Land utilization simulation method and system based on convolutional neural network

Abstract

The invention discloses a land utilization simulation method based on a convolutional neural network, and relates to the technical field of land utilization change simulation, and the method comprises the steps: collecting land utilization data, social economy and natural geography driving factors of a research region, carrying out the data preprocessing, training the land utilization data through employing an improved LeNet-5 convolutional neural network, and carrying out the simulation of the land utilization change. The method comprises the following steps: estimating growth probabilities of different land utilization types, quantifying a neighborhood effect through a Gaussian function, calibrating parameters by using a multi-objective optimization algorithm, describing a city spreading trend through the Gaussian function, capturing a macroscopic city expansion process, and comprehensively considering the growth probabilities, the neighborhood effect and the city spreading trend. According to the land utilization change simulation method based on the convolutional neural network and the multi-objective optimization, the precision and the reliability of land utilization change simulation are remarkably improved by combining microscopic and macroscopic spatial features.

Description

Technical Field

[0001] The present invention relates to the technical field of land use change simulation, and in particular to a land use simulation method and system based on convolutional neural network. Background Art

[0002] Taking Wuhan, China as an example, the study aims to reveal the differences in simulation accuracy and quality factor between the traditional PLUS model of land use change model and the SC-CNN-CA model proposed in this patent, so as to better predict the dynamic evolution trend of land use and predict the spatial explicit information of land use change. Therefore, it has attracted widespread attention from scholars at home and abroad.

[0003] Earlier studies combined top-down models, such as Markov chains and system dynamics models, to quantify macro-demand as the number of units of each land use type at the end of the study period. At the same time, combined with bottom-up CA models, the location of land use transition was determined. However, although macro-demand guides the trajectory of micro-allocation in iterative modeling, the interaction and feedback loop between the two subcomponents have been neglected. This oversight leads to a discrepancy between the evolution of land use in the simulation and reality, which increases the uncertainty of long-term land use change simulation. The coupled top-down and bottom-up urban land dynamic simulation model S-CMM-CA model is an exception. This model uses an inverse S-shaped function to derive urban land demand within each concentric ring of the study area to guide land use transition. Although the model changes the combined process between macro-demand and micro-allocation, the characteristics of urban sprawl are still constrained to a static target rather than a dynamic trend with time dependence, which poses a challenge to capturing its true spatiotemporal characteristics. In summary, there is an urgent need to incorporate the concept of time step into urban sprawl measurement. In the process of constructing the CA model, it is imperative to take into account the regularity of urban sprawl (macro) and the neighborhood effect (micro). On the one hand, it is effective and intuitive to describe various fluctuation trends. On the other hand, Yang Bohan's comparative analysis of urban sprawl patterns in 27 representative cities in China shows that the Gaussian function accurately depicts the trend of urban land sprawl in a specific period.

[0004] With the breakthrough of geographical measurement concepts and the introduction of intelligent algorithms, the research on land use change simulation has broken through the traditional qualitative method dominated by practical experience, and gradually moved towards the quantification of high efficiency of simulation process and high precision of simulation results. The introduction of intelligent algorithms provides a more efficient means for land use change simulation research. In recent years, the research on driving mechanisms based on intelligent algorithms has continued to increase. Summary of the invention

[0005] In view of the above problems existing in the prior art, the present invention is proposed.

[0006] Therefore, the prior art problem to be solved by the present invention is

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: a land use simulation method based on convolutional neural network, which includes collecting land use data, socio-economic and natural geographical driving factors of the study area, and performing data preprocessing;

[0008] The improved LeNet-5 convolutional neural network is used to train land use data to estimate the growth probability of different land use types;

[0009] The neighborhood effect is quantified by Gaussian function, and the parameters are calibrated using a multi-objective optimization algorithm;

[0010] The urban sprawl trend is described by Gaussian function, capturing the macro-urban expansion process;

[0011] The final land use type is determined through roulette wheel selection, taking into account growth probability, neighborhood effect and urban sprawl tendency.

[0012] As a preferred solution of the land use simulation method based on convolutional neural network described in the present invention, wherein: the collecting of land use data of the study area includes collecting land use data using a 30m resolution annual land cover dataset of China;

[0013] The data preprocessing includes processing the driving factor data set, the initial land use map and the land use expansion map.

[0014] As a preferred solution of the land use simulation method based on convolutional neural network described in the present invention, the method of using the improved LeNet-5 convolutional neural network to train the land use data and estimate the growth probability of different land use types includes:

[0015] A 21×21 sampling window is constructed at the input layer, which includes 14 driving factors and the initial land use map;

[0016] Extract image features through multiple convolutional layers;

[0017] The pooling layer is used to select the data features extracted by the convolution layer, and the function value is calculated for the data features in the pooling area. The formula is as follows:

[0018]

[0019] In the formula is the i-th dimension input variable of the l-th pooling layer, and They are the multiplication and addition bias terms of the j-th dimension variable in the l-th pooling layer; is the output variable of the j-th dimension variable in the l-th pooling layer; down is the pooling function;

[0020] The connection layer connects all the output variables of the pooling layer;

[0021] Each vector in the output layer represents the probability of growth of a land-use type at the corresponding cell.

