A land use spatio-temporal evolution simulation method and system

Abstract

The application relates to the technical field of land use simulation, and discloses a land use space-time evolution simulation method and system, which unifies multi-period land use grid data, driving factor data, ecological constraint data and carbon density data into a space-time feature tensor, extracts neighborhood spatial structure features by using convolution calculation, extracts the nonlinear trend of land use state evolution over time by using long short-term memory calculation, makes the initial growth probability of each pixel to different land use types simultaneously reflect spatial texture features and time evolution rules, then solves a target demand vector under multi-target constraints and feeds back to a patch competition allocation process, combines neighborhood effects, demand inertia and spatial restriction conditions to calculate comprehensive competition potential, makes the spatial allocation process be simultaneously controlled at the pixel level and the patch level, and finally improves the integrity, continuity and realizability of the simulation result through a demand deviation driven feedback updating mechanism.

Description

Technical Field

[0001] This invention relates to the field of land use simulation technology, and in particular to a method and system for simulating the spatiotemporal evolution of land use. Background Technology

[0002] Land use / cover change simulation is a crucial technical component in national spatial planning, ecological protection assessment, urban expansion analysis, and carbon emission scenario projection. Patch-based land use simulation models can predict regional land use spatial patterns through land expansion rule mining and patch evolution simulation, thus finding applications in regional planning and geographic information analysis. Among related technologies, land use simulation typically employs PLUS models or cellular automata-based spatial allocation models. These models often rely on random forest algorithms or suitability assessment methods in the rule mining stage to perform point-by-point probability calculations for pixels; in the demand forecasting stage, they often use Markov chains, linear regression, or manually defined scenario areas; and in the spatial allocation stage, the demand area is then allocated to specific pixels.

[0003] However, the aforementioned technical solutions still suffer from insufficient spatial neighborhood structure recognition. For processes such as urban boundary expansion, farmland fragmentation, and changes in ecological patch connectivity, single-pixel-level probability calculations are insufficient to fully express spatial information such as similar clustering, dissimilar contact, patch boundaries, and local textures, easily resulting in isolated pixels or fragmented patches. Furthermore, these technologies are weak in expressing long-term trends, lag effects, and sudden changes over time. Land use evolution is often influenced by policy adjustments, population flows, transportation construction, climate fluctuations, and changes in industrial structure; the assumption of a single stationary process cannot accurately reflect the nonlinear relationships between land use states across multiple time phases. In addition, macro-level land demand and micro-level spatial allocation are often disconnected in these technologies. The target land area may quantitatively meet the scenario setting, but lacks a spatially convertible basis; or some areas may have high spatial conversion potential, but cannot be directly converted due to total quantity control, ecological protection, farmland protection, and carbon storage targets.

[0004] Therefore, how to improve the synergistic expression of spatial neighborhood structure, temporal evolution law and macro-demand constraints in land use simulation remains a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This invention provides a method and system for simulating the spatiotemporal evolution of land use, aiming to solve the technical problems of insufficient capture of spatial neighborhood structure, weak temporal nonlinearity representation ability, and disconnect between macroscopic demand solution and microscopic spatial allocation in existing land use simulation technologies.

[0006] To achieve the above objectives, the present invention provides a method for simulating the spatiotemporal evolution of land use, the method comprising: Unified raster processing was performed on the multi-phase land use raster data and driving factor raster data of the study area to construct the spatiotemporal feature tensor of the target pixels; The convolutional long short-term memory network is invoked to perform spatial convolution and temporal memory calculations on the spatiotemporal feature tensor, and the growth tendency probability surface of each pixel facing the candidate land use type is output. The growth tendency probability surface is integrated with neighborhood effect, spatial constraint and land type inertia to form a comprehensive conversion potential quantification value, and candidate conversion units are determined. The comprehensive conversion potential quantification value is configured as a conversion parameter shared by multi-objective demand solution and patch competition allocation. Perform multi-objective demand solving on the candidate conversion units to obtain the target land demand vector; The target land demand vector is used to drive patch competition allocation, and cell state transitions and bias feedback updates are performed on the global raster to output the land use simulation pattern for the target year.

[0007] Furthermore, to achieve the above objectives, the present invention also provides a land use spatiotemporal evolution simulation system, comprising: The module is used to perform unified raster processing on multi-period land use raster data and driving factor raster data of the study area to construct the spatiotemporal feature tensor of the target pixel. The calling module is used to call the convolutional long short-term memory network to perform spatial convolution calculation and temporal memory calculation on the spatiotemporal feature tensor, and output the growth tendency probability surface of each pixel facing the candidate land use type. The determination module is used to fuse the growth tendency probability surface with neighborhood effect, spatial constraint and land type inertia into a comprehensive conversion potential quantification value, and to determine candidate conversion units, wherein the comprehensive conversion potential quantification value is configured as a conversion parameter shared by multi-objective demand solution and patch competition allocation; The solver module is used to perform multi-objective demand solving on candidate transformation units to obtain the target land demand vector; The update module is used to drive patch competition allocation using the target land demand vector, perform cell state transitions and bias feedback updates on the global raster, and output the land use simulation pattern for the target year.

[0008] One or more technical solutions proposed in this application have at least the following technical effects: Compared with related technologies that only use random forests, Markov chains, or static cellular automata for allocation, this application uses spatiotemporal feature tensors as a unified input carrier, so that historical land use status, spatial neighborhood structure, and multiple driving factors enter the same expression structure, which can reduce the spatial structure loss caused by isolated pixel-level judgments.

