Artificial intelligence-based county sustainable development land use prediction method and system

Abstract

The application discloses a county sustainable development land use prediction method and system based on artificial intelligence, and the method comprises the following steps: collecting county-level panel data and historical land use type data, and constructing a county city sustainable development index and a land use pattern index; taking the land use pattern index as an input feature and the county city sustainable development index as an output target, a GNNWR model is trained; a PLUS model is used to simulate different future land use scenarios, and the future land use pattern index is calculated according to the simulation result; and the future land use pattern index is input into the trained geographical neural network weighted regression model to obtain the county city sustainable development index prediction value under different future land use scenarios. The application breaks through the path dependence of traditional spatial econometric models on linear hypothesis and static weight matrix, realizes the collaborative analysis of complex nonlinear interaction, spatial dependence and spatio-temporal evolution trend of county man-land relationship.

Description

Technical Field

[0001] This invention belongs to the field of land use planning and relates to a land use prediction method and system, particularly to an artificial intelligence-based method and system for predicting land use for sustainable development in counties. Background Technology

[0002] Currently, quantitative simulation of land use change and its correlation analysis with the United Nations Sustainable Development Goals (SDGs) have become core technical directions in the field of land use planning. The core requirement is to establish a dynamic response relationship between land use patterns and sustainable development status through technological means to support precise regional policy formulation. However, existing technical systems still face significant bottlenecks in applications at the county / district level, specifically in three technical dimensions: data adaptation, modeling capabilities, and model performance.

[0003] (1) Insufficient compatibility between spatial scale and data processing technology

[0004] Existing sustainable development evaluation indicator systems are mostly designed based on the provincial scale, and their technical architecture does not fully consider the inherent characteristics of county-level data: In terms of data granularity, provincial indicators are mostly aggregated data, while county-level data requires more refined micro-indicators, resulting in a technical gap in data dimensions; In terms of data quality, county-level panel data generally suffers from problems such as low collection frequency, high missing rate, and inconsistent statistical periods. Existing data preprocessing techniques cannot effectively eliminate data heterogeneity, leading to a significant decrease in evaluation accuracy when provincial indicator systems are directly transferred to the county level, making it difficult to depict the true differences in sustainable development at the county level.

[0005] (2) Lack of technical functions for spatiotemporal dynamic modeling and scenario prediction

[0006] Existing evaluation systems are mostly static assessment frameworks, which have two major functional shortcomings in their technical design: First, they lack spatiotemporal dynamism, failing to integrate dynamic modeling modules based on time-series data and relying solely on static weighting with fixed weights, thus failing to reflect the spatiotemporal evolution of land use type transformation and sustainable development; Second, they lack scenario prediction capabilities, lacking technical coupling with land use simulation models, and existing research mostly stops at current status assessment, resulting in technical outputs that can only describe the current state and cannot provide forward-looking technical support for future land use planning in counties.

[0007] (3) Technical performance bottlenecks of traditional association models

[0008] The performance of traditional GWR and GTWR models is highly dependent on the rationality of the spatial / spatiotemporal weight matrix. However, existing construction methods are mostly based on preset rules and lack adaptive optimization algorithms, making it difficult to match the complex geographical features at the county scale, resulting in a disconnect between weight allocation and actual geographical processes. Traditional models have strict requirements for the completeness and temporal consistency of input data, but the sampling frequency differences that are common in county-level data can easily lead to technical problems such as gradient vanishing and poor convergence during model training, reducing prediction stability. Furthermore, traditional models are prone to multicollinearity, which increases the bias in parameter estimation and makes it difficult to accurately quantify the impact of land use on sustainable development.

[0009] In summary, existing technologies for analyzing the correlation between land use and sustainable development at the county and district levels suffer from problems such as poor data adaptation, lack of dynamic modeling, and weak model performance. There is an urgent need to construct a technical solution that takes into account both spatiotemporal dynamics and scenario prediction to meet the precision requirements of county-level land use planning. Summary of the Invention

[0010] To address the aforementioned problems, this invention provides an artificial intelligence-based method and system for predicting land use for sustainable development in counties.

[0011] The technical solution adopted in this invention is as follows:

[0012] An artificial intelligence-based method for predicting land use for sustainable development at the county level includes the following steps:

[0013] S1. Collect district and county-level panel data and historical land use type data, construct district and county-level urban sustainable development index using CRITIC weighting method, and construct district and county-level land use pattern index based on land use type data;

[0014] S2. Using the district / county-level land use pattern index as input features and the district / county-level urban sustainable development index as output targets, train a geographic neural network weighted regression model;

[0015] S3. Based on the historical land use type data, a patch primitive land use simulation model is used to simulate different future land use scenarios to obtain simulation results of land use patterns under different future scenarios. Based on the simulation results, the future county-level land use pattern index of the corresponding scenario is calculated.

[0016] S4. Input the future county-level land use pattern index into the trained geographic neural network weighted regression model to obtain the predicted values ​​of the sustainable development index of county-level cities under different future land use scenarios, thereby obtaining the future land use prediction results.

[0017] Furthermore, the district / county-level panel data includes multidimensional indicator data based on the United Nations Sustainable Development Goals, specifically including:

[0018] The indicators corresponding to the zero hunger target are: cultivated land area, output of major grains, and added value of the primary industry.