[0022] As a preferred solution of the land use simulation method based on convolutional neural network described in the present invention, the Gaussian function is used to quantify the attraction and neighborhood effect between land use types, and the neighborhood effect is defined as:

[0023]

[0024] in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that the initial land use type of cell i is transferred to k; V(i) is the set of all cells in the Moore neighborhood with i as the central cell; K(i') is the current land use status of any cell i' in the neighborhood; d(i,i') is the distance between cells i and i'; W d(i,i′),K(i′,t) represents the probability that cell i is transformed into land use k by the influence of land use K(i') on cell i' with a distance d(i,i') in the neighborhood space, and t represents the transfer time;

[0025] The Gaussian function quantifies the neighborhood effect by dividing the neighborhood effect into an inertia point and a relationship curve. The inertia point represents the inherent characteristics of the land use category and the strength of the pixel that maintains the original land use type. The relationship curve represents the force affected by the spatial distribution of other land use types in the neighborhood environment, which is defined as:

[0026] Where y(x) represents the influence of any cell with a different land use type from the central cell on the central cell; a, b, c are all control variables of the Gaussian function; when 0<x≤x c When x>x, there is an attractive force between the center cell and the target cell, and the Gaussian function shows the neighborhood interaction related to the distance. c When the neighborhood range x is exceeded c The influence of the cell on the central cell is zero, and the final neighborhood effect is defined as:

[0027]

[0028] in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that cell i is transferred to the target land use k. When the target land use k is the land use type K(i,t) of the current cell i, the probability that cell i maintains its original land use type is d, which is only related to the land use type. When the target land use k is inconsistent with the land use type of the current cell i, the probability that cell i is transferred to the target land use k is the sum of the forces exerted on the central cell i by all cells i' with land use type k in the neighborhood space, where V(i) is the set of all cells in the Moore neighborhood with i as the center unit; a, b, c are related to the initial land type K(i,t) and the target land type k; d(i,i') is the distance between cells i and i', and d(i,i') is greater than 0 and less than the radius of the Moore neighborhood; t represents the transfer time.

[0029] The Gaussian function quantifies the neighborhood effect by dividing the neighborhood effect into an inertia point and a relationship curve. The inertia point represents the inherent characteristics of the land use category and the strength of the pixel that maintains the original land use type. The relationship curve represents the force affected by the spatial distribution of other land use types in the neighborhood environment, which is defined as:

[0030]

[0031] Where y(x) represents the influence of any cell with a different land use type from the central cell on the central cell; a, b, c are all control variables of the Gaussian function; when 0<x≤x c When x>x, there is an attractive force between the center cell and the target cell, and the Gaussian function shows the neighborhood interaction related to the distance. c When the neighborhood range x is exceeded c The influence of the cell on the central cell is zero, and the final neighborhood effect is defined as:

[0032]

[0033] in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that cell i is transferred to the target land use k. When the target land use k is the land use type K(i,t) of the current cell i, the probability that cell i maintains its original land use type is d, which is only related to the land use type. When the target land use k is inconsistent with the land use type of the current cell i, the probability that cell i is transferred to the target land use k is the sum of the forces exerted on the central cell i by all cells i' with land use type k in the neighborhood space. Among them, V(i) is the set of all cells in the Moore neighborhood with i as the central cell; a, b, c are related to the initial land type K(i,t) and the target land type k; d(i,i') is the distance between cells i and i', and d(i,i') is greater than 0 and less than the radius of the Moore neighborhood; t represents the transfer time.

[0034] As a preferred solution of the land use simulation method based on convolutional neural network described in the present invention, the urban sprawl trend is described by a Gaussian function, and its formula is:

[0035]

[0036] is the probability of urban sprawl of cell i in the concentric circles, r represents the radius of the concentric circles from the city center, a represents the height of the peak, b represents the peak of the peak, describing the hotspot of urban sprawl, c represents the width of the function curve, which is approximately the spatial compactness of urban sprawl, and d·t describes the variation of the Gaussian function along the horizontal axis under the assumption that the macroscopic probability of urban sprawl moves uniformly toward the periphery of the city over time.

[0037] As a preferred embodiment of the land use simulation method based on convolutional neural network described in the present invention, the final land use allocation is determined by roulette wheel selection by comprehensively considering growth probability, neighborhood effect and urban sprawl trend, and the calculation formula is:

[0038]

[0039] in represents the overall conversion probability of each land use type, describes the expansion probability of k on cell i, represents the neighborhood effect of land use type k, represents the probability of urban sprawl, γ represents the random perturbation mechanism based on roulette, and t represents the number of iterations.

[0040] As a preferred solution of the land use simulation method based on convolutional neural network described in the present invention, wherein: a data acquisition module is used to collect land use data and driving factors;

[0041] Data processing module, used to preprocess and normalize data;

[0042] Model training module for growth probability estimation of land use types using convolutional neural networks;

[0043] Neighborhood effect calibration module, used to calibrate neighborhood effect parameters through Gaussian function and multi-objective optimization algorithm;

[0044] Urban sprawl trend simulation module, used to describe urban sprawl trends through Gaussian functions;

[0045] The land use assignment module is used to determine the final land use type through a roulette wheel selection.

[0046] As a preferred solution of the land use simulation method based on convolutional neural network described in the present invention, the data acquisition module includes a submodule for collecting China's annual land cover data set with a resolution of 30m and a submodule for collecting 14 driving factor data.

[0047] As a preferred solution of the land use simulation method based on convolutional neural network described in the present invention, the model training module adopts an improved LeNet-5 convolutional neural network model, and the input layer constructs a 21×21 sampling window.

[0048] As a preferred solution of the land use simulation method based on convolutional neural network described in the present invention, the land use allocation module determines the final land use type through a roulette wheel selection algorithm by comprehensively considering the growth probability, neighborhood effect and urban sprawl trend.

[0049] Beneficial effects of the present invention: The land use change simulation method based on convolutional neural network and multi-objective optimization of the present invention significantly improves the accuracy and reliability of land use change simulation by combining micro and macro spatial features. First, the growth probability of different land use types can be estimated quickly and efficiently through the improved LeNet-5 convolutional neural network model. Secondly, the neighborhood effect is quantified by using Gaussian function, and the parameters are calibrated by multi-objective optimization algorithm, which improves the accuracy of the model in the simulation process. The simulation of urban sprawl trend is parameterized by Gaussian function, which can effectively capture the macro process of urban expansion. Finally, by comprehensively considering the growth probability, neighborhood effect and urban sprawl trend, the allocation of land use types is determined by roulette selection algorithm, which significantly improves the accuracy of simulation results. The overall accuracy (OA) and quality factor (FOM) of the model reach 84.12% and 20.20% respectively, which are significantly improved compared with the traditional model. In addition, the simulation effect of this method in agricultural land, water area and urban land is particularly outstanding, showing its superior performance in practical applications. By accurately simulating urban sprawl and land use changes, the present invention provides more scientific decision-making support for urban planning and land management, and has broad application prospects and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0051] Figure 1 This is a framework diagram of the SC-CNN-CA model with multiple dynamic land use categories in the data and method examples of the present invention.