[0009] Compared with the separation of macro-level demand area and micro-level spatial probability in related technologies, this application uses the comprehensive conversion potential quantification value as a common conversion parameter for both land demand solution and patch competition allocation, so that the target land demand vector matches the spatial conversion capability and reduces the disconnect between macro-level quantitative targets and micro-level spatial realization.

[0010] Compared with the static allocation rules in related technologies, this application drives the inertia coefficient update by the gap in the number of land categories, so that each land use type can make feedback adjustments around the target land demand vector during the iterative allocation process, thereby improving the synchronous response capability of spatial pattern evolution to total volume control and patch morphology. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the first embodiment of the land use spatiotemporal evolution simulation method of this application.

[0014] Figure 2 This is a schematic diagram of the spatiotemporal feature tensor construction process in the land use spatiotemporal evolution simulation method of this application.

[0015] Figure 3 This is a schematic diagram of the CNN-LSTM growth tendency prediction process in the land use spatiotemporal evolution simulation method of this application.

[0016] Figure 4 This is a schematic diagram of the data flow process in the land use spatiotemporal evolution simulation method of this application.

[0017] Figure 5 This is a schematic diagram of the structure provided for Embodiment 2 of the land use spatiotemporal evolution simulation system of this application.

[0018] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0020] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0021] The main solution of this application embodiment is as follows: A unified raster processing is performed on multi-period land use raster data and driving factor raster data of the study area to construct a spatiotemporal feature tensor for the target pixels; a convolutional long short-term memory network is invoked to perform spatial convolution and temporal memory calculations on the spatiotemporal feature tensor, outputting the growth tendency probability surface of each pixel towards candidate land use types; the growth tendency probability surface is fused with neighborhood effect, spatial constraint, and land type inertia to represent a comprehensive conversion potential quantification value, and candidate conversion units are determined, wherein the comprehensive conversion potential quantification value is configured as a conversion parameter shared by multi-objective demand solution and patch competition allocation; multi-objective demand solution is performed on the candidate conversion units to obtain the target land demand vector; the target land demand vector is used to drive patch competition allocation, and pixel state transformation and bias feedback updates are performed on the global raster to output the land use simulation pattern for the target year.

[0022] In related technologies, land use simulation methods typically handle land demand forecasting and spatial allocation processes separately. The demand forecasting process focuses on the quantity of various types of land, while the spatial allocation process focuses on pixel conversion probabilities. The lack of a unified conversion parameter between the two can lead to the target demand area not being able to be successfully achieved spatially.

[0023] This application uses the spatiotemporal feature tensor as a unified input carrier, enabling historical land use status, spatial neighborhood structure, and multiple driving factors to enter the same expression structure. In the demand solution stage, a pre-set initial land type inertia is used to form a comprehensive conversion potential quantification value, which is then used for candidate conversion unit identification and target land demand solution. After obtaining the target land demand vector, the land type inertia updated by deviation feedback is used for pixel state transformation in the patch competition allocation stage. This reduces the disconnect between macro-quantitative targets and micro-spatial realization, and improves the synchronous response capability of spatial pattern evolution to total quantity control and patch morphology.

[0024] It should be noted that the executing entity in this embodiment can be a computing device with raster data processing, model calculation, spatial analysis, and result output functions, such as a server, workstation, personal computer, cloud computing platform, or electronic device with geographic information processing software installed. The following uses a land use spatiotemporal evolution simulation system as an example to describe this embodiment and the subsequent embodiments.

[0025] Example 1 Based on this, the embodiments of this application provide a method for simulating the spatiotemporal evolution of land use, referring to... Figure 1 , Figure 1This is a flowchart illustrating the first embodiment of the land use spatiotemporal evolution simulation method of this application.

[0026] In this embodiment, the land use spatiotemporal evolution simulation method includes steps S100 to S500.

[0027] Step S100: Perform unified raster processing on the multi-period land use raster data and driving factor raster data of the study area to construct the spatiotemporal feature tensor of the target pixel.

[0028] It should be noted that multi-period land use raster data refers to land use type raster layers formed in the study area over multiple historical periods. This land use raster data is denoted as... ,in This represents a historical period. Driving factor raster data can include data such as DEM, slope, population, GDP, transport accessibility, ecological constraints, and carbon density. Among these, DEM, slope, and carbon density are numerical data related to natural geography or the ecological environment, while population, GDP, transport accessibility, and ecological constraints can be represented as quantified spatial characteristic values ​​during calculation. These data serve as external conditional features influencing land use change, enabling the model to learn the spatial-temporal correlation between different driving conditions and changes in pixel state.

[0029] It should be noted that the driving factors describe the external conditions affecting land use change. Although their data sources or meanings are related to human activities, this application embodiment does not utilize their management rules or business rules attributes. Instead, it quantifies them as model input channels to participate in convolutional feature extraction, temporal memory calculation, and conversion potential calculation, so as to characterize the external conditions affecting land use change.

[0030] Specifically, the land use spatiotemporal evolution simulation system first reads at least three periods of land use raster data within the study area. and matching The land use raster data consists of several driving factors. Each period of land use raster data includes multiple pixels, and each pixel corresponds to a land use type label. The land use type can be cultivated land, forest land, grassland, water area, construction land, unused land, etc. The driving factor raster data corresponds spatially to the land use raster data and is used to provide each pixel with topographic, location, economic, transportation, ecological, or carbon density attributes.

