process, and significantly enhancing the robustness and credibility of the entire intelligent site selection system.

[0122] This approach introduces probabilistic and uncertainty quantification throughout the entire decision-making process. Traditional site selection methods typically provide deterministic ranking results but do not offer a measure of the reliability of these results, leaving decision-makers unaware of the risks hidden behind the assessment outcomes. This solution establishes a comprehensive uncertainty quantification framework by calculating an overall uncertainty index. To achieve:

[0123]

[0124] in, Data uncertainty Model uncertainty Spatial uncertainty; the theoretical basis of this formula comes from the error propagation theory, which synthesizes the errors of three orthogonal dimensions, namely: data uncertainty, model uncertainty and spatial uncertainty; this makes each site selection suggestion come with a clear confidence interval; when the overall uncertainty index exceeds the preset threshold, the system will suggest manual review, thus establishing a human-machine collaborative risk management mechanism; this is an essential progress from giving an answer to giving an answer with confidence.

[0125] The comprehensive resilience index in S6 is obtained by weighted summation of the normalized values ​​of multiple resilience-related indicators, with the weights of each indicator determined by domain experts.

[0126] As an example of using the Delphi method to determine weights, for the resilience-related indicators mentioned in step S6, a set of consensus weights reached by 10 soil science and ecology experts were used. Examples are as follows:

[0127] Biodiversity Index: ;

[0128] Soil aggregate stability index: ;

[0129] Landscape connectivity index: ;

[0130] Soil moisture regulation capacity index: ;

[0131] Ensure that the sum of all weights is 1. These weight values ​​make the calculation process of the comprehensive resilience index completely certain;

[0132] In this embodiment, the comprehensive resilience index in step S6 is... The calculation method is explained in detail; the purpose of the comprehensive resilience index is to quantify the ability of arable land ecosystems to maintain their key functions and structures and recover from external disturbances such as drought, floods, and pests and diseases; plots with strong resilience have better long-term stability and sustainability;

[0133] In this embodiment, The calculation is a multi-standard weighted summation process; to further clarify, a series of resilience-related indicators directly related to ecological resilience are selected; these indicators may include: biodiversity index, soil aggregate stability index, landscape connectivity index, soil moisture regulation capacity index, etc.; the raw data of these indicators are all obtained from the basic database constructed in step S1;

[0134] The raw measurements of these indicators were normalized to dimensionless values ​​for comparison and weighting; the comprehensive resilience index was calculated using the following weighted summation formula:

[0135]

[0136] in, For comprehensive resilience index; j: index of the j-th resilience-related indicator; For the first The weight of each indicator, For the first Normalized values ​​of each indicator;

[0137] in, It is the first The normalized values ​​of several resilience-related indicators; the key lies in determining the weights, i.e. The weights of each indicator are determined by experts in the relevant field. These experts are scholars or technicians with extensive knowledge and practical experience in fields such as ecology, soil science, and agronomy. The weights can be determined using established expert scoring techniques such as the Delphi method or the analytic hierarchy process. Multiple experts independently score and evaluate the relative importance of each indicator for long-term resilience, reaching a consensus and ensuring that the sum of all weights equals 1. ;

[0138] By defining the calculation method of the comprehensive resilience index in this way, the technical benefits brought by this invention are: transforming the complex ecological concept of resilience into an operable and quantifiable evaluation indicator; by integrating multiple dimensions of specific indicators whose importance has been verified by experts, introducing a new evaluation dimension for site selection decisions that focuses on long-term stability and risk resistance, thereby making the final forward-looking decision score more comprehensive and far-sighted, and helping to select high-quality plots that are not only of excellent current quality, but can also maintain a healthy state in the long term;

[0139] This approach extends the decision-making perspective from current state assessment to future trend prediction and long-term resilience considerations. Existing technologies often only focus on the current snapshot of a site, lacking an assessment of long-term sustainability. This solution utilizes a Long Short-Term Memory (LSTM) network model to analyze the future quality evolution trend of a site. Predictions were made; furthermore, a comprehensive resilience index was constructed.

