A land use type prediction method and system based on multi-source data fusion

Abstract

The application discloses a land use type prediction method and system based on multi-source data fusion, and the method comprises the following steps: acquiring historical land use and multi-source driving factor grid data of a target region; generating a shape self-adaptive conversion probability through a neural network coupled with deformable convolution; determining future land demand based on scenario and policy coupling; iteratively simulating by using a dynamic double feedback cellular automaton model, the model synchronously adjusts the competition inertia coefficient and the spatial neighborhood weight, and dynamically integrates spatial constraints such as ecological protection red lines in a three-dimensional lookup table; finally, the prediction result is converted into standardized land cover data through a probabilistic multi-level classification mapping engine, and a spatial uncertainty distribution map and an area confidence interval are synchronously generated. The method realizes the deep internalization of regulatory rules, provides complete decision support information containing uncertainty quantification, and provides a scientific and reliable technical tool for land space planning and management.

Description

Technical Field

[0001] This invention belongs to the field of land use type prediction technology, and in particular relates to a land use type prediction method and system based on multi-source data fusion. Background Technology

[0002] Land use / cover change is one of the core factors influencing regional climate, ecological environment, and sustainable development. Accurately predicting future land use patterns is of great significance for national spatial planning, ecological protection and restoration, and climate change response.

[0003] Current land use prediction models, such as those based on cellular automata (CA), typically require the integration of multiple driving factors (such as topography, climate, and transportation) to simulate spatial competition and allocation of land types. Traditional CA models often employ fixed transformation rules or simple inertia coefficients during simulation, making it difficult to dynamically respond to the complex feedback between macro-level land demand and micro-level spatial competition. This results in large cumulative errors in long-term simulations and insufficient rationality in spatial allocation.

[0004] On the other hand, existing predictive studies often stop at generating future land use distribution maps, which has the following limitations: First, the prediction results mostly use model-defined or study area-specific classification systems, which cannot directly meet the input requirements of climate models (such as WRF) and hydrological models for standardized land cover classification data (such as IGBP and MODIS classifications), thus limiting the application value of the prediction results. Second, there is a lack of refined quantitative analysis tools for the land use change process in the prediction results, making it difficult to clearly reveal the transformation relationships and competitive trade-offs between different types of land, thus limiting its ability to provide decision support for precise spatial management.

[0005] Therefore, there is an urgent need for a land use prediction technology that can improve the dynamic adaptability of simulations and achieve standardized output and in-depth analysis of prediction results. Summary of the Invention

[0006] This invention provides a land use type prediction method and system based on multi-source data fusion, which addresses the problems of insufficient dynamic adaptability of the simulation process, poor compatibility of prediction results, and weak quantitative analysis of change processes in existing technologies.

[0007] In a first aspect, the present invention provides a land use type prediction method based on multi-source data fusion, comprising:

[0008] Acquire historical land use raster data for the target area, as well as multi-source driving factor raster data including topographic factors, climate factors, and infrastructure distance factors;

[0009] The historical land use raster data and the multi-source driving factor raster data are input into a feature enhancement neural network coupled with a deformable convolutional kernel to generate the morphological adaptive transformation probability of various land use types on each raster unit. The feature enhancement neural network adaptively adjusts the sampling point position through deformable convolutional layers to accurately match the geometric morphology of different land use patches.

[0010] Based on pre-defined future development scenario data, a scenario-policy coupled correction algorithm is used to determine the demand area of ​​various land use types in the predicted target year;

[0011] A dynamic dual-feedback cellular automata model is used for iterative simulation. The morphological adaptive transformation probability, the demand area, and the spatial transformation cost dynamically generated by the rigid constraint layer are used as inputs. After iterating until the convergence condition is met, a predicted map of future land use spatial distribution is output. In each iteration, the first feedback and the second feedback are executed simultaneously. The first feedback is to dynamically adjust the competition inertia coefficient of each land type according to the difference between the current simulated area and the target demand area of ​​each land type according to a preset nonlinear rule. The second feedback is to dynamically fine-tune the neighborhood influence weight of each land type according to the spatial clustering degree of each land type in the simulation results of this round of iteration.