[0031] When reading data, the system performs an integrity check on each raster layer. If a layer has inconsistencies in coordinate system, raster range, cell size, or number of rows and columns, the system marks it as a layer requiring unified processing. This check ensures that each data layer participating in tensor assembly can be aligned at the same cell location, avoiding spatial misalignment between the target cell and neighboring cells due to differences in data sources.

[0032] Subsequently, the system performs resampling processing on all land use raster data and driving factor raster data. This resampling process transforms data from different sources, with different spatial resolutions, or different raster ranges to a unified spatial reference. For categorical land use raster data, a resampling method that maintains stable category labels can be used; for continuous driving factor data, a resampling method suitable for continuous variables can be used. After resampling, pixels at the same row and column position can simultaneously correspond to historical land use status and various driving factor values.

[0033] After resampling, the system performs Z-score standardization on the resampled continuous driving factors to bring data with different dimensions to the same numerical scale. For example, driving factors such as population, GDP, DEM, slope, and accessibility have different physical meanings and numerical ranges. If directly input into the model, factors with large numerical ranges can easily have an excessive impact on model training. Therefore, the system standardizes the resampled continuous driving factors so that they can be used as input data in the same feature channel system.

[0034] For categorical variables, the system uses one-hot encoding or equivalent discrete encoding. For example, historical land use types are inherently categorical; the system can convert them into several category channels or use discrete encoding that can distinguish between different land use types. For restricted areas such as ecological protection zones, basic farmland, and construction restriction zones, the system uses masks or restriction factors for labeling. These masks or restriction factors are used in subsequent spatial allocation processes to indicate whether pixels are subject to development bans, restricted conversions, or protection constraints.

[0035] After completing the unified raster processing, the system extracts a neighborhood window of a preset size centered on the target pixel. This neighborhood window is used to represent the spatial structure within a certain range around the target pixel. If there are a large number of similar pixels around the target pixel, it indicates that the target pixel has a strong spatial clustering relationship with similar patches; if the target pixel is located at the boundary of multiple land cover types, it indicates that the target pixel is located at the boundary of a patch or in a region where different land cover types meet. Through neighborhood window extraction, the system can preserve local spatial pattern information in the tensor.

[0036] Specifically, for each target pixel, the system sequentially extracts land use status from multiple historical periods, the corresponding driving factor values, and spatial constraints, and assembles this information according to the time dimension, attribute channel dimension, and spatial neighborhood dimension. The time dimension is used to arrange data from multiple historical periods; the attribute channel dimension is used to arrange features such as land use status, topography, socio-economic conditions, transportation accessibility, ecological constraints, and carbon density; and the spatial neighborhood dimension is used to record the local spatial structure of each location within the window surrounding the target pixel.

[0037] The spatiotemporal feature tensor obtained through the above processing is as follows:

[0038] in, For the spatiotemporal feature tensor, Represents a multidimensional array of real numbers. For time steps, The number of feature channels, and This refers to the spatial dimensions of the neighboring window. For example... Figure 2 As shown, through this tensor, the historical evolution information, local neighborhood association information, and external driving condition information of the target pixel are assembled into unified input data.

[0039] Step S200: Invoke the convolutional long short-term memory network to perform spatial convolution calculation and temporal memory calculation on the spatiotemporal feature tensor, and output the growth tendency probability surface of each pixel facing the candidate land use type.

[0040] It should be noted that the convolutional long short-term memory network is used to extract spatial neighborhood structure and multi-temporal evolution patterns. Specifically, the convolutional neural network part is used to identify spatial texture, patch boundaries, homogeneous clustering, and heterogeneous contact relationships within the neighborhood of a target cell; the long short-term memory network part is used to process multi-temporal feature sequences and identify the persistence, lag, and abruptness of land use status evolution over time.

[0041] Before calling the convolutional long short-term memory network to predict the target year, the system uses the spatiotemporal feature tensor of historical periods as input samples and the land use type label or land use conversion result of the next period as supervision labels to train the weights of the convolutional layer, the weights of the long short-term memory network, and the bias parameters. After training, the system fixes or calls the trained model parameters to calculate the growth tendency probability surface of the target year.

[0042] like Figure 3 As shown, the system inputs the spatiotemporal feature tensor obtained in step S100 into the CNN-LSTM structure. For each time step in the tensor, the system reads the input data of the current time step from the spatiotemporal feature tensor. The input data includes the land use status, driving factor channels, and constraint labels within the target cell's neighborhood window at the current time step. The system first feeds the input data into a convolutional layer and performs spatial convolution calculations.

[0043] In convolutional computation, the convolutional kernel slides within a neighborhood window, weighting the attribute channels of the target pixel and its surrounding pixels. This process can extract local spatial pattern features. For example, if built-up land continuously expands around the target pixel, the convolutional layer can identify the spatial texture of the outward expansion of the built-up land boundary; if woodland or grassland is continuously distributed, the convolutional layer can identify the similar clustering structure of ecological patches; if the pixel is located at the boundary of built-up land, cultivated land, and unused land, the convolutional layer can capture heterogeneous contact relationships.