[0019] Indicators corresponding to the goals of sustainable cities and communities: GDP per capita, added value of the tertiary sector, built-up area, and population density;

[0020] The indicators corresponding to terrestrial ecological targets are: CO2 emissions, annual average PM2.5 concentration, and normalized vegetation index.

[0021] Furthermore, the specific steps for constructing the district / county-level urban sustainable development index using the CRITIC weighting method include:

[0022] The original data of each indicator are standardized using the range method to eliminate the dimensional differences between different indicators;

[0023] Based on the degree of dispersion of the standardized data of each indicator, a variability index is calculated, and the degree of dispersion is quantified by the fluctuation of the data deviating from the mean.

[0024] Calculate the correlation coefficient between any two indicators, and based on the correlation coefficient, quantify the degree of conflict between the indicators to obtain the representative value of conflict for each indicator.

[0025] Multiply the variability index of each indicator by the conflict representative value to obtain the information content of each indicator; determine the weight of each indicator based on the information content ratio of all indicators.

[0026] Based on the weights of each indicator and the corresponding standardized data, the urban sustainable development index of each district and county is obtained by weighted summation.

[0027] Furthermore, the district / county-level land use pattern index includes:

[0028] Indicators related to the zero hunger target include: farmland concentration index, intensity of farmland transfer in or out, and proportion of farmland used for land purposes.

[0029] Indicators related to sustainable urban and community goals: land expansion intensity, land use patch density, and average patch size of built-up land;

[0030] Indicators related to terrestrial ecological objectives include: proportion of ecological land, land use diversity, and largest ecological land patch.

[0031] Furthermore, the geographic neural network weighted regression model is built on the PyTorch framework and includes an input layer, several hidden layers, and an output layer. Except for the last hidden layer, all other hidden layers are followed by a ReLU activation function to introduce nonlinear characteristics into the model.

[0032] Furthermore, the training process of the geographic neural network weighted regression model includes:

[0033] A training dataset is constructed based on the district / county-level urban sustainable development index and the district / county-level land use pattern index. The training dataset is divided into a training set and a test set according to a preset ratio.

[0034] Using the county-level land use pattern index as the input feature and the county-level urban sustainable development index as the output target, the Adam optimizer and mean squared error loss function are used to iteratively train the model using the training set;

[0035] After each round of training, the model performance is verified using a test set until the model's decision coefficient on the test set meets a preset acceptable threshold, at which point training stops.

[0036] Furthermore, the different future land use scenarios include a natural growth scenario, a farmland protection scenario, and a sustainable development scenario. The natural growth scenario only restricts the conversion of impermeable land to other land types without adding additional constraints. The farmland protection scenario restricts the conversion of farmland to non-farmland types, but allows other land types to be converted to farmland. The sustainable development scenario superimposes the constraints of the ecological protection scenario and the farmland protection scenario, and restricts the conversion of farmland, forest land, and water areas to non-target types. The ecological protection scenario adds an ecological restriction zone on the basis of the natural growth scenario, and the ecological restriction zone is selected from the vector data of national and provincial nature reserves.

[0037] An artificial intelligence-based land use prediction system for sustainable development at the county level, used to implement the above method, includes:

[0038] Data acquisition and processing module: used to collect district and county-level panel data and historical land use type data, construct district and county-level urban sustainable development index using CRITIC weighting method, and construct land use pattern index based on land use type data;

[0039] GNNWR model training module: used to train a geographic neural network weighted regression model with the district / county-level land use pattern index as input features and the district / county-level urban sustainable development index as output target;

[0040] Land use simulation module: Based on the historical land use type data, it uses a patch primitive land use simulation model to simulate different future land use scenarios, obtain land use pattern simulation results, and calculate the future county-level land use pattern index for the corresponding scenario based on the simulation results;

[0041] Sustainable Development Prediction Module: This module is used to input the future county-level land use pattern index into a trained geographic neural network weighted regression model to obtain the predicted values ​​of the sustainable development index of county-level cities under different future land use scenarios.

[0042] A computer device, the computer device comprising:

[0043] One or more processors;

[0044] Memory, used to store one or more programs;

[0045] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described artificial intelligence-based method for predicting land use for sustainable development in counties.

[0046] A computer-readable storage medium storing computer instructions that, when executed by one or more processors, cause the one or more processors to perform the steps in the method described above.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] This invention establishes a county-level "land-SDG synergy index" for the first time by integrating multi-dimensional indicators from SDG2, SDG11, and SDG15, thus solving the problem of the lack of spatiotemporal dynamism in traditional assessment systems. Compared to the city-level SDG index, this invention refines the spatial scale to the county level, effectively compensating for the shortcomings of urban sustainable development research at the county level. Methodologically, this invention, for the first time, achieves a ternary coupling of artificial intelligence, urban sustainable development, and future land use simulation by training a Geographic Neural Network Weighted Regression (GNNWR) model. It also, for the first time, couples artificial intelligence with land use models to achieve a closed loop of "data modeling-scenario prediction-policy quantification." By predicting the sustainable development index under different future land use patterns using GNNWR, it breaks through the path dependence of traditional spatial econometric models on linear assumptions and static weight matrices. This enables the synergistic analysis of the complex nonlinear interaction, spatial dependence, and spatiotemporal evolution trends of human-land relationships at the county level. It can quantitatively evaluate the implementation effects of different policies, reveal the feedback mechanism between resource and environmental carrying capacity and the spatial pattern of the land in the process of county-level urbanization, and provide a decision-making inference mechanism with both spatial accuracy and temporal granularity for coordinating urban-rural integration and optimizing the spatial layout of the land. Furthermore, it provides planning guidance and suggestions for land spatial planning and urban sustainable development. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method in an embodiment of the present invention. Detailed Implementation