[0052] Figure 2 This is the location map of the study area in the data and method examples of the present invention and the land use map in 2005 (upper right).

[0053] Figure 3 It is the city center and concentric rings of the test area in 2005 and 2015 in the data and method examples of the present invention.

[0054] Figure 4 Illustration of the CNN architecture in the data and method examples of the present invention.

[0055] Figure 5 In the data and method examples of the present invention, a Gaussian function is used to parameterize the neighborhood effect.

[0056] Figure 6 This is a vectorization process diagram of the SC-CNN-CA model in the result example of the present invention.

[0057] Figure 7 It is the ROC curve and AUC value diagram of a single land use type based on LeNet network fitting in the result example of the present invention.

[0058] Figure 8 1 is a Pareto front diagram of NSGA-III after self-calibration in an embodiment of the present invention.

[0059] Fig. 9 It is a violin plot of the inertia point of each land use category in the embodiment of the present invention.

[0060] Fig.10 1 and 2 are different neighborhood effects sorted by land use conversion type in an embodiment of the present invention: (a) conversion parameters between agricultural land and forest land; (b) conversion parameters between agricultural land and water land; (c) conversion parameters between agricultural land and urban land; (d) conversion parameters between grassland and urban land; (e) conversion parameters between water land and agricultural land; (f) conversion parameters between water land and urban land; (g) conversion parameters between urban land and water land; (h) conversion parameters between unused land and urban land.

[0061] Fig.11 This is a diagram of urban expansion trends based on a Gaussian function in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0063] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0064] Example 1

[0065] Reference Figure 1 , Figure 2 , which is the first embodiment of the present invention, and provides a land use simulation method based on a convolutional neural network, comprising:

[0066] Collect land use data, socioeconomic and physical geographic drivers of the study area and perform data preprocessing;

[0067] The improved LeNet-5 convolutional neural network is used to train land use data to estimate the growth probability of different land use types;

[0068] The neighborhood effect is quantified by Gaussian function, and the parameters are calibrated using a multi-objective optimization algorithm;

[0069] The urban sprawl trend is described by Gaussian function, capturing the macro-urban expansion process;

[0070] The final land use type is determined through roulette wheel selection, taking into account growth probability, neighborhood effect and urban sprawl tendency.

[0071] The collection of land use data in the study area includes collecting land use data using the annual land cover dataset of China with a resolution of 30 m;

[0072] Data preprocessing includes processing the driving factor dataset, initial land use map and land use expansion map.

[0073] The improved LeNet-5 convolutional neural network is used to train the land use data to estimate the growth probability of different land use types including:

[0074] A 21×21 sampling window is constructed at the input layer, which includes 14 driving factors and the initial land use map;

[0075] Extract image features through multiple convolutional layers;

[0076] The pooling layer is used to select the data features extracted by the convolution layer, and the function value is calculated for the data features in the pooling area. The formula is as follows:

[0077]

[0078] In the formula is the i-th dimension input variable of the l-th pooling layer, and They are the multiplication and addition bias terms of the j-th dimension variable in the l-th pooling layer; The output variable of the j-th dimension variable in the l-th pooling layer; down is the pooling function;

[0079] The connection layer connects all the output variables of the pooling layer;

[0080] Each vector in the output layer represents the probability of growth of a land-use type at the corresponding cell.

[0081] Gaussian functions are used to quantify the attraction and neighborhood effects between land use types. The neighborhood effect is defined as:

[0082]

[0083] in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that the initial land use type of cell i is transferred to k; V(i) is the set of all cells in the Moore neighborhood with i as the central cell; K(i') is the current land use status of any cell i' in the neighborhood; d(i,i') is the distance between cells i and i'; W d(i,i′),K(i′,t) It represents the probability that cell i is transformed into land use k by the land use K(i') on cell i' with a distance d(i,i') in the neighborhood space, and t represents the transfer time.

[0084] The Gaussian function quantifies the neighborhood effect by dividing it into inertia points and relationship curves. The inertia points represent the inherent characteristics of the land use category and the intensity of the pixels that maintain the original land use type. The relationship curve represents the force affected by the spatial distribution of other land use types in the neighborhood environment, which is defined as:

[0085] Where y(x) represents the influence of any cell with a different land use type from the central cell on the central cell; a, b, c are all control variables of the Gaussian function; when 0<x≤x c When x>x, there is an attractive force between the center cell and the target cell, and the Gaussian function shows the neighborhood interaction related to the distance. c When the neighborhood range x is exceeded c The influence of the cell on the central cell is zero, and the final neighborhood effect is defined as:

[0086]

[0087] in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that cell i is transferred to the target land use k. When the target land use k is the land use type K(i,t) of the current cell i, the probability that cell i maintains its original land use type is d, which is only related to the land use type. When the target land use k is inconsistent with the land use type of the current cell i, the probability that cell i is transferred to the target land use k is the sum of the forces exerted on the central cell i by all cells i' with land use type k in the neighborhood space. Among them, V(i) is the set of all cells in the Moore neighborhood with i as the central cell; a, b, c are related to the initial land type K(i,t) and the target land type k; d(i,i') is the distance between cells i and i', and d(i,i') is greater than 0 and less than the radius of the Moore neighborhood; t represents the transfer time.