[0044] In one feasible implementation, the system uses convolutional layers to perform spatial feature extraction, and the calculation formula is as follows:

[0045] in, For spatial features, It is the ReLU activation function. For convolutional layer weights or convolutional kernels, This represents the convolution operation. For time steps Input data, These are the bias parameters for the convolutional layer.

[0046] The convolutional layer weights or convolutional kernels These are parameters that are continuously adjusted during model training to learn the combination patterns of different land use types and driving factors in the spatial neighborhood. The bias parameters... The ReLU activation function is used to provide a base offset to the convolution calculation results, enabling the model to fit more complex spatial nonlinear relationships. It performs a nonlinear transformation on the convolution results, thereby enhancing the spatial feature representation capability.

[0047] Furthermore, the system inputs the spatial features extracted by the convolutional layers into long short-term memory (LSM) network units in chronological order. These LSM units process multi-temporal sequences through a gating mechanism, enabling the model to retain long-term information related to land use evolution while weakening historical information that contributes less to target prediction. Thus, the system can handle temporal features in land use change such as continuous expansion, phased lags, and local abrupt changes.

[0048] In a Long Short-Term Memory (LSTM) network unit, the system first determines the proportion of information that needs to be retained from the previous cell state through a forgetting gate, then determines the extent to which current input information is written into the cell state through an input gate, and finally updates the current cell state. The calculation formula is as follows:

[0049]

[0050]

[0051] in, This is a forget gate, used to control the information retained from the cell's state at the previous moment; This is the input gate, used to control the extent to which the current input information is written into the cell state; This represents the current state of the cell. This represents the cell state at the previous moment; This is the hidden state from the previous moment; This is the current input; This is the concatenation result of the hidden state from the previous moment and the current input; , , These are the weight matrices for the forget gate, input gate, and candidate cell states, respectively. , , These are the bias vectors for the forget gate, input gate, and candidate cell state, respectively. This indicates element-wise multiplication.

[0052] Specifically, if a land use type shows a continuous expansion trend in multiple historical periods, the LSTM unit can retain this continuous information through cell state; if a short-term abnormal change occurs in a certain period, but it is inconsistent with the long-term evolution trend, the gating structure can reduce the impact of this abnormal information on the final output; if a certain driving condition changes in the later period and causes a sudden change in land use state, the input gate can write the change into the current state, so that the model can respond to the sudden change trend.

[0053] After completing the temporal memory computation, the system uses the updated cell state or the hidden representation formed by that cell state as the temporal memory computation result, and sends this result to the Softmax layer to calculate the growth tendency probability of the cell turning to the candidate land type:

[0054] in, For pixels Turning to land category The probability of growth tendency For pixels Land Class The output value, For pixels For the Output values ​​for land use types, The total number of candidate land use types. This is the base of the natural index.

[0055] It should be noted that the growth tendency probability serves as a fundamental parameter for subsequent demand constraint solutions and patch competition allocation. This probability expresses the relative likelihood that a target pixel will convert to a certain land use type under the combined influence of the current driving environment, historical evolution trends, and neighborhood structure.

[0056] Step S300: The growth tendency probability surface is fused with neighborhood effect, spatial constraint and land type inertia to form a comprehensive conversion potential quantification value, and candidate conversion units are determined. The comprehensive conversion potential quantification value is configured as a conversion parameter shared by multi-objective demand solution and patch competition allocation.

[0057] Understandably, the growth tendency probability surface primarily expresses the relative probability of a target pixel transforming into a certain land use type under the influence of historical evolution trends and local spatial structures. To integrate the neural network output into the spatial evolution process, the system further introduces neighborhood effects, spatial constraints, and land use inertia to form a comprehensive conversion potential quantification value that can be used for CARS competitive allocation.

[0058] The neighborhood effect is used to express the proportion of similar land use types around a pixel. If a certain land use type has a high proportion around a target pixel, the target pixel is more likely to form continuous patches after moving to that land use type; if there are few similar pixels around, the conversion is more likely to result in isolated pixels. By introducing the neighborhood effect, the system can tend to form a more complete patch structure in spatial allocation, rather than simply selecting the pixel with the highest probability.

[0059] Spatial constraints are used to express restrictive conditions such as prohibited development zones, ecological protection red lines, and basic farmland. When a target pixel is located in a prohibited transition zone, the spatial constraint factor can suppress or prevent it from undergoing a state transition that does not meet the constraints; when the target pixel is not in a restricted zone, the spatial constraint factor does not block its transition process. Thus, the system can ensure that the simulation process meets the spatial control requirements of ecological protection, arable land protection, and construction restrictions.

[0060] Land use inertia is used to express the competitive intensity of a certain land use type relative to the target demand in the current iteration stage. If the current quantity of a certain land use type is significantly lower than the target demand, its weight in subsequent spatial competition needs to be increased; if a certain land use type is close to the target demand or has experienced excessive expansion, its competitive weight needs to be reduced. By introducing land use inertia, the system can feed back macro-demand differences to the micro-pixel competition process.

[0061] The growth tendency probability surface first enters the comprehensive conversion potential quantification stage during data flow, and then supports the demand solving process and the patch competition allocation process respectively. Thus, it simultaneously connects the shared conversion parameters of macroscopic target structure and microscopic spatial allocation.

[0062] For the calculation of the comprehensive conversion potential quantification value, in the demand solving stage, the system reads the preprocessed growth tendency probability surface, neighborhood effect, spatial constraint factor and preset initial land use inertia coefficient to screen candidate conversion units; in the patch competition allocation stage, the system reads the current global raster state and the updated land use inertia coefficient in each iteration, and then performs the comprehensive conversion potential quantification value calculation for each target pixel.