[0050] The technical solution of the present invention will be further described clearly and in detail below with reference to the accompanying drawings and specific examples.

[0051] This invention is mainly divided into three parts: constructing county-level land use and urban sustainable development indices, simulating land use under different future scenarios, and predicting urban sustainable development indices under different future scenarios. The first part involves collecting county-level panel data and constructing a historical urban sustainable development index based on the CRITIC weighting method. Simultaneously, a historical land use pattern index is constructed using historical land use data. The second part uses a GNNWR geographic neural network weighted regression model to train the two indices and establish the relationship between urban sustainable development and land use. The third part uses a PLUS model combined with historical land use data to simulate land use under different future scenarios. Finally, the land use simulation results are used to obtain the county-level urban sustainable development index under different future scenarios through the trained GNNWR model. This invention effectively fills the gaps in county-level urban sustainable development research and, for the first time, achieves a ternary coupling of artificial intelligence, urban sustainable development, and future land use simulation by training a neural network GNNWR model. It breaks through the path dependence of traditional spatial econometric models on linear assumptions and static weight matrices, and realizes the synergistic analysis of the complex nonlinear interaction, spatial dependence, and spatiotemporal evolution trends of human-land relations at the county level. It reveals the feedback mechanism between resource and environmental carrying capacity and national land spatial pattern in the process of county-level urbanization, providing a decision-making and extrapolation mechanism with both spatial accuracy and temporal granularity for coordinating urban-rural integration and optimizing national land spatial layout. Furthermore, it provides planning guidance and suggestions for national land spatial planning and urban sustainable development.

[0052] Example

[0053] An artificial intelligence-based method for predicting land use for sustainable development at the county level includes the following steps:

[0054] S1. Collect district and county-level panel data and historical land use type data, construct district and county-level urban sustainable development index using the CRITIC weighting method, and construct district and county-level land use pattern index based on land use type data.

[0055] In one embodiment of the present invention, based on the land-related development goals and expectations of SDG2 (Zero Hunger), SDG11 (Sustainable Cities and Communities), and SDG15 (Terrestrial Ecology) in the SDG Sustainable Development Goals, and combined with the theoretical basis of SDGs and the characteristics of actual development, an indicator system for the synergistic and coordinated development index of urban sustainable development goals is constructed from the perspective of land-based development and the coordinated development of economic, social, and ecological systems.

[0056] As shown in Table 1, the district and county-level panel data selected in this embodiment includes multi-dimensional indicator data based on the United Nations Sustainable Development Goals.

[0057] In constructing the SDG2 (Zero Hunger) target, this embodiment comprehensively selects arable land area, major grain output, and primary industry added value as core indicators from three dimensions: the resource base, output efficiency, and economic contribution of the agricultural production system. Arable land area reflects the land resource endowment for sustainable agricultural development in a region; major grain output directly measures the actual output capacity of the food production system and is related to the food accessibility target proposed by the Food and Agriculture Organization of the United Nations; and primary industry added value, through the proportion of agricultural GDP to regional GDP, systematically assesses the contribution of the agricultural economy to regional development.

[0058] In constructing the SDG11 (Sustainable Cities and Communities) indicators, this embodiment is based on the theory of coordinated urban system development, comprehensively considering two interactive dimensions: economic development and spatial carrying capacity (New Urban Academy of United Nations Human Settlements Program (UN-Habitat)). In the economic development dimension, per capita GDP and the added value of the tertiary sector are selected as core indicators to measure the efficiency of urban economic output and the progress of industrial upgrading, respectively. In the spatial carrying capacity dimension, built-up area and population density are selected as core indicators.

[0059] Regarding the construction of SDG15 (terrestrial ecology) targets, this embodiment constructs an indicator system based on the dual dimensions of environmental pressure regulation and ecological resource conservation, selecting carbon emissions, annual average PM2.5 concentration, and normalized difference vegetation index (NDVI) as core indicators to quantify the intensity of human activities' interference with the ecosystem.

[0060] Table 1. Sustainable Development Level Indicator System

[0061]

[0062] In this embodiment, the specific steps for constructing the county-level urban sustainable development index using the CRITIC weighting method include:

[0063] (1) Data standardization

[0064] The raw data for each indicator are standardized using the range method to eliminate the dimensional differences between different indicators and achieve data discretization. This can be expressed by the formula:

[0065] ,

[0066] in, It is the raw data. For standardized data, This represents the minimum value of the corresponding indicator. This represents the maximum value of the corresponding indicator.