[0012] A probabilistic multi-level classification mapping engine is used to convert the land categories in the future land use spatial distribution prediction map into standardized land cover classification data. The probabilistic multi-level classification mapping engine constructs a probability transition matrix from predicted classification to standard classification based on historical data, and generates an area allocation scheme that conforms to the probability distribution through Monte Carlo simulation.

[0013] Secondly, the present invention provides a land use type prediction system based on multi-source data fusion, comprising:

[0014] The acquisition module is configured to acquire historical land use raster data of the target area, as well as multi-source driving factor raster data including topographic factors, climate factors, and infrastructure distance factors.

[0015] The generation module is configured to input the historical land use raster data and the multi-source driving factor raster data into a feature enhancement neural network coupled with a deformable convolutional kernel to generate the morphological adaptive transformation probability of various land use types on each raster unit. The feature enhancement neural network adaptively adjusts the sampling point position through deformable convolutional layers to accurately match the geometric morphology of different land use patches.

[0016] The module is configured to determine the demand area of ​​various land use types in the predicted target year based on preset future development scenario data and using a scenario-policy coupling correction algorithm.

[0017] The output module is configured to use a dynamic dual-feedback cellular automaton model for iterative simulation. It takes the morphological adaptive transformation probability, the demand area, and the spatial transformation cost dynamically generated by the rigid constraint layer as input. After iterating until the convergence condition is met, it outputs a predicted map of future land use spatial distribution. In each iteration, the first feedback and the second feedback are executed simultaneously. The first feedback is to dynamically adjust the competition inertia coefficient of each land type according to the difference between the current simulated area and the target demand area of ​​each land type according to a preset nonlinear rule. The second feedback is to dynamically fine-tune the neighborhood influence weight of each land type according to the spatial clustering degree of each land type in the simulation results of this round of iteration.

[0018] The conversion module is configured to use a probabilistic multi-level classification mapping engine to convert the land categories in the future land use spatial distribution prediction map into standardized land cover classification data. The probabilistic multi-level classification mapping engine constructs a probability transition matrix from predicted classification to standard classification based on historical data and generates an area allocation scheme that conforms to the probability distribution through Monte Carlo simulation.

[0019] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the land use type prediction method based on multi-source data fusion according to any embodiment of the present invention.

[0020] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the land use type prediction method based on multi-source data fusion according to any embodiment of the present invention.

[0021] This application presents a land use type prediction method and system based on multi-source data fusion. It employs an adaptive neural network to accurately extract land use morphological features and combines this with a dual-feedback cellular automata to simultaneously optimize total area and spatial pattern, effectively improving prediction accuracy and rationality. A three-dimensional constraint lookup table is designed to achieve pixel-level dynamic embedding of control rules such as ecological red lines and basic farmland, ensuring strict compliance of prediction results. A probabilistic multi-level mapping engine is constructed to directly output standardized data that can drive international climate models. Based on spatial visualization of uncertainty and quantification of confidence intervals, it provides complete decision support including credibility assessment. This technical system deeply integrates multidisciplinary methods, achieving an integrated breakthrough from accurate prediction and rule internalization to risk quantification, providing a reliable scientific tool for refined governance of national land space. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a land use type prediction method based on multi-source data fusion, as provided in an embodiment of the present invention;

[0024] Figure 2 This is a structural block diagram of a land use type prediction system based on multi-source data fusion, provided in an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figure 1 The diagram shows a flowchart of a land use type prediction method based on multi-source data fusion according to this application.

[0028] like Figure 1 As shown, the land use type prediction method based on multi-source data fusion specifically includes the following steps:

[0029] Step S101: Obtain historical land use raster data of the target area, as well as multi-source driving factor raster data including topographic factors, climate factors, and infrastructure distance factors.