[0063] In one feasible implementation, the system targets pixels. At the iteration time Shift to candidate land categories The comprehensive conversion potential is calculated using quantitative values:

[0064] in, For pixels At any moment Land Classification The comprehensive conversion potential quantification value, For pixels Land Classification The probability of growth tendency For a moment lower pixel Surrounding land types Neighborhood effect For a moment Lower Earth The coefficient of inertia, For pixels Spatial constraint factor.

[0065] Step S400: Perform multi-objective demand solution on the candidate conversion unit to obtain the target land demand vector.

[0066] It should be noted that the target land demand vector is used to limit the quantitative structure of each land use type in the target year. In this embodiment, the demand solution is based on the comprehensive conversion potential quantification value, identifying high-confidence candidate conversion units for each land use type in space, and then solving the target demand structure in combination with constraints.

[0067] Specifically, the system first reads the corresponding comprehensive conversion potential quantification value layer for each land use type, and then filters pixels whose comprehensive conversion potential quantification values ​​meet the preset information requirements to form candidate conversion units for each land use type. The candidate conversion unit represents a set of pixels with a basis for conversion under the combined effects of the current historical evolution trend, local spatial structure, driving conditions, and limiting conditions.

[0068] Subsequently, the system statistically analyzes the number, spatial distribution, and connectivity of candidate conversion units. The number represents the potential scale of spatial conversion for that land use type; the spatial distribution represents the concentrated and dispersed areas of candidate conversion units within the study area; and the connectivity represents whether candidate conversion units can form continuous patches. Through these statistics, the system can determine the area range within which each land use type has the potential for realization during the target period.

[0069] When solving multi-objective requirements, the system uses the area of ​​each land use type as the upper and lower limits of the corresponding target quantity, and introduces a total area conservation constraint to ensure that the target quantity of all land use types matches the total area of ​​the study area; it introduces farmland protection constraints to ensure that the farmland target quantity meets the farmland red line requirements; it introduces ecological protection constraints to prevent the disorderly compression of ecological land such as forest land, grassland, and water area; it introduces construction land control constraints to ensure that the expansion of construction land is subject to planning constraints; and it introduces carbon storage improvement targets and economic output targets to ensure that the demand results can take into account carbon emissions and regional development.

[0070] The objective function for carbon reserves is as follows:

[0071] in, Let be the objective function for carbon storage. For land use type serial number, The total number of land use types. For the first Aboveground carbon density of land-like soil For the first The underground carbon density of land-like soil For the first Soil carbon density of land-like soils, For the first Carbon density of dead matter in land-like soil For the first The predicted area of ​​land-like species.

[0072] To address the impact of climate differences, the system uses precipitation correction relationships to process soil carbon density or corresponding carbon density parameters:

[0073] in, To correct relevant quantities for precipitation, The system uses precipitation parameters to correct soil carbon density or carbon density parameters for the corresponding land use type, and then incorporates the corrected carbon density parameters into the carbon storage objective function, thereby matching the carbon storage calculation with the climate differences in the study area.

[0074] After obtaining the constraints and objective function, the system uses the non-dominated sorting genetic algorithm NSGA-II or an equivalent multi-objective optimization algorithm to search for the Pareto optimal solution set, and outputs the target land demand vectors under different scenarios from the Pareto optimal solution set. .

[0075] It is understandable that the target land demand vector forms a quantitative structure under the combined influence of candidate transformation units, spatial feasibility, carbon storage targets, ecological protection targets, and economic output targets. Through this processing, the system can avoid the problem of macroscopic area targets being numerically reasonable but spatially unattainable.

[0076] Step S500: Using the target land demand vector to drive patch competition allocation, perform cell state transition and bias feedback update on the global raster to output the land use simulation pattern for the target year.

[0077] It should be noted that the CARS patch competitive allocation process is performed on a global raster base map using an iterative spatial competition. The system inputs the target land demand vector and growth tendency probability surface into the CARS module, and after obtaining the target land demand vector, updates the land use inertia coefficient based on the current allocation deviation, and then calculates the comprehensive conversion potential quantification value of the candidate land use for each target pixel.

[0078] Furthermore, the system employs a random patch seed mechanism and a descending threshold mechanism to trigger the generation of new patches. When the quantified comprehensive conversion potential value of a candidate land type exceeds the descending threshold... When the candidate land type enters the competition set, the candidate land type's comprehensive conversion potential quantification value does not reach the descending threshold. When this threshold is reached, the candidate land category will not participate in the conversion competition for the current pixel. Through this threshold control, the system can reduce the participation of land categories that clearly do not meet the conversion conditions in invalid competition.

[0079] If multiple candidate land use types for the same pixel meet the triggering conditions, the system only performs a roulette wheel selection among the candidate land use types that meet the threshold requirements. Specifically, the system divides the competition probability according to the quantified value of the comprehensive conversion potential of each candidate land use type, generates random numbers, and performs a roulette wheel selection, ultimately updating the land use type of the target pixel to the selected land use type. This method selects candidate land use types based on their relative competition intensity, giving the spatial evolution process the randomness required for patch generation.