[0067] (2) Calculate the variability index

[0068] Based on the dispersion of standardized data for each indicator, the standard deviation of the data is used as the cost of indicator variability to calculate the variability index. This can be expressed by the formula:

[0069] ,

[0070] in, Let be the variability index of the j-th indicator, which is the standard deviation. A larger value indicates that the data values ​​of each sample fluctuate more and are more dispersed under the j-th indicator; a smaller value indicates that the data is more concentrated and more stable. This is the value of the j-th indicator of the i-th sample after standardization. "Standardization" is to eliminate differences in units and orders of magnitude between different indicators, so that all indicators can be compared and calculated under the same standard. The subscript ij indicates that this is the value of the i-th row (sample) and j-th column (indicator) in the data matrix. It is the average (also called mean) of all sample values ​​of the j-th indicator, which represents the "center position" of the indicator data. It represents the deviation of a single sample value from its mean. It reflects how close a particular sample point is to the overall average level. This represents the summation operation over all values ​​from the first sample (i=1) to the s-th sample (i=s). Since the deviation was previously squared, the variance is in squared units of the original data. Taking the square root reverts the standard deviation to the same units as the original data, making the interpretation more intuitive.

[0071] (3) Calculate the conflict index

[0072] The correlation coefficient between any two indicators is calculated using the Pearson correlation coefficient method. Based on this correlation coefficient, the degree of conflict between the indicators is quantified inversely to obtain the representative conflict value for each indicator. This can be expressed by the formula:

[0073] ,

[0074] ,

[0075] The first formula is used to calculate the degree of linear correlation between the j-th indicator and the k-th indicator, and its value range is [-1, 1]. Let be the Pearson correlation coefficient between the j-th and k-th indicators. A value closer to 1 indicates a stronger linear correlation (positive or negative) between the two indicators; a value closer to 0 indicates a weaker linear correlation. i and m represent the total sample size being calculated for the i-th sample, with i cycling from 1 to m. In this study, m represents the total number of all participating districts, cities, or individuals. After standardization... Let be the value of the j-th indicator for the i-th sample. (') usually indicates that the data is standardized or has been normalized to eliminate the influence of dimensions. This is the value of the k-th index of the i-th sample after standardization. It is the arithmetic mean of all sample values ​​of the k-th indicator. is the arithmetic mean of all sample values ​​for the j-th indicator. These represent the deviations from the mean for the j-th and k-th indicators, respectively, which are the deviations between the value of each sample and its mean.

[0076] The second one is used to calculate the total conflict between the j-th indicator and all other indicators. This represents the conflict value of the j-th indicator. The larger this value, the weaker the correlation between the j-th indicator and other indicators. The smaller the value, the more independent the information it provides, and the less it can be replaced by other indicators. Therefore, its weight should be greater in the evaluation.

[0077] (4) Calculate the information content and weights

[0078] The information content of each indicator is obtained by multiplying its variability index by the conflict representative value. This can be expressed by the formula:

[0079] = =

[0080] in, This represents the amount of information contained in the j-th indicator. To measure the degree of variation (volatility) within the indicator. To measure the external independence of the indicator (the degree of conflict / difference with other indicators).

[0081] The weight of each indicator is determined based on the information content ratio of all indicators. This can be expressed by the formula:

[0082]

[0083] The above formula is a normalized formula used in the weighted comprehensive evaluation method to calculate the final weight of a certain indicator. Its purpose is to transform the originally calculated "importance" values ​​into formal weights that sum to 1. Among them, This represents the weight of the j-th indicator.

[0084] (5) Calculate the sustainable development index

[0085] Based on the weights of each indicator and their corresponding standardized data, the urban sustainable development index for each district / county is obtained through weighted summation. This can be expressed as:

[0086]

[0087] Let be the sustainable development index of the i-th district / county. Let be the weight of the j-th indicator. n is the number of evaluation indicators.

[0088] In existing technologies, methods for determining indicator weights include subjective weighting methods based primarily on expert experience, such as the analytic hierarchy process (AHP), the Delphi method, and comparative ranking, as well as objective weighting methods based on data characteristics and qualitative and quantitative analysis, such as principal component analysis (PCA), entropy methods, and the CRITIC method. The entropy method, based on the principle of information entropy, considers the discreteness of data and determines weights by calculating the information entropy of each indicator. It is suitable for situations with high data dispersion and can reflect the information content of each indicator, but it ignores the correlation and conflict between indicators. Principal component analysis focuses on considering the correlation between data, merging highly correlated indicators through dimensionality reduction to reduce redundant information, improve computational efficiency, and simplify the problem. However, it may lose the uniqueness of some individual indicators and has poor applicability to nonlinear data. The CRITIC method, proposed by Diakoulaki, considers both the correlation of indicator data and the comparative strength between data. It effectively handles the correlation and conflict between indicators and fully considers the variability of data. Unlike traditional weighting methods that only consider data variability, the CRITIC method chosen in this invention ensures a more objective and realistic weight allocation by quantifying conflicts between standards. Compared to traditional subjective weighting methods, the CRITIC method provides a more comprehensive allocation of indicator weights and achieves a better balance among multiple standards.