[0088] The urban sprawl trend is described by the Gaussian function, whose formula is:

[0089]

[0090] is the probability of urban sprawl of cell i in the concentric circles, r represents the radius of the concentric circles from the city center, a represents the height of the peak, b represents the peak of the peak, describing the hotspot of urban sprawl, c represents the width of the function curve, which is approximately the spatial compactness of urban sprawl, and d·t describes the variation of the Gaussian function along the horizontal axis under the assumption that the macroscopic probability of urban sprawl moves uniformly toward the periphery of the city over time.

[0091] The final land use allocation is determined by roulette wheel selection by comprehensively considering growth probability, neighborhood effect and urban sprawl trend. The calculation formula is:

[0092]

[0093] in represents the overall conversion probability of each land use type, describes the expansion probability of k on cell i, represents the neighborhood effect of land use type k, represents the probability of urban sprawl, γ represents the random perturbation mechanism based on roulette, and t represents the number of iterations.

[0094] A land use change simulation system, comprising:

[0095] Data collection module to collect land use data and drivers;

[0096] Data processing module, used to preprocess and normalize data;

[0097] Model training module for growth probability estimation of land use types using convolutional neural networks;

[0098] Neighborhood effect calibration module, used to calibrate neighborhood effect parameters through Gaussian function and multi-objective optimization algorithm;

[0099] Urban sprawl trend simulation module, used to describe urban sprawl trends through Gaussian functions;

[0100] The land use assignment module is used to determine the final land use type through a roulette wheel selection.

[0101] The data collection module includes a submodule for collecting China's annual land cover dataset with a resolution of 30m and a submodule for collecting data on 14 driving factors.

[0102] The model training module adopts the improved LeNet-5 convolutional neural network model, and the input layer constructs a 21×21 sampling window.

[0103] The land use allocation module determines the final land use type through a roulette wheel selection algorithm by comprehensively considering growth probability, neighborhood effect, and urban sprawl trend.

[0104] Example 2

[0105] Reference Figure 3 to Figure 6 , which is the second embodiment of the present invention, is based on the previous embodiment and takes Wuhan, China as an example, and includes the following steps:

[0106] Step 1: Land use data

[0107] Step 1-1: Land use data area: This patent selects Wuhan, the capital city of Hubei Province, China, as the study area. Due to its comprehensive geographical advantages and talent advantages, it has a strategic position in the rise of Central China. The total area of ​​the city is about 8569.15km2, and the water surface area is 2117.6km2, accounting for nearly 25% of the city. In addition to its superior geographical location and environmental advantages, Wuhan has a total registered population of 9,341,016, a permanent population of 13,648,900, and a permanent population density of 1,593 people / square kilometer in 2021. Therefore, Wuhan plays a leading role in the Yangtze River Economic Belt and the "1+8" urban circle. Wuhan has experienced rapid urbanization in the past decade, and the construction land has increased by 133.43% compared with 2020 (658.64km2). There are 13 districts in Wuhan City. This patent selects 11 districts as the research objects, including the 7 central districts of Jianghan, Jiangan, Qiaokou, Wuchang, Hongshan, Qingshan and Hanyang, as well as Jiangxia District, Caidian District, Hannan District and Dongxihu District.

[0108] Step 1-2: Land use data collection and land use levels:

[0109] The data in this patent include the historical land use map of Wuhan City and the socio-economic and physical geographic driving factors. The land use data were collected using the 30m resolution China Land Cover Dataset (CLCD) in this patent, with an overall accuracy closeness of 79.31%. 14 driving factors related to physical geography and socio-economic characteristics (i.e., land use, precipitation, temperature, elevation, slope, distance to water bodies, population, GDP, distance to highways, distance to main roads, distance to primary roads, distance to secondary roads, distance to tertiary roads, and distance to railways) were selected to evaluate the conversion probability of the CA transformation rule. Six land use types were set in this patent, including agricultural land, forest land, grassland, water area, urban land, and idle land. Since the city center was identified as the central business district (CBD) in 1990, the study area of ​​Wuhan City was divided into 50 concentric rings (each ring was 1km wide) extending from the city center to the rural areas to analyze the macro trend of urban expansion. In view of the purpose of the study, data accuracy and model effectiveness, all spatial data in this study were uniformly resampled to a spatial resolution of 100 m.

[0110] Step 1-3: Data preprocessing:

[0111] The driving factor dataset, initial land use map and land use expansion map are put into the CNN model for training. After multiple training and feedback cycles, the optimal CNN model is selected. The Gaussian function in the patent is used to quantify the spatiotemporal characteristics of urban sprawl and neighborhood effects. Finally, the multi-objective optimization (MOO) algorithm is used to automatically adjust the Gaussian function to improve the simulation accuracy of the CA model during the calibration process.

[0112] Step 2: This patent simulates land use change based on cellular automata

[0113] Step 2-1: Use convolutional neural networks to estimate the growth probability of different land use types

[0114] Step 2-1-1: Input layer. A lightweight CNN based on the improved LeNet-5 is used. This CNN network is known for its simplicity and computational efficiency, which enables faster training and reduced inference time. The input layer constructs a 21×21 sampling window in the neighborhood of each sample unit, and it also includes the driving factor (N=14) and the initial land use map.

[0115] Step 2-1-2: Convolutional layer. A convolution operation with a filter size of 24×3×3 is applied to the input, resulting in an equal-sized feature map of 24×21×21 with edge padding. The second layer is defined by the same convolution size (3×3) and 48 filters, resulting in a feature map of 48×19×19. A max pooling layer with a 2×2 kernel is then used to produce the output image. The fourth and fifth layers are similar to the first two 3×3 convolutional layers. They are used to explore nonlinear factors that affect land use evolution. These two layers output 96×7×7 and 128×5×5 feature maps, respectively. The convolutional network input data first passes through the convolutional layer, which acts as a feature extractor. It plays a key role in processing the neighborhood features of the input data into an output result with high-level features, and learns the feature representation of the input variable through the convolution kernel. Through multiple convolutional layers, the convolutional network model gradually parses the abstract features of different dimensions of the image to obtain deeper feature representations.