[0080] After the cell state is updated, the system writes the updated target cell back to the global raster base map. Cell state transition here refers to a change in the land use type label represented by the cell. For example, if the land use type of the target cell was cultivated land in the previous time step, and after the comprehensive conversion potential quantification value calculation and roulette wheel competition, if construction land is selected, then the land use type label of that cell in the current iteration result will be updated to construction land.

[0081] It should be noted that the relationship between patches and cells is one of globality and locality. A cell is the basic unit in the raster space, and all probability calculations and state updates are performed on the cell level; a patch, on the other hand, is a collection of multiple spatially connected cells with the same land use type. The system does not require additional stitching of evolved patches; instead, the patch structure is naturally formed on the global raster base map through continuous cell updates.

[0082] To make the simulation results closer to the actual land use pattern, the system controls the formation, expansion, and boundary extension of patches during the pixel transformation process through neighborhood effects and connectivity constraints. If a candidate transform pixel is located at the edge of an existing similar patch, its transformation can promote the outward expansion of the patch; if a candidate transform pixel is far from a similar patch and has weak neighborhood support, its comprehensive transformation potential quantification value is suppressed, thereby reducing the probability of isolated pixels.

[0083] After each round of spatial allocation, the system calculates the difference between the currently allocated quantity of each land use type and the target land demand vector, as well as the absolute value of this difference. It then updates the land use type inertia coefficient based on the direction and absolute value changes of the difference. If the allocated quantity of a certain land use type exceeds the target quantity and the excess deviation widens, the inertia coefficient for that land use type is reduced according to the following formula:

[0084] in, For a moment Lower Earth The coefficient of inertia, For the previous iteration, the land class The coefficient of inertia, For the previous iteration, the land class The difference in quantity For the previous iteration, the land class The difference in quantity For land use type serial number, This is the iteration time.

[0085] It is understood that the quantity difference corresponds to the difference between the total number of pixels included in a specific land use type and the target number, which geographically corresponds to the difference in land area. In the above formula... and This represents the absolute value of the excess deviation; when the excess deviation increases... Less than 1, making This reduces the threshold, thereby curbing further excessive expansion of that land category. For land categories whose current quantity is below the target demand, the system enhances their chances of participating in subsequent competition by lowering the expansion threshold or increasing their competition priority.

[0086] Specifically, if the number of allocated pixels for construction land in the current iteration is lower than the target number of construction land in the target land demand vector, the system can lower the expansion threshold of construction land or increase its priority in the competition set, making it easier for it to complete expansion in areas that meet spatial constraints and neighborhood effects; if a certain land type has exceeded the target number and the excess deviation has further increased compared to the previous round, the system reduces the inertia coefficient of that land type according to the inertia coefficient update formula to inhibit its continued expansion; if forest land or grassland is close to its target number, the system reduces its tendency to continue expanding to avoid causing an imbalance in the quantity structure.

[0087] Furthermore, the system continuously executes a loop of calculating the comprehensive conversion potential quantification value and updating the inertia coefficient during the iteration process until the quantity requirement corresponding to the target land demand vector is met, or a preset iteration termination condition is reached. The preset iteration termination condition may be meeting the simulation requirements of the target year, reaching a preset number of iterations, or the quantity difference between each land use type meeting the allowable range.

[0088] After completing all iterations, the system outputs a simulated land use pattern for the target year. This simulated pattern is a land use type layer corresponding to a unified raster base map of the study area, with each cell labeled with the land use type for the target year. This simulated pattern can be used for land spatial planning scenario analysis, ecological protection pattern assessment, construction land expansion analysis, and carbon storage change assessment.

[0089] Through the above steps, this embodiment can solve the technical problems in existing land use simulation technologies, such as insufficient capture of spatial neighborhood structure, weak temporal nonlinearity representation ability, and the disconnect between macroscopic demand solving and microscopic spatial allocation. Figure 4As shown, the original raster data is preprocessed and then enters the spatiotemporal feature tensor. The spatiotemporal feature tensor is calculated by CNN-LSTM to form the growth tendency probability. The growth tendency probability is further involved in the calculation of the comprehensive conversion potential quantification value. The comprehensive conversion potential quantification value serves both the target demand solution and the patch competition allocation, and finally forms the simulated pattern in the feedback output stage.

[0090] The technical effects of this embodiment are as follows: on the one hand, the system improves the accuracy of land use conversion potential expression through spatiotemporal feature tensors and CNN-LSTM structure; on the other hand, the system connects demand solving and spatial allocation through comprehensive conversion potential quantification value, so that the target land demand matches the spatial feasibility; furthermore, the system reduces the probability of isolated pixels, fragmented patches and unreasonable mutations through patch competition allocation and bias feedback update.

[0091] Example 2 Based on the first embodiment of this application, in the second embodiment of this application, the same or similar content as the first embodiment described above can be referred to the above description, and will not be repeated hereafter.

[0092] This application also provides a land use spatiotemporal evolution simulation system, please refer to... Figure 5 The land use spatiotemporal evolution simulation system includes: Module 10 is used to perform unified raster processing on multi-period land use raster data and driving factor raster data of the study area to construct the spatiotemporal feature tensor of the target pixel. Module 20 is invoked to call the convolutional long short-term memory network to perform spatial convolution calculation and temporal memory calculation on the spatiotemporal feature tensor, and output the growth tendency probability surface of each pixel facing the candidate land use type; The determination module 30 is used to fuse the growth tendency probability surface with neighborhood effect, spatial constraint and land type inertia into a comprehensive conversion potential quantification value, and to determine candidate conversion units, wherein the comprehensive conversion potential quantification value is configured as a conversion parameter shared by multi-objective demand solution and patch competition allocation; Solver module 40 is used to perform multi-objective demand solving on candidate conversion units to obtain the target land demand vector; The update module 50 is used to drive patch competition allocation using the target land demand vector, perform cell state transition and bias feedback update on the global raster, and output the land use simulation pattern for the target year.