[0030] In this step, historical land use raster data for the target area at a resolution of 1km for 2015 and 2020 are acquired. Simultaneously, multi-source driving factor raster data for the same period are also acquired, including: a 30m digital elevation model (DEM) and its derived slope and aspect data (resampled to 1km); climate raster data for annual mean temperature and annual precipitation; and Euclidean distance raster data to railways, highways, provincial roads, city centers, and first-order rivers calculated based on OpenStreetMap data. All driving factor data are then standardized in terms of spatial reference, resolution, and extent, and normalized.

[0031] Step S102: Input the historical land use raster data and the multi-source driving factor raster data into a feature enhancement neural network coupled with deformable convolutional kernels to generate the morphological adaptive transformation probability of various land use types on each raster unit. The feature enhancement neural network adaptively adjusts the sampling point position through deformable convolutional layers to accurately match the geometric shape of different land use patches.

[0032] In this step, the multi-source driving factor raster data is preprocessed by normalization. The preprocessed raster data is then used as input for preliminary feature extraction via a feature extraction backbone network. A deformable convolutional module is introduced onto the feature map output from an intermediate feature layer of the backbone network. This module predicts a spatial offset vector for each standard convolution sampling position based on the feature map, using a parallel offset prediction network composed of convolutional layers. Based on the spatial offset vector, the sampling grid of the feature map is warped, and bilinear interpolation is used to calculate the feature values ​​at non-integer coordinate positions, thereby outputting a deformably adapted target feature map. This target feature map is then input into subsequent network layers for processing, and finally, a Softmax classifier outputs the morphological adaptive transformation probability of each raster unit for various land use types.

[0033] In one specific embodiment, training samples are constructed by using the 2020 land use map (one-hot encoded) as the label and the multidimensional raster formed by the aforementioned multi-source driving factors as the feature to construct a training sample set.

[0034] Network Structure and Training: Design a feature extraction backbone network (such as a ResNet variant). A deformable convolutional module is connected in parallel after the feature map F (size H×W×C) output from one of its intermediate layers. This module first predicts the offset field Δ (size H×W×2k, where k is the kernel size; for example, k=9 for a 3×3 convolution) based on the feature map F using a lightweight offset prediction network (composed of several 1×1 convolutional layers). Then, the original regular convolutional kernel sampling grid P (e.g., the 9 center coordinates of a 3×3 grid) is warped according to the offset Δ to obtain a new sampling position P' = P + Δ. Since the coordinates of P' are decimals, bilinear interpolation is used to calculate the feature values.

[0035] Output probability: After deformable convolution and other layer processing, the feature map is processed by global average pooling and fully connected layers, and finally the Softmax layer outputs the probability that each grid cell belongs to each land use type, i.e., the morphological adaptive transformation probability Ps. Because deformable convolution can make its receptive field fit the land use boundary, this Ps is more accurate in the patch edge region.

[0036] Step S103: Based on the preset future development scenario data, the scenario-policy coupling correction algorithm is used to determine the demand area of ​​various land use types in the predicted target year.

[0037] In this step, we obtain global LUH2 future land demand ratio data at a resolution of 0.25° under future development scenarios. Using an area-weighted statistical downscaling method, we process this data to obtain preliminary demand ratios for the study area at a resolution of 1 km. Then, we read the vector data of the overall land use plan for the study area and extract constraint indicators such as arable land area and ecological protection space area. Using these localized indicators as benchmarks, we correct the downscaled ratios and finally calculate the demand area for various types of land, including arable land, forest land, grassland, and construction land, within the study area for the predicted target year.

[0038] Step S104: Iterative simulation is performed using a dynamic dual-feedback cellular automata model. The inputs are the morphological adaptive transformation probability, the required area, and the spatial transformation cost dynamically generated by the rigid constraint layer. After iterating until the convergence condition is met, the predicted map of future land use spatial distribution is output.

[0039] In this step, the first feedback and the second feedback are executed synchronously in each iteration. The first feedback is to dynamically adjust the competition inertia coefficient of each type of land according to the difference between the current simulated area and the target demand area based on a preset nonlinear rule. The second feedback is to dynamically fine-tune the neighborhood influence weight of each type of land based on the spatial clustering degree of each type of land in the simulation results of this iteration.