[0116] Step 2-1-3: Maximum pooling layer. After the convolution layer operation is completed, although the order of magnitude of parameters is reduced, the amount of data is still large. The work of the pooling layer is mainly to select the data features after convolution extraction, and screen out the features that meet the requirements according to the amount of information, thereby eliminating unimportant features, while reducing the data dimension and avoiding overfitting. This patent extracts the maximum value within the 2×2 kernel neighborhood and reshapes the input to 128×2×2. In order to further enhance the network's ability to extract nonlinear features, the seventh layer consists of 256 1×1 convolution kernels, and the final output is a 256×2×2 map. The principle of the pooling process is to calculate the function value of the data features in the pooling area, and the function value varies depending on the parameters. The mathematical form of the pooling layer is as follows:

[0117]

[0118] In the formula It is the i-th dimension input variable of the l-th pooling layer; and They are the multiplication and addition bias terms of the j-th dimension variable in the l-th pooling layer; The output variable of the j-th dimension variable in the l-th pooling layer; down is the pooling function.

[0119] Step 2-1-4: Fully connected layer. The fully connected layer is usually located at the end of the convolutional network, connecting all output variables of the previous layer (maximum pooling layer). In this patent, the input is converted into a 6×1 feature vector through three fully connected layers, and then softmax regression is performed.

[0120] Step 2-1-5: Output layer. Each vector in the output layer represents the growth probability of the land use type at the corresponding pixel.

[0121] Step 2-2: Use multi-objective optimization to test the parameters of domain effects.

[0122] The neighborhood effect is used to estimate the probability that a grid cell retains its original land use type or converts to a potential target land use type, which is affected by the spatial distribution of the surrounding neighborhood land use. Usually the neighborhood effect is defined as:

[0123]

[0124] in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that the initial land use type of cell i is transferred to k; V(i) is the set of all cells in the Moore neighborhood with i as the central cell; K(i') is the current land use status of any cell i' in the neighborhood; d(i,i') is the distance between cells i and i'; W d(i,i′),K(i′,t) It represents the probability that cell i is transformed into land use k by the land use K(i') on cell i' with a distance d(i,i') in the neighborhood space, and t represents the transfer time.

[0125] To explain the computational cost in detail, we make some simplifying assumptions that apply to this study: if the neighborhood radius is equal to 5 and the number of land use classes is equal to 6, there will be 108 parameters for calibration.

[0126] Therefore, a Gaussian function is used in this patent to quantify the neighborhood effect. We divide the neighborhood effect into inertia points and relationship curves. The inertia points represent the inherent characteristics of the land use category while describing the strength of the pixels that maintain their original land use type. This parameter is not affected by the aggregation pattern of different land use types or multiple driving factors and is only defined by a specific land use category. This setting reduces the difficulty of land use conversion because the overexpansion of the main soil type is limited. On the other hand, the relationship curve describes the force affected by the spatial distribution of other land use types in the neighborhood environment, which is defined as:

[0127]

[0128] Where y(x) represents the influence of any cell with a different land use type from the central cell on the central cell; a, b, c are all control variables of the Gaussian function; when 0<x≤x c When x>x, there is an attractive force between the center cell and the target cell, and the Gaussian function shows the neighborhood interaction related to the distance. c When the neighborhood range x is exceeded c The influence of the cell on the central cell is zero, and the final neighborhood effect is defined as:

[0129]

[0130] in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that cell i is transferred to the target land use k. When the target land use k is the land use type K(i,t) of the current cell i, the probability that cell i maintains its original land use type is d, which is only related to the land use type. When the target land use k is inconsistent with the land use type of the current cell i, the probability that cell i is transferred to the target land use k is the sum of the forces exerted on the central cell i by all cells i' with land use type k in the neighborhood space. Among them, V(i) is the set of all cells in the Moore neighborhood with i as the central cell; a, b, c are related to the initial land type K(i,t) and the target land type k; d(i,i') is the distance between cells i and i', and d(i,i') is greater than 0 and less than the radius of the Moore neighborhood; t represents the transfer time.

[0131] Using the above Gaussian function, the total number of parameters related to neighborhood effects is reduced from 180 to 96. This advantage becomes more obvious as the neighborhood size and land use level increase, which helps to enrich application scenarios across multiple spatial scales and land use levels.

[0132] Step 2-3: Urban expansion trends

[0133] To capture the macro process of LUCC, we incorporate urban sprawl trends into the CA model and use a Gaussian function. Urban sprawl trends summarize land urbanization, reflecting the trend of new urban land expansion and its evolution over time. Sprawl trends are defined as:

[0134]

[0135] is the probability of urban sprawl of cell i in the concentric circles, r represents the radius of the concentric circles from the city center, a represents the height of the peak, b represents the peak of the peak, describing the hotspot of urban sprawl, c represents the width of the function curve, which is approximately the spatial compactness of urban sprawl, and d·t describes the variation of the Gaussian function along the horizontal axis under the assumption that the macroscopic probability of urban sprawl moves uniformly toward the periphery of the city over time.

[0136] By parameterizing the urban sprawl trend using a Gaussian function, we can explain the generation of urban patterns in distant areas and urban gaps in central areas. The urban sprawl trend also reflects the dynamics and complexity of land use, providing an opportunity to develop flexible transition rules and more accurate opportunities for simulating and predicting urban development.

[0137] Step 2-4: Roulette Selection

[0138] Taking into account the growth probability, neighborhood effect and urban sprawl trend, the overall probability is expressed as follows:

[0139]

[0140] in represents the overall conversion probability of each land use type, describes the expansion probability of k on cell i, represents the neighborhood effect of land use type k, represents the probability of urban sprawl, γ represents the random perturbation mechanism based on roulette, and t represents the number of iterations.

[0141] Specifically, the iterative operation of CA is Before adding random disturbances, the preliminary land use transition probability is obtained through network and parameter correction. It is defined as:

[0142]

[0143] After considering the above factors and calculating the combined probability for each cell, a roulette wheel selection is used to determine the final land use type for cell i. This approach describes land use competition because it allows the propagation of disadvantaged land use types with the lowest total probability. After iteration, the winning land use type is assigned. If the land use type reaches its quantity requirement in the target year, the CA iteration ends, if not, the next iteration is performed.