[0093] The land use spatiotemporal evolution simulation system provided in this application, employing the land use spatiotemporal evolution simulation method described in the above embodiments, can solve the technical problems of insufficient capture of spatial neighborhood structure, weak temporal nonlinearity representation ability, and the disconnect between macroscopic demand solution and microscopic spatial allocation in land use simulation. Compared with the prior art, the beneficial effects of the land use spatiotemporal evolution simulation system provided in this application are the same as those of the land use spatiotemporal evolution simulation method provided in the above embodiments, and will not be elaborated here.

[0094] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the land use spatiotemporal evolution simulation method and system of this application. Equivalent substitutions or conventional transformations based on the technical concept of this application should be included within the protection scope of this application.

[0095] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for simulating the spatiotemporal evolution of land use, characterized in that, The method includes: Unified raster processing was performed on the multi-phase land use raster data and driving factor raster data of the study area to construct the spatiotemporal feature tensor of the target pixels; The convolutional long short-term memory network is invoked to perform spatial convolution and temporal memory calculations on the spatiotemporal feature tensor, outputting the growth tendency probability surface of each pixel facing the candidate land use type; wherein, the execution of spatial convolution and temporal memory calculations specifically includes: reading input data from the spatiotemporal feature tensor step by step, performing spatial convolution calculations on the input data at each time step to obtain spatial features, and inputting the spatial features at each time step into the long short-term memory network in chronological order to perform temporal memory calculations; The growth tendency probability surface is integrated with neighborhood effect, spatial constraint and land type inertia to form a comprehensive conversion potential quantification value, and candidate conversion units are determined. The comprehensive conversion potential quantification value is configured as a conversion parameter shared by macro-level land multi-objective demand solution and micro-level patch competition allocation. The comprehensive conversion potential quantification value is formed by a preset initial land type inertia coefficient. Multi-objective demand solution is performed on the candidate conversion units to obtain the target land demand vector; specifically, this includes: statistically analyzing the number, spatial distribution and connectivity of the candidate conversion units to form the area realization range of each land use type, and using the area realization range as the upper and lower limits of the target number of each land use type to perform multi-objective search. The target land demand vector is used to drive patch competition allocation, and cell state transition and deviation feedback update are performed on the global raster to output the land use simulation pattern for the target year. Specifically, the land use inertia coefficient is updated based on the deviation between the number of currently allocated land use cells and the corresponding target number of land use cells in the target land demand vector, and the updated land use inertia coefficient is used to form a comprehensive conversion potential quantification value.

2. The land use spatiotemporal evolution simulation method as described in claim 1, characterized in that, Unified raster processing was performed on multi-period land use raster data and driving factor raster data of the study area to construct the spatiotemporal feature tensor of the target pixels, specifically including: Obtain at least three phases of land use raster data within the study area. and matching Raster data of driving factors; Resampling is performed on the land use raster data and the driving factor raster data. Z-score normalization is performed on the resampled continuous driving factor raster data. One-hot coding or equivalent discrete coding is performed on the categorical land use status in the land use raster data. A neighborhood window of a preset size is extracted centered on the target pixel, and historical land use status, driving factor values ​​and restrictive conditions are encoded into the same data structure; The encoded time series, attribute channels, and spatial neighborhood are assembled into a spatiotemporal feature tensor: ； in, For the spatiotemporal feature tensor, Represents a multidimensional array of real numbers. For time steps, The number of feature channels, and The spatial dimensions of the neighboring window.

3. The land use spatiotemporal evolution simulation method as described in claim 1, characterized in that, The convolutional long short-term memory network is invoked to perform spatial convolution and temporal memory computations on the spatiotemporal feature tensor, outputting the growth tendency probability surface of each pixel facing the candidate land use type, specifically including: Input data is read from the spatiotemporal feature tensor step by step. Convolutional layers are used to process the input data. Execution spatial feature extraction: ； in, For spatial features, It is the ReLU activation function. For convolutional layer weights or convolutional kernels, This represents the convolution operation. For time steps Input data, These are the bias parameters for the convolutional layer; The spatial features are input into a long short-term memory network unit to perform gated update calculations: ； in, For the Gate of Oblivion For input gate, This represents the current state of the cell. This represents the cell state at the previous moment. This is the hidden state from the previous moment. For the current input, For the feature splicing result, , , These are the weight matrices for the forget gate, input gate, and candidate cell states, respectively. , , These are the bias vectors for the forget gate, input gate, and candidate cell states, respectively. This indicates element-wise multiplication; The results of the temporal memory computation are fed into the Softmax layer, and the pixels are output. For candidate land categories Probability of growth tendency: ； in, For pixels Land Classification The probability of growth tendency For pixels Land Class The output value, For pixels For the Output values ​​for land use types, The total number of candidate land use types. This is the base of the natural index.