[0089] This embodiment classifies the sustainable development index of county-level cities according to ArcGIS natural breakpoint classification method, and the results are shown in Table 2.

[0090] Table 2. Division of Sustainable Development Index among 4 Cities in Counties and Districts

[0091] Urban Sustainable Development Index Urban Sustainable Development Level 0.001-0.303 Very low 0.304-0.351 Low 0.352-0.402 medium 0.403-0.473 high 0.474-0.929 Very high

[0092] In this embodiment, a land use pattern index based on SDG targets is also constructed, transforming the sustainable development goals from abstract concepts into a quantifiable and operable land spatial pattern index.

[0093] Regarding the SDG2 (Zero Hunger) goal, with agricultural sustainability at its core, indicators such as the Farmland Agglomeration Index (LUAI), farmland transfer-out / transfer intensity, and farmland use ratio have been constructed. These indicators collectively quantify the degree of intensification and stability of agricultural land use, providing a data foundation for ensuring food security. Spatial analysis optimizes farmland layout, promotes agricultural production efficiency and sustainable resource management, and helps achieve agricultural resilience under the "Zero Hunger" goal.

[0094] The calculation method for the Farmland Agglomeration Index (LUAI) is as follows:

[0095]

[0096] It refers to the area of ​​arable land. It represents the total area. This index reflects the spatial concentration of arable land in a certain region. High concentration means that arable land is concentrated in a large area, while low concentration means that it is scattered.

[0097] Intensity of farmland transfer out / in The calculation method is as follows:

[0098]

[0099] This index represents the percentage of arable land converted to other types (such as construction land and forest land) or other types converted to arable land.

[0100] The method for calculating the proportion of arable land is as follows:

[0101] Cultivated land ratio = 100%.

[0102] Regarding the SDG11 (Sustainable Cities) goals, with urban land use efficiency as the core, indicators such as land expansion intensity, land use patch density, and average construction land patch are adopted to incorporate urban expansion rate, land use structure transformation, and spatial morphology changes into a unified analysis framework, thereby promoting the intensive use of land resources and the synergy between urban sustainable development and urban development.

[0103] The method for calculating land expansion intensity is as follows:

[0104] 100%

[0105] I represents the intensity of urban built-up area expansion; and The areas of the urban built-up areas at the beginning and end of the study period are respectively; S is the total land area of ​​the study area; and T is the study duration.

[0106] The method for calculating land use patch density (PD) is as follows:

[0107]

[0108] NP is the number of patches, A is the total area of ​​the region (in square meters), multiplied by This is to convert the units to .

[0109] The method for calculating the average patch size (MPS) of construction land is as follows:

[0110]

[0111] Let n be the area of ​​the first type of construction land, and n be the total number of patches. For example, when counting construction land patches, the areas of all construction patches are added together and then divided by the total number of construction land patches to obtain the average area of ​​the construction land patches.

[0112] Regarding the SDG 15 (terrestrial ecology) goals, the focus is primarily on ecological protection and restoration. Based on indicators such as the proportion of ecological land, land use diversity, and the largest ecological land patch, a systematic assessment of the spatial integrity and heterogeneity of natural ecosystems is conducted. By quantifying the service value and carrying capacity of ecosystems, scientific support is provided for the subsequent implementation of ecological restoration projects, promoting a dynamic balance between terrestrial ecological protection and land resource development.

[0113] The method for calculating the proportion of ecological land is as follows:

[0114] 100%

[0115] The method for calculating land use diversity is as follows:

[0116]

[0117] It is a land use structure diversity index; Let be the proportion of the area of ​​the i-th land use type to the total area; n is the number of land use types. It is the natural logarithm. The natural logarithm of the proportion of the i-th land type is its logarithm. Perform the conversion. Let be the contribution of the i-th land type to the overall diversity, and its functional properties make it possible to distinguish rare types ( Small changes are more sensitive.

[0118] The method for calculating the Largest Patch Index (LPI) of ecological land is as follows:

[0119]

[0120] A is the area. The area of ​​the largest patch in the landscape. This represents the total area of ​​the landscape. A high LPI value (e.g., >50%) indicates that the landscape is dominated by a few large patches and may have high ecological stability (e.g., large areas of continuous forest or wetlands). A low LPI value (e.g., <10%) indicates severe landscape fragmentation, with small and scattered patches (e.g., scattered woodlands in farmland). This may lead to problems such as biodiversity loss and ecological degradation.

[0121] S2. Using the land use pattern index as the input feature and the county-level urban sustainable development index as the output target, train a geographic neural network weighted regression model.

[0122] Geographically Neural Network Weighted Regression (GNNWR) is an innovative spatial statistical model that combines neural networks with geographic weighted regression. It aims to address the limitations of traditional Geographically Weighted Regression (GWR) models due to spatial heterogeneity and nonlinear relationships. Its core principle is to construct a dynamic spatial weight matrix through the adaptive learning capability of neural networks, replacing the static weight calculation method based on fixed kernel functions in traditional GWR. This allows for a more accurate capture of the spatial non-stationarity and complex nonlinear characteristics of geographic element relationships. SWNN is the core component of the GNNWR model. Its function is to dynamically generate a spatially heterogeneous weight matrix through deep learning, replacing the static weight calculation based on fixed kernel functions (such as Gaussian kernels or bisquare kernels) in traditional GWR.