[0040] Specifically, in the dynamic dual-feedback cellular automaton model, the dynamic adjustment formula for the competitive inertia coefficient of the first feedback is:

[0041]

[0042] in, Let be the competition inertia coefficient of the k-th land class in the t-th iteration. Let be the competitive inertia coefficient of the k-th land class in the (t-1)-th iteration, and η be the inertia adjustment intensity coefficient. For the target required area, Let be the current simulated area after the previous iteration, and tanh() be the hyperbolic tangent function;

[0043] In the second feedback, the neighborhood influence weight of land type k. The dynamic adjustment is based on the spatial Moran index of the current simulation results for land type k. Perform, the expression is:

[0044] ,

[0045] in, The neighborhood influence weights used in the t-th iteration. The preset base weights are ρ, where ρ is the spatial feedback coefficient. The spatial Moran index is calculated based on the results of the previous simulation for the k-th land class. This is the preset spatial aggregation baseline value.

[0046] The steps involved in dynamically generating the spatial transformation cost based on the rigid constraint layer are as follows:

[0047] Integrate the vector layers of the ecological protection red line, permanent basic farmland, and urban development boundary of the target area, and convert the vector layers into raster layers that are spatially aligned with the historical land use raster data;

[0048] Based on the preset land use control rules, a three-dimensional conversion cost lookup table is constructed. Its three dimensions are: land category c before conversion, land category i after conversion, and the constraint zone type z where the raster unit is located. The land use control rules are the rigid control requirements in the land spatial planning policy text, such as the provisions on permanent basic farmland in official documents.

[0049] During the iterative simulation of cellular automata, for any grid cell, the spatial transformation cost corresponding to the current transformation is dynamically obtained by querying the three-dimensional transformation cost lookup table based on the land class c before transformation, the land class i after transformation, and the constraint region type z in which the grid cell is located. .

[0050] The principles for constructing a 3D conversion cost lookup table include:

[0051] For grids located within ecological protection red line areas, if the conversion involves changing ecological land to non-ecological land, then the corresponding... The value is set to a prohibitive high value that approaches 1;

[0052] For grids located within permanent basic farmland protection zones, if the conversion involves changing arable land to non-agricultural land, then the corresponding... The value is set to a prohibitive high value that approaches 1;

[0053] For grid cells located within the urban development boundary, if the conversion conforms to the planning direction, then the corresponding... The value is set to the first preset value, and otherwise set to the second preset value, wherein the second preset value is greater than the second preset value;

[0054] For grid cells located within the urban development boundary, if the conversion conforms to the planning direction, then the corresponding... The value is set to a first preset value, and otherwise set to a second preset value, wherein the first preset value is greater than the second preset value;

[0055] For grids not located in any rigid constraint zones, the conversion cost is set based on the difficulty of the basic land use conversion. Rigid constraint zones include ecological protection red line zones, permanent basic farmland protection zones, and urban development boundaries.

[0056] Step S105: Using a probabilistic multi-level classification mapping engine, the land categories in the future land use spatial distribution prediction map are converted into standardized land cover classification data.

[0057] In this step, a three-level mapping relationship table is constructed. The three-level mapping relationship table defines the category correspondence between the first-level classification system, the second-level classification system and the third-level classification system. The first-level classification system is the classification system output by the dynamic dual-feedback cellular automata model, the second-level classification system is the standard classification system used to drive the climate or hydrological model, and the third-level classification system is the detailed classification system used in the local historical land survey.

[0058] Based on the three-level mapping table and historical data, the following steps are performed: According to the correspondence between the first-level classification system and the third-level classification system, obtain all third-level category patches historically belonging to the same first-level category; then, according to the correspondence between the third-level classification system and the second-level classification system, calculate the area distribution of third-level category patches on each subcategory in the second-level classification system; calculate the percentage of the total area of ​​each second-level subcategory to the total area of ​​the first-level category, as the historical conversion ratio of the first-level category to each subcategory in the second-level classification system.

[0059] For each first-level category of land use type in the future land use spatial distribution prediction map, obtain the total land area of ​​the first-level category in the future land use spatial distribution prediction map.