4. The land use spatiotemporal evolution simulation method as described in claim 1, characterized in that, The growth tendency probability surface is integrated with neighborhood effects, spatial constraints, and land cover inertia to form a comprehensive conversion potential quantification value, and candidate conversion units are determined, specifically including: For each target pixel At the iteration time Calculate its candidate land use type for relocation Quantitative value of comprehensive conversion potential: ； in, For pixels At any moment Land Classification The comprehensive conversion potential quantification value, For pixels Land Classification The probability of growth tendency For a moment lower pixel Surrounding land types Neighborhood effect For a moment Lower Earth The coefficient of inertia, For pixels Spatial constraint factor; Pixels whose comprehensive conversion potential quantification value meets the preset information requirements are selected to form candidate conversion units for each land use type.

5. The land use spatiotemporal evolution simulation method as described in claim 1, characterized in that, Perform multi-objective demand solving on the candidate transformation units to obtain the target land demand vector, specifically including: The number, spatial distribution, and connectivity of the candidate conversion units are statistically analyzed to determine the area realization range for each land use type. Under the constraints of total area conservation, arable land protection, ecological protection, construction land control, carbon storage enhancement, and economic output, a multi-objective search is performed using the area range as the upper and lower limits of the target quantity for each land use type, and the carbon storage objective function is calculated: ； in, Let be the objective function for carbon storage. For land use type serial number, This represents the total number of land use types. For the first Aboveground carbon density of land-like soil For the first The underground carbon density of land-like soil For the first Soil carbon density of land-like soils, For the first Carbon density of dead matter in land-like soil For the first Predicted area of ​​land class 2; Soil carbon density or its corresponding parameters are corrected using precipitation correction relationships to account for the effects of climate differences. ； in, To correct relevant quantities for precipitation, These are precipitation parameters; Search for the Pareto optimal solution set and output the target land demand vector. .

6. The land use spatiotemporal evolution simulation method as described in claim 5, characterized in that, The target land demand vector is configured as a set of target quantities for each land use type that matches the spatial convertibility capability. The number of land use type targets in the target land demand vector is limited by the area range formed by the candidate conversion units, and is jointly limited by carbon storage targets, ecological protection constraints, arable land protection constraints, construction land control constraints, and economic output constraints.

7. The land use spatiotemporal evolution simulation method as described in claim 1, characterized in that, Using the target land demand vector to drive patch competition allocation, cell state transitions and bias feedback updates are performed on the global raster, specifically including: Input the target land demand vector and the growth tendency probability surface into the CARS module; The comprehensive conversion potential quantification value is higher than the descending order threshold. The candidate land use type is added to the competition set, a roulette wheel competition is performed on the competition set, and the land use type of the target cell is updated to the land use type selected in the competition.

8. The land use spatiotemporal evolution simulation method as described in claim 7, characterized in that, The deviation feedback update adjusts the inertia coefficient based on the land category quantity deviation. The land category quantity deviation includes the difference between the currently allocated number of land category pixels and the target number of the corresponding land category in the target land demand vector, as well as the absolute value of this difference. When the allocated quantity of a certain land category exceeds the target quantity and the excess deviation increases, the inertia coefficient of that land category is reduced according to the following formula: ； in, For a moment Lower Earth The coefficient of inertia, For the previous iteration, the land class The coefficient of inertia, For the previous iteration, the land class The difference in quantity For the previous iteration, the land class The difference in quantity For land use type serial number, This is the iteration time.

9. The land use spatiotemporal evolution simulation method as described in claim 1, characterized in that, The global grid completes state updates pixel by pixel on the same grid base map. Multiple spatially adjacent pixels with the same land use type after state updates constitute a patch.

10. A land use spatiotemporal evolution simulation system, characterized in that, The system includes: The module is used to perform unified raster processing on multi-period land use raster data and driving factor raster data of the study area to construct the spatiotemporal feature tensor of the target pixel. The calling module is used to call the convolutional long short-term memory network to perform spatial convolution calculation and temporal memory calculation on the spatiotemporal feature tensor, and output the growth tendency probability surface of each pixel facing the candidate land use type; wherein, the execution of spatial convolution calculation and temporal memory calculation specifically includes: reading input data from the spatiotemporal feature tensor step by step, performing spatial convolution calculation on the input data at each time step to obtain spatial features, and inputting the spatial features at each time step into the long short-term memory network in chronological order to perform temporal memory calculation; The determination module is used to integrate the growth tendency probability surface with neighborhood effect, spatial constraint and land type inertia to form a comprehensive conversion potential quantification value, and to determine candidate conversion units. The comprehensive conversion potential quantification value is configured as a conversion parameter shared by macro-level land multi-objective demand solution and micro-level patch competition allocation. The comprehensive conversion potential quantification value is formed by a preset initial land type inertia coefficient. The solution module is used to perform multi-objective demand solving on the candidate conversion units to obtain the target land demand vector; specifically, it includes: counting the number, spatial distribution and connectivity of the candidate conversion units to form the area realization range of each land use type, and using the area realization range as the upper and lower limits of the target number of each land use type to perform multi-objective search; The update module is used to drive patch competition allocation using the target land demand vector, perform cell state transition and deviation feedback update on the global raster, and output the land use simulation pattern for the target year. Specifically, it includes: updating the land use inertia coefficient based on the deviation between the number of currently allocated land use cells and the corresponding target number of land use cells in the target land demand vector, and using the updated land use inertia coefficient to form a comprehensive conversion potential quantification value.

Images