[0140]

[0141] in, The j-th resilience index, : Corresponding weights, used to quantify the land parcel's ability to resist and recover from external disturbances; the final forward-looking decision score. It is about the current quality Future Trends With resilience Linear weighting of the three;

[0142]

[0143] in, Strategic parameters; the practical significance of this model lies in providing a highly flexible decision-making framework, allowing decision-makers to adjust strategic parameters. The value of site selection involves weighing and optimizing multiple strategic goals, such as ensuring current production, pursuing future production potential, and ensuring long-term ecological stability, so that site selection decisions are not only scientific but also precisely aligned with macroeconomic policy objectives.

[0144] The reliability of the location recommendation in S7 includes:

[0145] S71. Compare the overall uncertainty index with the preset uncertainty threshold;

[0146] S72. When the overall uncertainty index is less than the uncertainty threshold, the site selection recommendation is determined to be reliable.

[0147] S73. When the overall uncertainty index is not less than the uncertainty threshold, the site selection recommendation will be marked as high uncertainty and manual review is recommended.

[0148] In this embodiment, the logic for determining the reliability of the location recommendation in step S7 is explained; this determination process is a key step in the system output quality control.

[0149] In step S71, the overall uncertainty index is compared with a preset uncertainty threshold; the overall uncertainty index is the one calculated in step S5. The value comprehensively reflects the uncertainty in three aspects: data, model, and space; the preset uncertainty threshold. It is a key decision parameter; the threshold is not set arbitrarily, but determined based on rigorous methods; in this embodiment, It is calibrated through backtesting analysis of a large number of historical site selection cases; the goal is to strike a balance between the efficiency of automated screening and the risk of false positives and false negatives; for example, it can... The setting is to correctly identify 95% of successful site selection cases as reliable in the historical case dataset. Statistical quantiles of values ​​are used to control decision-making risk;

[0150] In steps S72 and S73, explicit binary decision logic is executed:

[0151] When the overall uncertainty index is less than the uncertainty threshold, i.e. The system automatically determines that the site selection recommendation for the plot is reliable; this means that based on the currently available data and model analysis, the evaluation result of the plot, namely the forward-looking decision score, has a high degree of confidence and can directly proceed to the subsequent sorting and list generation stages.

[0152] When the overall uncertainty index is not less than, i.e., greater than or equal to the uncertainty threshold, that is... The system then marks the site selection suggestion as high uncertainty and recommends manual review. This does not mean that the site is necessarily unsuitable, but rather that the current automated assessment process cannot provide a sufficiently credible conclusion. After being marked as high uncertainty, the system will list it separately and may provide the main sources of uncertainty, such as data quality issues or excessive model prediction variance, so that domain experts can conduct targeted on-site inspections, supplement data collection, or model validation to make the final manual judgment.

[0153] Through the aforementioned clear reliability assessment process, a risk management mechanism combining automated evaluation and human supervision was established. This avoids blind trust in the model's output results and defines the boundaries of machine decision-making through quantified and evidence-based thresholds. This not only ensures that the final output list of site selection recommendations has high reliability, but also greatly improves the efficiency and relevance of expert manual review by accurately targeting cases with high uncertainty, thus optimizing human-machine collaborative decision-making.

[0154] Example 2:

[0155] A smart site selection system for balancing farmland occupation and compensation includes:

[0156] The data acquisition and preprocessing module is used to collect and preprocess multi-source spatiotemporal data to build a basic database;

[0157] The spatiotemporal weight adaptive learning module is used to construct a spatial adjacency graph and input data from the basic database into the graph neural network model to generate dynamic evaluation weight vectors for candidate plots.

[0158] The multi-scale ecological balance assessment module is used to calculate the final utility value of each assessment indicator and, in combination with the dynamic assessment weight vector, uses a multiplicative aggregation model to calculate the multi-scale ecological balance score of the plot.

[0159] The arable land quality evolution prediction module is used to predict the future quality evolution trend of land parcels using a long short-term memory network model.