[0144] Where B is the total number of pixels representing a hit (i.e., the observed land-use change was consistent with the model predicted as a change); A is the total number of pixels observed that were inconsistent with the prediction; C is the total number of pixels observed that were inconsistent with the prediction, i.e., changed to an incorrect land-use type; and D is the total number of pixels observed that were inconsistent with the predicted change.

[0145] Step 3: Simulation results.

[0146] Step 3-1: Figure of Merit (FOM) and Overall Accuracy (OA): The location consistency between the actual and simulated land use maps of the CA model is measured using the Figure of Merit (FOM). The FOM index can be expressed as follows:

[0147]

[0148] Where B is the total number of pixels that are hits (i.e., the observed land use changes are consistent with the model predicted as changes); A is the total number of pixels observed that are inconsistent with the predictions; C is the total number of pixels observed that are inconsistent with the predictions, i.e., changed to incorrect land use types; and D is the total number of pixels observed that are inconsistent with the predicted changes. In addition to the FOM index used in the above steps to evaluate accuracy, this patent also uses overall accuracy (OA) to measure precision.

[0149] Step 3-2: Multi-objective optimization - Non-dominated sorting genetic algorithm III: Non-dominated sorting genetic algorithm III is adopted based on its ability to train and correct CA model parameters. The algorithm contains fast non-dominated sorting technology, implicit elite selection mechanism based on Pareto dominance level, and secondary selection method based on target distance, which greatly enhances its performance on challenging multi-objective problems. Because NSGA-III retains a special group of individuals based on the reference point of idealized simulation output, a high-performance LUCC model with excellent simulation accuracy is produced. In this patent, the population size and maximum number of evolution generations of NSGA-III are set to 100 and 200 respectively.

[0150] Step 3-3: Cellular Automata Parameter Verification. The calibration model based on NSGA-III uses 200 population optimizations, generating approximately 20,000 optimal solutions during the calibration process. Each solution corresponds to a set of parameters and is colored to identify the overall optimization round to which it belongs. The final population optimization is given a dark blue color, which constitutes the final Pareto frontier. Figure 8As shown, OA ranges from 0.790 to 0.847, while FOM values ​​range from 0.173 to 0.222. On the Pareto front, the overall accuracy (OA) and quality factor (FOM) values ​​are negatively correlated, which may be due to the large study area and the relatively small proportion of actual LUCC. Therefore, NSGA-III has a weak ability to obtain high OA values ​​through optimization. Finally, three types of parameters are obtained through the peak of the Pareto front: (1) Neighborhood effect inertia points for each land use type; the values ​​of inertia points vary greatly between categories. The inertia point values ​​of farmland and water bodies are the largest, with medians of 8.94 and 8.67, respectively. The study shows that agricultural land and water bodies are restricted by transformation and affected by urbanization. The inertia points of urban land are relatively small and more concentrated than other inertia points, with a median value of 2.91. The inertia points of unused land are the lowest, with a median value of only 1.28, indicating that the transformation of unused land is more likely to occur. (2) Neighborhood effect relationship curve of land use transformation; In this patent, a Gaussian function is fitted to describe the distance-dependent neighborhood effect of each possible land use transformation. (3) Relationship curve of urban sprawl trend; The macro trend of urban expansion is also demonstrated by the Gaussian function. Therefore, since the Gaussian function reflects the hot spots of urban sprawl, the value is 18.20. The values ​​that determine the probability of macro-sprawl of urban land are 0.01 and 6.24 respectively. Using the Gaussian function curve, the spatial compactness of new urban development can be approximated. Applying the urban zoning theory, Wuhan City is divided into four circular areas according to the distance from the city center: urban core area, inner city, suburbs and urban edge. The urban core in this patent includes concentric rings with a distance of less than 11.96km. It is the most densely distributed urban area with limited expansion space, so the probability of macro-expansion is low, less than 0.009. The distance of the inner suburban area ranges from 11.96km to 24.44km. In particular, the probability of urban sprawl increases rapidly with the increase of distance, reaching a maximum value of 0.015 at 18.20km. Suburban expansion mainly occurs at the intersection of cities and suburbs, and the probability quickly decreases as the distance from the city center increases. The probability of urban development sprawl at the city's edge shows similar variation characteristics as that of the suburbs. When the distance exceeds 30.68 kilometers, the rate of decrease in probability slows down until it finally approaches zero.

[0151] Step 4: Model analysis and comparison based on land use change simulation.

[0152] Step 4-1: Improve accuracy through self-calibration. An automatic calibration convolutional neural network CA (AC-CNN-CA) proposed in this patent is compared with the Patch-generating land use simulation (PLUS) model that takes into account micro- and macro-spatial characteristics. The PLUS model uses land expansion mechanisms and patch generation strategies to effectively describe the generation of distant urban patches and is widely used in ecosystem service research and carbon sequestration prediction. The confusion matrix aims to quantify the performance of the model, from which OA and FOM are calculated. Because the overall accuracy (OA) and quality factor (FOM) of the SC-CNN-CA model are 84.12% and 20.20%, respectively, which are much higher than those of the PLUS model (OA = 83.59% and FOM = 14.54%). The model performs better in agricultural land, water area, and urban land, with PAs improved by 0.71%, 1.57%, and 2.24% respectively compared with the PLUS model. Overall, the SC-CNN-CA model achieves higher accuracy (FOM is improved by 5.66%) and has a significant advantage in modeling urban sprawl (Producer Accuracy is improved by 2.24%).