[0060] The total land area of ​​the first-level classification is allocated to the corresponding subcategories of the second-level classification system according to the historical conversion ratio.

[0061] The allocated future land use data that conforms to the category definition and area composition of the second-level classification system will be output as the standardized land use classification data.

[0062] In summary, the method presented in this application significantly improves the overall efficiency of land use prediction through a series of technological innovations, including deformable convolution feature enhancement, dual-feedback spatial simulation, three-dimensional constraint dynamic coupling, and probabilistic uncertainty quantification. The method employs an adaptive neural network to accurately extract land cover morphological features and combines this with a dual-feedback cellular automata to simultaneously optimize total area and spatial pattern, effectively improving prediction accuracy and rationality. A three-dimensional constraint lookup table is designed to achieve pixel-level dynamic embedding of control rules such as ecological red lines and basic farmland, ensuring strict compliance of prediction results. A probabilistic multi-level mapping engine is constructed to directly output standardized data that can drive international climate models. Furthermore, it pioneers uncertainty spatial visualization and confidence interval quantification, providing comprehensive decision support including credibility assessment. This technological system deeply integrates multidisciplinary methods, achieving an integrated breakthrough from accurate prediction and rule internalization to risk quantification, providing a reliable scientific tool for refined governance of national land space.

[0063] Please see Figure 2 The diagram shows a structural block diagram of a land use type prediction system based on multi-source data fusion according to this application.

[0064] like Figure 1 As shown, the land use type prediction system 200 includes an acquisition module 210, a generation module 220, a determination module 230, an output module 240, and a conversion module 250.

[0065] The acquisition module 210 is configured to acquire historical land use raster data of the target area, as well as multi-source driving factor raster data including topographic factors, climate factors, and infrastructure distance factors; the generation module 220 is configured to input the historical land use raster data and the multi-source driving factor raster data into a feature enhancement neural network coupled with deformable convolutional kernels to generate morphological adaptive transformation probabilities for various land use types on each raster cell, wherein the feature enhancement neural network adaptively adjusts the sampling point positions through deformable convolutional layers to accurately match the geometric morphology of different land use patches; the determination module 230 is configured to determine the demand area of ​​various land use types for the predicted target year based on preset future development scenario data and using a scenario-policy coupling correction algorithm; the output module 240 is configured to perform iterative simulation using a dynamic dual-feedback cellular automata model to obtain the morphological adaptive transformation probabilities. The system takes the land use rate, required area, and spatial conversion cost dynamically generated by the rigid constraint layer as inputs. After iterating until the convergence condition is met, it outputs a predicted map of future land use spatial distribution. In each iteration, a first feedback and a second feedback are executed simultaneously. The first feedback dynamically adjusts the competition inertia coefficient of each land type according to a preset nonlinear rule based on the difference between the current simulated area and the target required area. The second feedback dynamically fine-tunes the neighborhood influence weight of each land type based on the spatial clustering degree of each land type in the simulation results of this iteration. The conversion module 250 is configured to use a probabilistic multi-level classification mapping engine to convert the land categories in the predicted map of future land use spatial distribution into standardized land cover classification data. The probabilistic multi-level classification mapping engine constructs a probability transition matrix from predicted classification to standard classification based on historical data and generates an area allocation scheme that conforms to the probability distribution through Monte Carlo simulation.

[0066] It should be understood that Figure 2 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 2 The various modules in the document will not be described in detail here.

[0067] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the land use type prediction method based on multi-source data fusion in any of the above method embodiments.

[0068] In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows:

[0069] Acquire historical land use raster data for the target area, as well as multi-source driving factor raster data including topographic factors, climate factors, and infrastructure distance factors;

[0070] The historical land use raster data and the multi-source driving factor raster data are input into a feature enhancement neural network coupled with a deformable convolutional kernel to generate the morphological adaptive transformation probability of various land use types on each raster unit. The feature enhancement neural network adaptively adjusts the sampling point position through deformable convolutional layers to accurately match the geometric morphology of different land use patches.