[0160] The uncertainty quantification module is used to quantify data uncertainty, model uncertainty and spatial uncertainty respectively, and to perform comprehensive calculations on the above three uncertainties to generate an overall uncertainty index;

[0161] The decision support module is used to calculate the comprehensive resilience index, generate a forward-looking decision score, judge reliability based on the overall uncertainty index, and output a final list of site selection recommendations based on the forward-looking decision score.

[0162] The present invention also provides a smart site selection system for implementing the above method; the system can be deployed on a server or cloud platform and consists of a series of functionally independent but data-coupled software modules that jointly execute the smart site selection process;

[0163] Data Acquisition and Preprocessing Module: As the unified data entry point of the system, it is responsible for executing step S1 in the method. This module has multiple built-in data interfaces, which can access databases from different sources, such as land and resources databases, meteorological databases, and file formats such as Shapefile, GeoTIFF, and CSV. It also includes a series of data processing tools to achieve automated data cleaning, coordinate registration, time-series alignment, and normalization, generating a standardized basic database that can be called by other modules.

[0164] Spatiotemporal weight adaptive learning module: This is the system's weight dynamic generation unit, responsible for executing step S2. The core of this module is a weight learning engine based on graph neural networks. It first automatically constructs a spatial adjacency graph based on the geographic information of the plots, then calls a pre-trained GNN model to process the spatiotemporal data in the basic database, dynamically generates an evaluation weight vector for each candidate plot that changes with time and spatial context, and passes it to the multi-scale ecological balance assessment module.

[0165] Multi-scale ecological balance assessment module: responsible for executing step S3; this module receives dynamic weight vectors from the spatiotemporal weight module and multi-scale index data from the basic database; internally, it implements a multiplicative aggregation algorithm and embeds nonlinear utility functions for key indicators; the core function of this module is to calculate the multi-scale ecological balance score that comprehensively reflects the current quality of the land parcel.

[0166] Farmland quality evolution prediction module: responsible for executing step S4, providing a forward-looking perspective for decision-making; the core of this module is a pre-trained long short-term memory network model; it extracts historical time series data of land parcels from the basic database, predicts their future quality evolution trend, and outputs the quantified trend results to the decision support module;

[0167] Uncertainty Quantification Module: Responsible for executing step S5, this module is a technical component that ensures the reliability of the system output. It integrates three uncertainty assessment algorithms, corresponding to data uncertainty, based on data source metadata and model uncertainty, and based on Monte Carlo dropout technology and spatial uncertainty, respectively, and based on spatial statistics. It calculates the overall uncertainty index through a combined formula and provides it to the decision support module.

[0168] Decision Support Module: As the system's integration and decision output unit, it is responsible for executing steps S6 and S7. This module calculates the comprehensive resilience index based on expert weights. As an integrator, it linearly weights the current quality score, future trend, and resilience index input from other modules to generate the final forward-looking decision score. Simultaneously, it receives the output from the uncertainty quantification module and judges the reliability of the evaluation results for each plot based on preset thresholds. This module sorts all reliable site selection recommendations according to the forward-looking decision score, generates and visualizes the final list of site selection recommendations.

[0169] The system provided by this invention, through the aforementioned modular structural design, decomposes the complex intelligent site selection task into a series of clear and achievable functional units; the modules work collaboratively to fully realize the above-mentioned technical solution; the system deeply integrates advanced artificial intelligence algorithms with geographic information systems and ecological models, forming an automated, intelligent, and highly reliable decision support tool; the overall technical effect is: it provides an end-to-end intelligent site selection solution for arable land occupation and compensation balance, transforming the site selection process from a traditional model relying on subjective experience to a data-driven, model-supported scientific decision-making model, significantly improving the efficiency, accuracy, comprehensiveness, and foresight of site selection work.