[0153] In addition, this patent compares the actual map of 2015 with the land use simulation output based on the two models. The two regional zooms provide more details, as the A2_Erro, B2_Erro, A3_Erro, and B3_Erro panels show the error patches from the two models. According to the output of SC-CNN-CA, the simulation results in 2015 are highly consistent with the actual results, especially the simulation results of the edge of urban land. For the SC-CNN-CA model, the simulation errors are evenly distributed at the edge of the actual urban sprawl area. In contrast, the simulation errors of the PLUS model are clustered, and the error expansion pattern is mainly at the edge and periphery.

[0154] Step 4-2: Policy impact. The SC-CNN-CA model proposed in this patent found that the conversion of agricultural land and waters is restricted and not easy to convert. This shows that regional policies have a positive impact on the protection of agriculture and waters. However, compared with agricultural land and waters, forests and grasslands are more vulnerable and have lower inertia points. It reminds urban planners and policymakers to protect and manage urban green spaces, because the spatial distribution of forests and grasslands is closely related to various socioeconomic and environmental phenomena, including urban heat islands, carbon emissions, and mental health. The potential application of the SC-CNN-CA model proposed in this patent is spatial optimization and land use prediction based on scenarios designed in a more microscopic way. The urban pattern of optimizing cultivated land protection requires careful adjustment of the inertia points and relationship curves of neighborhood effects. The higher the inertia point value of agricultural land, the greater the possibility of maintenance. Afterwards, in order to prevent the transition from agricultural land to urban land, the amplitude of the Gaussian-based relationship curve should be designed to a lower value. In addition, the model provides a potential interface for simulating comprehensive scenarios to explore the trade-offs and synergistic interactions between different ecosystem services. Researchers can combine the calibration targets of MOO with the CA model to study the optimal spatial distribution of land use. Therefore, it provides more possibilities for scenario-based research on future sustainable land use.

[0155] Step 4-3: Limitations.

[0156] Although the SC-CNN-CA model has high accuracy results in the simulation of land use change in Wuhan, the estimation of the amount of land use change is also important. In future studies, we will combine future land use change scenarios that support land resource allocation and urban management. In addition, there is much work to be done to strengthen the conversion of agricultural land to forest land. During the study period, the forest land in Wuhan showed a uniform rather than clustered spatial configuration, which was difficult to simulate through neighborhood effects. Therefore, in subsequent studies, the stochastic mechanism of the model should be improved so that it is also applicable to uniformly dispersed land expansion patterns. In addition, adding mid-term calibration can obtain a more natural land evolution process, making the model more suitable for future long-term land use prediction. In summary, the model proposed in this study needs to be further optimized for different regions or applications, adding mid-term corrections and optimizing random mechanisms to make the model more reliable in a wider range of applications.

[0157] Example 3

[0158] Reference Figure 3 to Figure 6 , which is the third embodiment of the present invention, and is based on the previous embodiment, step one, land use data.

[0159] Step 1-1: Land use data area: This patent selects Wuhan, the capital city of Hubei Province, China, as the study area. Due to its comprehensive geographical advantages and talent advantages, it has a strategic position in the rise of Central China. The total area of ​​the city is about 8569.15km2, and the water surface area is 2117.6km2, accounting for nearly 25% of the city. In addition to its superior geographical location and environmental advantages, Wuhan has a total registered population of 9,341,016, a permanent population of 13,648,900, and a permanent population density of 1,593 people / square kilometer in 2021. Therefore, Wuhan plays a leading role in the Yangtze River Economic Belt and the "1+8" urban circle. Wuhan has experienced rapid urbanization in the past decade, and the construction land has increased by 133.43% compared with 2020 (658.64km2). There are 13 districts in Wuhan City. This patent selects 11 districts as the research objects, including the 7 central districts of Jianghan, Jiangan, Qiaokou, Wuchang, Hongshan, Qingshan and Hanyang, as well as Jiangxia District, Caidian District, Hannan District and Dongxihu District.

[0160] Step 1-2: Land use data collection and classification:

[0161] The data in this patent include the historical land use map of Wuhan City and the socio-economic and physical geographic driving factors. The land use data were collected using the 30m resolution China Land Cover Dataset (CLCD) in this patent, with an overall accuracy closeness of 79.31%. 14 driving factors related to physical geography and socio-economic characteristics (i.e., land use, precipitation, temperature, elevation, slope, distance to water bodies, population, GDP, distance to highways, distance to main roads, distance to primary roads, distance to secondary roads, distance to tertiary roads, and distance to railways) were selected to evaluate the conversion probability of the CA transformation rule. Six land use types were set in this patent, including agricultural land, forest land, grassland, water area, urban land, and idle land. Since the city center was identified as the central business district (CBD) in 1990, the study area of ​​Wuhan City was divided into 50 concentric rings (each ring was 1km wide) extending from the city center to the rural areas to analyze the macro trend of urban expansion. In view of the purpose of the study, data accuracy and model effectiveness, all spatial data in this study were uniformly resampled to a spatial resolution of 100 m.

[0162] Step 1-3: Data preprocessing:

[0163] The driving factor dataset, initial land use map and land expansion map are put into the CNN model for training. After multiple training and feedback cycles, the optimal CNN model is selected. The Gaussian function is used in the patent to measure the spatiotemporal characteristics of urban sprawl and neighborhood effects. Finally, the multi-objective optimization (MOO) algorithm is used to automatically adjust the Gaussian function to improve the simulation accuracy of the CA model during the calibration process.

[0164] Step 2: Convolutional network construction.

[0165] This study includes a total of one input layer, five convolutional layers, two maximum pooling layers, followed by three fully connected layers and an output layer. The first layer is a 24×21×21 feature map output by the input layer; the second layer is a convolutional layer with a 48×19×19 feature map defined by the same convolution size (3×3) and 48 filters; the third layer uses a maximum pooling layer with a 2×2 convolution kernel to produce a 48×9×9 image; 16×5×5; the fourth and fifth layers are similar to the first two convolutional layers with 3×3 kernels. They are used to explore other nonlinear factors affecting land use evolution, but these two layers output 96×7×7 and 128×5×5 feature maps respectively; then the maximum value within the 2×2 kernel neighborhood is extracted, and the input is reshaped into a 128×2×2 sixth layer. The seventh layer consists of 256 1×1 convolution kernels and finally outputs a 256×2×2 map. Finally, the input is converted into a 6×1 feature vector through three fully connected layers.