[0071] Based on pre-defined future development scenario data, a scenario-policy coupled correction algorithm is used to determine the demand area of ​​various land use types in the predicted target year;

[0072] A dynamic dual-feedback cellular automata model is used for iterative simulation. The morphological adaptive transformation probability, the demand area, and the spatial transformation cost dynamically generated by the rigid constraint layer are used as inputs. After iterating until the convergence condition is met, a predicted map of future land use spatial distribution is output. In each iteration, the first feedback and the second feedback are executed simultaneously. The first feedback is to dynamically adjust the competition inertia coefficient of each land type according to the difference between the current simulated area and the target demand area of ​​each land type according to a preset nonlinear rule. The second feedback is to dynamically fine-tune the neighborhood influence weight of each land type according to the spatial clustering degree of each land type in the simulation results of this round of iteration.

[0073] A probabilistic multi-level classification mapping engine is used to convert the land categories in the future land use spatial distribution prediction map into standardized land cover classification data. The probabilistic multi-level classification mapping engine constructs a probability transition matrix from predicted classification to standard classification based on historical data, and generates an area allocation scheme that conforms to the probability distribution through Monte Carlo simulation.

[0074] Computer-readable storage media may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application program required for at least one function; the data storage area may store data created based on the use of the land use type prediction system based on multi-source data fusion. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely configured relative to a processor, which can be connected to the land use type prediction system based on multi-source data fusion via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0075] Figure 3This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 3 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 3 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the land use type prediction method based on multi-source data fusion as described in the above embodiment. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the land use type prediction system based on multi-source data fusion. The output device 340 may include a display screen or other display device.

[0076] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0077] In one implementation, the above-described electronic device is applied to a land use type prediction system based on multi-source data fusion, and is used as a client. It includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to:

[0078] Acquire historical land use raster data for the target area, as well as multi-source driving factor raster data including topographic factors, climate factors, and infrastructure distance factors;

[0079] The historical land use raster data and the multi-source driving factor raster data are input into a feature enhancement neural network coupled with a deformable convolutional kernel to generate the morphological adaptive transformation probability of various land use types on each raster unit. The feature enhancement neural network adaptively adjusts the sampling point position through deformable convolutional layers to accurately match the geometric morphology of different land use patches.

[0080] Based on pre-defined future development scenario data, a scenario-policy coupled correction algorithm is used to determine the demand area of ​​various land use types in the predicted target year;

[0081] A dynamic dual-feedback cellular automata model is used for iterative simulation. The morphological adaptive transformation probability, the demand area, and the spatial transformation cost dynamically generated by the rigid constraint layer are used as inputs. After iterating until the convergence condition is met, a predicted map of future land use spatial distribution is output. In each iteration, the first feedback and the second feedback are executed simultaneously. The first feedback is to dynamically adjust the competition inertia coefficient of each land type according to the difference between the current simulated area and the target demand area of ​​each land type according to a preset nonlinear rule. The second feedback is to dynamically fine-tune the neighborhood influence weight of each land type according to the spatial clustering degree of each land type in the simulation results of this round of iteration.

[0082] A probabilistic multi-level classification mapping engine is used to convert the land categories in the future land use spatial distribution prediction map into standardized land cover classification data. The probabilistic multi-level classification mapping engine constructs a probability transition matrix from predicted classification to standard classification based on historical data, and generates an area allocation scheme that conforms to the probability distribution through Monte Carlo simulation.