[0123] This invention simulates the spatial nonstationary relationship between the land use pattern index and the CLSDI:

[0124]

[0125] The formula represents the sustainable development index of the i-th district / county. Equal to its spatial position The determined benchmark value , For the k-th land use pattern index of the i-th district / county, there are three land use pattern indices ( , , ) respectively with their corresponding elastic coefficients that vary with spatial location ( , , The sum of the products of ) This is the random error term. Let be the spatial elasticity coefficient generated by the neural network, k=0,1,2,3…….

[0126] To accurately represent spatial distance, GNNWR uses a spatially weighted neural network (SWNN) to represent the weight kernel function. The SWNN takes the spatial distance between the point to be estimated and all modeled sample points as its input layer and the spatial weight matrix as its output layer. The CLSDI index is not directly output. The connection between them is indirect and consists of two steps. The first step is training, where the GNNWR model uses the spatial weight matrix generated by SWNN. To calibrate (fit) historical data, thereby obtaining coefficients that vary spatially. The weight matrix determines that geographically closer sample points have a greater impact on the coefficients calibrating the current point i. The second step is prediction. During prediction, the model no longer needs to recalculate the weights, but directly uses the historical coefficients trained in the first step, which already contain spatial relationships. , and future independent variables Multiply the results to calculate the future CLSDI exponent. Finally, select an appropriate number of hidden layers based on modeling requirements. The spatial weights of the point i to be estimated are calculated as follows:

[0127]

[0128] i spatial weight vector , is its geographical coordinates The function is used to calculate the vector, which is obtained through a spatially weighted neural network (SWNN). Let s represent the spatial distance between point i and sample point j, where i is the index of the point to be estimated and j is the index of the sample point (j ranges from 1 to n). The superscript s is an abbreviation for "spatial," emphasizing that this is a spatial distance. The input to this neural network is the spatial distance vector from point i to all n modeling sample points. The output is the desired spatial weight vector.

[0129] After establishing the relationship between the county-level urban sustainable development index and the land use pattern index using the GNNWR model, the future relationship between the county-level urban sustainable development index and the land use pattern index can be predicted based on the model:

[0130]

[0131] in, Let CLSDI be the predicted value for county i in a future year. Estimating coefficients for historical GNNWR, This represents the land use pattern index for a future year.

[0132] In this embodiment, the parameter settings of the geographic neural network weighted regression model are specifically as follows:

[0133] Dataset: The training dataset used in this embodiment contains land pattern data for 300 districts and counties in a certain region from 1999 to 2022. This dataset captures the spatiotemporal dynamics of land use change, which is crucial for understanding regional land use patterns and their influencing factors. The dataset covers a variety of features related to land use types, socioeconomic indicators, and geographical characteristics, providing a comprehensive basis for model training and evaluation.

[0134] Model Architecture: The model is implemented using the PyTorch framework. The neural network architecture includes an input layer, several hidden layers, and an output layer, with the hidden layers configured as [32, 64, 128, 64]. Specifically, except for the last layer, each hidden layer is followed by a ReLU (Rectified Linear Unit) activation function to introduce non-linear characteristics into the model, enabling it to learn complex patterns in the data. The input layer is designed to adapt to the features of the land pattern data, while the output layer generates predicted values ​​corresponding to the target variable (which may be a continuous land use indicator such as land use intensity or land cover change rate).

[0135] Training configuration: The model was trained for 4000 epochs. The Adam optimizer was used to minimize the loss function, and the learning rate was set to 1×10⁻⁶. -5 This value was determined through prior experiments to balance training speed and convergence. The loss function during training is the mean squared error (MSE), which is suitable for regression problems and measures the squared mean of the error between the predicted and actual values.

[0136] Evaluation Metric: The coefficient of determination (R² score) is used as the primary metric for evaluating model performance. The R² score measures the proportion of the variance of the dependent variable that can be explained by the independent variables. In this embodiment, the model's final R² score on the validation set is 0.78, indicating that the model fits the data well. The validation set is typically a subset of the dataset (e.g., 20% of the data) and is not used during training to ensure an unbiased evaluation of the model's generalization ability.

[0137] S3. Based on the historical land use type spatial data, the Patch-based Land-Use Simulation model (PLUS) is used to simulate different future land use scenarios to obtain land use pattern simulation results. Based on the simulation results, the future land use pattern index of the corresponding scenario is calculated.