[0170] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A smart site selection method for balancing arable land occupation and compensation, characterized in that, Includes the following steps: S1. Collect and preprocess multi-source spatiotemporal data to build a basic database; S2. Construct a spatial adjacency graph and input the data from the basic database into the graph neural network model to generate dynamic evaluation weight vectors for the candidate plots; S3. Calculate the final utility value of each evaluation indicator, and use a multiplicative aggregation model in conjunction with the dynamic evaluation weight vector to calculate the multi-scale ecological balance score of the plot. S4. Use a long short-term memory network model to predict the future quality evolution trend of the land parcel; S5. Quantify data uncertainty, model uncertainty and spatial uncertainty separately, and perform a comprehensive calculation on the above three uncertainties to generate an overall uncertainty index; S6. Calculate the comprehensive resilience index, and linearly weight the multi-scale ecological balance score, future quality evolution trend and comprehensive resilience index according to strategic parameters to generate a forward-looking decision score; S7. Determine the reliability of the site selection recommendations based on the overall uncertainty index, and rank the candidate sites according to the forward-looking decision score to output the final list of site selection recommendations. Multi-source spatiotemporal data include: geospatial information of historical land reclamation projects, soil physicochemical properties at the plot scale, landscape pattern indices based on remote sensing image interpretation, and regional-scale ecosystem service function data and climate change data.

2. The intelligent site selection method for balancing arable land occupation and compensation according to claim 1, characterized in that, S3 specifically includes: S31. For key indicators with threshold effects, obtain the normalized measured values ​​and the corresponding ecological critical thresholds. S32. When the measured value is lower than the ecological critical threshold, the final utility value with exponential decay is calculated using the first preset function; S33. When the measured value is not lower than the ecological critical threshold, the final utility value with non-linear growth is calculated using the second preset function.

3. The intelligent site selection method for balancing arable land occupation and compensation according to claim 1, characterized in that, S5 specifically includes: S51. Quantify data uncertainty based on the reliability of the data source; S52. Using the Monte Carlo dropout technique, the uncertainty of the model is estimated by repeatedly deactivating a portion of neurons in the graphical neural network model and observing the output variance. S53. Calculate the normalized standard deviation of the key indicators of the target plot in the spatial neighborhood of all plots, as the spatial uncertainty. S54. Perform square root calculations on the data uncertainty, model uncertainty, and spatial uncertainty to generate an overall uncertainty index.

4. The intelligent site selection method for balancing arable land occupation and compensation according to claim 1, characterized in that, The comprehensive resilience index in S6 is derived by weighted summation of the normalized values ​​of multiple resilience-related indicators, with the weights of each indicator determined by domain experts.

5. A smart site selection method for balancing arable land occupation and compensation according to claim 1, characterized in that, The reliability of the location recommendation in S7 includes: S71. Compare the overall uncertainty index with the preset uncertainty threshold; S72. When the overall uncertainty index is less than the uncertainty threshold, the site selection recommendation is determined to be reliable. S73. When the overall uncertainty index is not less than the uncertainty threshold, the site selection recommendation will be marked as high uncertainty and manual review is recommended.

6. A smart site selection system for balancing arable land occupation and compensation, applied to the smart site selection method for balancing arable land occupation and compensation as described in any one of claims 1-5, characterized in that, include: The data acquisition and preprocessing module is used to collect and preprocess multi-source spatiotemporal data to build a basic database; The spatiotemporal weight adaptive learning module is used to construct a spatial adjacency graph and input data from the basic database into the graph neural network model to generate dynamic evaluation weight vectors for candidate plots. The multi-scale ecological balance assessment module is used to calculate the final utility value of each assessment indicator and, in combination with the dynamic assessment weight vector, uses a multiplicative aggregation model to calculate the multi-scale ecological balance score of the plot. The arable land quality evolution prediction module is used to predict the future quality evolution trend of land parcels using a long short-term memory network model. The uncertainty quantification module is used to quantify data uncertainty, model uncertainty and spatial uncertainty respectively, and to perform comprehensive calculations on the above three uncertainties to generate an overall uncertainty index; The decision support module is used to calculate the comprehensive resilience index, generate a forward-looking decision score, judge reliability based on the overall uncertainty index, and output a final list of site selection recommendations based on the forward-looking decision score.

Images