[0083] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A land use type prediction method based on multi-source data fusion, characterized in that, The method comprises the following steps: acquiring historical land use grid data of a target area and multi-source driving factor grid data including terrain factors, climate factors and infrastructure distance factors; inputting the historical land use grid data and the multi-source driving factor grid data into a feature enhancement neural network coupled with a deformable convolution kernel to generate shape adaptive conversion probabilities of each type of land use on each grid cell, wherein the feature enhancement neural network adaptively adjusts the sampling point positions through a deformable convolution layer to accurately match the geometric shapes of different land class patches; based on preset future development scenario data, using a scenario-policy coupled correction algorithm to determine the demand area of each type of land use in the predicted target year; using a dynamic double feedback cellular automata model for iterative simulation, taking the shape adaptive conversion probabilities, the demand area and the spatial conversion cost dynamically generated by the rigid constraint layer as inputs, iterating until the convergence condition is met, and outputting a future land use spatial distribution prediction map, wherein in each iteration, a first feedback and a second feedback are simultaneously performed, the first feedback is to dynamically adjust the competition inertia coefficient of each type of land according to the difference between the current simulation area and the target demand area of each type of land according to a preset nonlinear rule, and the second feedback is to dynamically fine-tune the neighborhood influence weight of each type of land according to the spatial aggregation degree of each type of land in the current iteration simulation result; using a probabilistic multi-level classification mapping engine to convert the land categories in the future land use spatial distribution prediction map into standardized land cover classification data, wherein the probabilistic multi-level classification mapping engine constructs a probability transition matrix from the prediction classification to the standard classification based on historical data, and generates an area allocation scheme conforming to the probability distribution through Monte Carlo simulation. 2.The land use type prediction method based on multi-source data fusion according to claim 1, characterized in that, The step of inputting the historical land use grid data and the multi-source driving factor grid data into a feature enhancement neural network coupled with a deformable convolution kernel to generate shape adaptive conversion probabilities of each type of land use on each grid cell comprises the following steps: normalizing and preprocessing the multi-source driving factor grid data; taking the preprocessed multi-source driving factor grid data as input and performing preliminary feature extraction through a feature extraction backbone network; introducing a deformable convolution module on the feature map output by an intermediate feature layer of the feature extraction backbone network, wherein the deformable convolution module predicts a spatial offset vector for each standard convolution sampling position according to the feature map through an offset prediction network in parallel composed of convolution layers; distorting the sampling grid of the feature map according to the spatial offset vector, and calculating the feature values of non-integer coordinate positions using bilinear interpolation to output a target feature map that has been deformed and adapted; inputting the target feature map into subsequent network layers for processing, and finally outputting the shape adaptive conversion probabilities of each grid cell for each type of land use through a Softmax classifier. 3.The land use type prediction method based on multi-source data fusion according to claim 1, characterized in that, In the dynamic double feedback cellular automata model, the dynamic adjustment formula of the competition inertia coefficient of the first feedback is: wherein, is the competition inertia coefficient of the kth type of land at the tth iteration, is the competition inertia coefficient of the kth type of land at the t-1th iteration, and η is an inertia adjustment intensity coefficient, is the target demand area, is the current simulation area after the end of the last iteration, and tanh() is a hyperbolic tangent function. In the second feedback, the neighborhood influence weight of the kth land type is dynamically adjusted based on the spatial Moran's index of the current simulation result of the kth land type is performed, and the expression is: ， wherein, the neighborhood influence weight used for the tth iteration, is a preset base weight, and p is a spatial feedback coefficient, is the spatial Moran's index of the kth type of land calculated based on the simulation results of the previous round, is a preset spatial aggregation degree reference value.

4. The land use type prediction method based on multi-source data fusion according to claim 1, characterized in that, wherein, the step of dynamically generating the spatial conversion cost based on the rigid constraint layer is: integrate vector layers of ecological protection red line, permanent basic farmland, and urban development boundary of a target region, and convert the vector layers into raster layers that are aligned with the historical land use raster data in space; construct a three-dimensional conversion cost lookup table according to a preset land use control rule, three dimensions of the three-dimensional conversion cost lookup table being: land class c before conversion, land class i after conversion, and constraint region type z in which a grid cell is located; In the iteration simulation of the cellular automaton, for any grid cell, according to the land class c before conversion, the land class i after conversion and the constraint region type z where the grid cell is located, the three-dimensional conversion cost lookup table is inquired to dynamically obtain the spatial conversion cost corresponding to the current conversion .