[0138] In this invention, the PLUS model is mainly used to simulate future land use. As an important tool for land use simulation, the PLUS model combines cellular automata (CA) model with a multi-agent system to capture spatial heterogeneity and complex interactions between patches. The CA component simulates local land use changes based on neighborhood rules, while the multi-agent system considers the behavioral decisions of different land use agents (such as urban development and agricultural activities) to more realistically present land use dynamics. Compared with other relevant land use simulation models: FLUS, while based on the fuzzy logic CA framework, excels at handling uncertain information, but lacks explicit consideration of subject behavior and is insufficient in simulating human-driven changes; Markov chain models, while simple and intuitive, ignore spatial interactions and treat pixels independently; Mixed-cell automata (Mixed-cellCA) models significantly improve the simulation capability of complex geospatial heterogeneity by introducing heterogeneous cells and mixing rules, and are suitable for dynamic modeling of land use change, urban growth, ecological processes, etc. under multi-objective and multi-constraint conditions. However, its advantages are based on detailed data support, complex parameter calibration, and efficient computing resources, and are more suitable for micro- and meso-scale scenarios with rich data and diverse objectives, and are not applicable to this invention. The PLUS model, with its ability to integrate spatial, temporal, and agent-related factors, can more comprehensively represent the real-world decision-making process in complex land use simulations, achieving spatially explicit and accurate simulations. The simulation process encompasses data preparation (collecting and processing historical land use maps, socioeconomic and environmental variables, etc.), parameter calibration (using historical data to determine parameters such as migration probabilities), and simulation prediction (iteratively updating pixel states according to rules and agent behavior to simulate future scenarios). Its advantages are significant; it not only overcomes the scaling effect of traditional CA models, making it suitable for large-scale, fine-grained simulations, but also dynamically incorporates multiple driving factors, adapts to regional characteristics, accurately handles spatial heterogeneity, and captures the spatial variability of land use change.

[0139] In this embodiment, the different future land use scenarios include natural growth scenario, farmland protection scenario, and sustainable development scenario.

[0140] The Natural Growth Scenario (NIS) serves as a benchmark reflecting the future spatial structure of land use in a region under the current model. It is based on historical land use data and does not consider the guiding or constraining effects of policy factors. In the matrix setup, only the conversion of impervious surfaces to other land use types is restricted, while the conversion of other land use types is not subject to spatial constraints.

[0141] The Farmland Protection Scenario (CPS) aims to protect farmland resources. Based on the existing farmland area in the 2022 land use data, it extracts farmland restriction zones. Under this scenario, the conversion of farmland to other land types is restricted, while other land types are allowed to be converted into farmland.

[0142] The Sustainable Development Scenario (SDS) adds ecological restriction zones to the natural growth scenario, balancing ecological protection, farmland conservation, and economic development. This embodiment overlays the restriction zones from the ecological protection and farmland conservation scenarios, integrating them as a new constraint factor. Simultaneously, based on the restrictions on land type conversion in the cost matrices of the ecological protection and farmland conservation scenarios, other parameter settings are appropriately adjusted according to the actual conditions of each zone. The restricted zones for nature reserves are selected from vector surface data of the distribution of national and provincial nature reserves in 2020. Based on the parameter settings of the natural growth scenario, the conversion of forest land and water areas to other land types is restricted.

[0143] S4. Input the future land use pattern index into the trained geographic neural network weighted regression model to obtain the predicted values ​​of the county-level city sustainable development index under different future land use scenarios, thereby obtaining the future land use prediction results.

[0144] In one embodiment of the present invention, OLS, GWR, GTWR, and GNNWR models are applied to a simulated dataset for analysis. A test set containing 15% of the dataset is randomly selected to fairly evaluate the applicability of the models, while the remaining 85% of the dataset is used for training and validation to construct regression models. This embodiment uses four performance evaluation metrics—R², RSS, RMSE, and MSE—to quantitatively evaluate the regression accuracy and fit of the models, and compares the performance of the four models. Table 3 shows the specific regression evaluation metrics for the four models.

[0145] Table 3 Evaluation Indicators for Different Models

[0146] Model R² RSS RMSE MSE OLS 0.299 26.690 0.300 0.090 GWR 0.606 14.989 0.225 0.050 GTWR 0.686 11.969 0.201 0.040 GNNWR 0.785 0.540 0.043 0.002

[0147] The closer the R² (coefficient of determination) value is to 1, the stronger the model's explanatory power. For the same dataset, the GNNWR model has the highest R² value (0.785), indicating its strongest explanatory power. The smaller the RSS (residual sum of squares) and MSE (mean squared error) values, the smaller the model's prediction error. The GNNWR model has the smallest RSS and MSE values ​​(11.969 and 0.002, respectively), indicating its smallest error. RMSE (root mean squared error) measures the average absolute deviation between the predicted and actual values; the smaller the value, the higher the model's accuracy. The GNNWR model has the smallest RMSE, therefore its accuracy is the highest. In summary, the GNNWR model performs best in terms of explanatory power, fit, and prediction error. Therefore, it can be considered that the GNNWR model outperforms traditional OLS, GWR, and GTWR models in both training and testing.

[0148] This invention constructs a county-level urban sustainable development index and a land use index, trains a GNNWR model, and establishes the relationship between the two indices. Then, a PLUS model is introduced to simulate land use under different future scenarios. The simulation results are combined with the GNNWR model to predict the CLSDI under different future scenarios. This invention provides an AI-driven framework that integrates the GNNWR and PLUS models, achieving the coupling of urban sustainable development and land use simulation. This invention overcomes the limitations of traditional regression models due to spatial heterogeneity and nonlinear relationships, clarifies the synergistic relationship between land spatial patterns and sustainable development goals, and completes a paradigm shift from static assessment to AI-driven adaptive decision-making, providing support for refined national spatial planning.