5. The land use type prediction method based on multi-source data fusion according to claim 4, characterized in that, the construction principle of the three-dimensional conversion cost lookup table includes: For the grid located in the ecological protection red line area, if the conversion involves converting ecological land to non-ecological land, the corresponding value is set to a high value of prohibition close to 1; For a cell located within a permanent basic farmland protection zone, if the conversion involves converting cultivated land into non-agricultural land, the corresponding value is set to a high value of prohibition approaching 1; For the grid located within the town development boundary, if the conversion conforms to the planning direction, the corresponding value is set to a first preset value, and otherwise is set to a second preset value, wherein the first preset value is greater than the second preset value; for a grid not located in any rigid constraint region, the conversion cost of the grid is set according to a basic land class conversion difficulty, the rigid constraint region including an ecological protection red line region, a permanent basic farmland protection region, and a region within an urban development boundary.

6. The land use type prediction method based on multi-source data fusion according to claim 1, characterized in that, the conversion of the land class in the future land use spatial distribution prediction map into the standardized land cover classification data by using the probabilistic multi-level classification mapping engine includes: construct a three-level mapping relationship table, the three-level mapping relationship table defining a class correspondence relationship between a first-level classification system, a second-level classification system, and a third-level classification system, wherein the first-level classification system is a classification system output by the dynamic double-feedback cellular automaton model, the second-level classification system is a standard classification system used to drive a climate or hydrological model, and the third-level classification system is a detailed classification system used in local historical land investigation; based on the three-level mapping relationship table and historical period data, the following steps are performed: according to the correspondence relationship between the first-level classification system and the third-level classification system, all third-level class patches belonging to the same first-level class in history are obtained; and according to the correspondence relationship between the third-level classification system and the second-level classification system, the area distribution of the third-level class patches in each sub-class of the second-level classification system is counted; the percentage of the total area of each second-level sub-class to the total area of the first-level class is calculated as the historical conversion proportion of the first-level class to each sub-class of the second-level classification system; for each first-level land use type in the future land use spatial distribution prediction map, the total first-level classification land area thereof in the future land use spatial distribution prediction map is obtained; the total first-level classification land area is distributed to each corresponding sub-class of the second-level classification system according to the historical conversion proportion; the future land use data obtained by the distribution and meeting the class definition and area composition of the second-level classification system is output as the standardized land use classification data.

7. A land use type prediction system based on multi-source data fusion, characterized in that, includes: an acquisition module configured to acquire historical land use raster data of a target region, and multi-source driving factor raster data including a terrain factor, a climate factor, and an infrastructure distance factor; a generation module configured to input the historical land use raster data and the multi-source driving factor raster data into a feature enhancement neural network coupled with a deformable convolution kernel to generate a shape-adaptive conversion probability of each land use type on each grid cell, wherein the feature enhancement neural network adaptively adjusts a sampling point position through a deformable convolution layer to accurately match the geometric shape of different land class patches; The determining module is configured to determine the demand area of each type of land use in the target year based on the preset future development scenario data and using a scenario-policy coupling correction algorithm. The output module is configured to perform iterative simulation using a dynamic double feedback cellular automaton model, take the form adaptive conversion probability, demand area, and spatial conversion cost dynamically generated by the rigid constraint layer as input, and output a future land use spatial distribution prediction map after iteration until the convergence condition is met, wherein the first feedback and the second feedback are executed simultaneously in each iteration, the first feedback is to dynamically adjust the competition inertia coefficient of each type of land according to the difference between the current simulation area and the target demand area of each type of land and according to a preset nonlinear rule, and the second feedback is to dynamically fine-tune the neighborhood influence weight of each type of land according to the spatial aggregation degree of each type of land in the current iteration simulation result. The conversion module is configured to convert the land categories in the future land use spatial distribution prediction map into standardized land cover classification data using a probabilistic multi-level classification mapping engine, wherein the probabilistic multi-level classification mapping engine constructs a probability transition matrix from the prediction classification to the standard classification based on historical data, and generates an area allocation scheme conforming to the probability distribution through Monte Carlo simulation.

8. An electronic device, comprising: Comprise: At least one processor, and a memory connected in communication with the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1 to 6.

Images