[0149] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0150] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0151] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0152] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0153] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.

Claims

1. A method for predicting land use for sustainable development at the county level based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect district / county-level panel data and historical land use type data, construct a district / county-level urban sustainable development index using the CRITIC weighting method, and construct a district / county-level land use pattern index based on the land use type data; the district / county-level panel data includes multi-dimensional indicator data based on the United Nations Sustainable Development Goals, specifically including: The indicators corresponding to the zero hunger target are: cultivated land area, output of major grains, and added value of the primary industry. Indicators corresponding to the goals of sustainable cities and communities: GDP per capita, added value of the tertiary sector, built-up area, and population density; The indicators corresponding to terrestrial ecological targets are: CO2 emissions, annual average PM2.5 concentration, and normalized vegetation index. The county-level land use pattern index includes: Indicators related to the zero hunger target include: farmland concentration index, intensity of farmland transfer in or out, and proportion of farmland used for land purposes. Indicators related to sustainable urban and community goals: land expansion intensity, land use patch density, and average patch size of built-up land; Indicators related to terrestrial ecological objectives: proportion of ecological land, land use diversity, and largest ecological land patch; The specific steps for constructing the county-level urban sustainable development index using the CRITIC weighting method include: The original data of each indicator are standardized using the range method to eliminate the dimensional differences between different indicators; Based on the degree of dispersion of the standardized data of each indicator, a variability index is calculated, and the degree of dispersion is quantified by the fluctuation of the data deviating from the mean. Calculate the correlation coefficient between any two indicators, and based on the correlation coefficient, quantify the degree of conflict between the indicators to obtain the representative value of conflict for each indicator. Multiply the variability index of each indicator by the conflict representative value to obtain the information content of each indicator; determine the weight of each indicator based on the information content ratio of all indicators. Based on the weights of each indicator and the corresponding standardized data, the urban sustainable development index of each district and county is obtained by weighted summation; S2. Using the district / county-level land use pattern index as the input feature and the district / county-level urban sustainable development index as the output target, train a geographic neural network weighted regression model; the geographic neural network weighted regression model is built based on the PyTorch framework and includes an input layer, several hidden layers and an output layer. Except for the last hidden layer, all other hidden layers are followed by the ReLU activation function to introduce nonlinear characteristics into the model. S3. Based on the historical land use type data, a patch primitive land use simulation model is used to simulate different future land use scenarios, obtaining simulation results of land use patterns under different future scenarios. Based on the simulation results, the future county-level land use pattern index for the corresponding scenario is calculated. The different future land use scenarios include a natural growth scenario, a farmland protection scenario, and a sustainable development scenario. The natural growth scenario only restricts the conversion of impermeable land to other land types without adding additional constraints. The farmland protection scenario restricts the conversion of farmland to non-farmland types, but allows other land types to be converted to farmland. The sustainable development scenario superimposes the constraints of the ecological protection scenario and the farmland protection scenario, and restricts the conversion of farmland, forest land, and water areas to non-target types. The ecological protection scenario adds an ecological restriction zone on the basis of the natural growth scenario. The ecological restriction zone is selected from the vector data of national and provincial nature reserves. S4. Input the future county-level land use pattern index into the trained geographic neural network weighted regression model to obtain the predicted values ​​of the sustainable development index of county-level cities under different future land use scenarios, thereby obtaining the future land use prediction results.

2. The artificial intelligence-based land use prediction method for sustainable development at the county level according to claim 1, characterized in that, The training process of the geographic neural network weighted regression model includes: A training dataset is constructed based on the district / county-level urban sustainable development index and the district / county-level land use pattern index. The training dataset is divided into a training set and a test set according to a preset ratio. Using the county-level land use pattern index as the input feature and the county-level urban sustainable development index as the output target, the Adam optimizer and mean squared error loss function are used to iteratively train the model using the training set; After each round of training, the model performance is verified using a test set until the model's decision coefficient on the test set meets a preset acceptable threshold, at which point training stops.

3. An artificial intelligence-based county-level sustainable land use prediction system, used to implement the method as described in any one of claims 1 to 2, characterized in that, include: Data acquisition and processing module: used to collect district and county-level panel data and historical land use type data, construct district and county-level urban sustainable development index using CRITIC weighting method, and construct district and county-level land use pattern index based on land use type data; GNNWR model training module: used to train a geographic neural network weighted regression model with the district / county-level land use pattern index as input features and the district / county-level urban sustainable development index as output target; Land use simulation module: Based on the historical land use type data, it uses a patch primitive land use simulation model to simulate different future land use scenarios, obtain land use pattern simulation results, and calculate the future county-level land use pattern index for the corresponding scenario based on the simulation results; Sustainable Development Prediction Module: This module is used to input the future county-level land use pattern index into a trained geographic neural network weighted regression model to obtain the predicted values ​​of the sustainable development index of county-level cities under different future land use scenarios.

4. A computer device, characterized in that, The computer device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the artificial intelligence-based land use prediction method for sustainable development in counties as described in any one of claims 1 to 2.

5. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed by one or more processors, the one or more processors are caused to perform the steps of the method according to any one of claims 1 to 2.

Images