[0166] Step 3: Simulate land use change based on CA model

[0167] Step 3-1: Use multi-objective optimization to test the parameters of domain effects.

[0168] The neighborhood effect is used to estimate the probability of a grid cell retaining its original land use type or converting to a potential target land use type, which is affected by the spatial distribution of land uses in the surrounding neighborhood.

[0169] In this patent, a Gaussian function is used to quantify the neighborhood effect. We divide the neighborhood effect into inertia points and relationship curves. The inertia points represent the inherent characteristics of the land use category while describing the strength of the pixels that maintain their original land use type. This parameter is not affected by the aggregation pattern of different land use types or multiple driving factors and is only defined by a specific land use category. This setting reduces the difficulty of land use conversion because the overexpansion of the main soil type is limited. On the other hand, the relationship curve describes the force affected by the spatial distribution of other land use types in the neighborhood environment.

[0170] After using the Gaussian function in this study, the total number of parameters related to neighborhood effects was reduced from 180 to 96. This advantage becomes more obvious as the neighborhood size and land use level increase, which helps to enrich application scenarios across multiple spatial scales and land use levels.

[0171] Step 3-2: Urban expansion trends

[0172] To capture the macroscopic process of LUCC, we incorporate the urban sprawl trend into the CA model and utilize the Gaussian function. The urban sprawl trend summarizes the urbanization of land, reflecting the trend of new urban land expansion and its pattern of evolution over time. By parameterizing the urban sprawl trend using the Gaussian function, we can explain the generation of urban patterns in distant areas and urban gaps in central areas. The urban sprawl trend also reflects the dynamics and complexity of LUCC, providing an opportunity to develop flexible transition rules and more accurate opportunities for simulating and predicting urban development.

[0173] Step 4: Simulation results.

[0174] Step 4-1: Calculate the quality factor (FOM) and overall accuracy (OA): The location consistency between the actual and simulated land use maps of the CA model is measured using the quality factor (FOM). In this study, the OA values ​​of the SC-CNN-CA model and the PLUS model were 84.12% and 83.59%, respectively, and the FOM values ​​were 20.20% and 14.54%, respectively, reflecting the superiority of the SC-CNN-CA model in accuracy assessment.

[0175] Step 4-2: Multi-objective optimization - Non-dominated sorting genetic algorithm III: Non-dominated sorting genetic algorithm III was adopted based on its ability to train and correct the parameters of the CA model. The algorithm contains a fast non-dominated sorting technique, an implicit elite selection mechanism based on Pareto dominance level, and a secondary selection method based on target distance, which greatly enhances its performance on challenging multi-objective problems. Because NSGA-III retains a special group of individuals based on the reference point of the idealized simulation output, it produces a high-performance LUCC model with excellent simulation accuracy.

[0176] Step 4-3: Cellular automaton parameter verification. Cellular automaton parameter verification. The calibration model based on NSGA-III uses 200 population optimizations, generating approximately 20,000 optimal solutions during the calibration process. Each solution corresponds to a set of parameters and is colored to identify the overall optimization round to which it belongs. The final population optimization is given a dark blue color, which constitutes the final Pareto frontier. Figure 8As shown, OA ranges from 0.790 to 0.847, while FOM values ​​range from 0.173 to 0.222. On the Pareto front, the overall accuracy (OA) and quality factor (FOM) values ​​are negatively correlated, which may be due to the large study area and the relatively small proportion of actual LUCC. Therefore, NSGA-III has a weak ability to obtain high OA values ​​through optimization. Finally, three types of parameters are obtained through the peak of the Pareto front: (1) Neighborhood effect inertia points for each land use type; the values ​​of inertia points vary greatly between categories. The inertia point values ​​of farmland and water bodies are the largest, with medians of 8.94 and 8.67, respectively. The study shows that agricultural land and water bodies are restricted by transformation and affected by urbanization. The inertia points of urban land are relatively small and more concentrated than other inertia points, with a median value of 2.91. The inertia points of unused land are the lowest, with a median value of only 1.28, indicating that the transformation of unused land is more likely to occur. (2) Neighborhood effect relationship curve of land use transformation; In this patent, a Gaussian function is fitted to describe the distance-dependent neighborhood effect of each possible land use transformation. (3) Relationship curve of urban sprawl trend; The macro trend of urban expansion is also demonstrated by the Gaussian function. Therefore, since the Gaussian function reflects the hot spots of urban sprawl, the value is 18.20. The values ​​that determine the probability of macro-sprawl of urban land are 0.01 and 6.24 respectively. Using the Gaussian function curve, the spatial compactness of new urban development can be approximated. Applying the urban zoning theory, Wuhan City is divided into four circular areas according to the distance from the city center: urban core area, inner city, suburbs and urban edge. The urban core in this patent includes concentric rings with a distance of less than 11.96km. It is the most densely distributed urban area with limited expansion space, so the probability of macro-expansion is low, less than 0.009. The distance of the inner suburban area ranges from 11.96km to 24.44km. In particular, the probability of urban sprawl increases rapidly with the increase of distance, reaching a maximum value of 0.015 at 18.20km. Suburban expansion mainly occurs at the intersection of cities and suburbs, and the probability quickly decreases as the distance from the city center increases. The probability of urban development sprawl at the city's edge shows similar variation characteristics as that of the suburbs. When the distance exceeds 30.68 kilometers, the rate of decrease in probability slows down until it finally approaches zero.

[0177] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the invention, which should be included in the scope of the claims of the present invention.

Claims

1. A land use simulation method based on convolutional neural network, characterized in that: include, Collect land use data, socioeconomic and physical geographic drivers of the study area and perform data preprocessing; The improved LeNet-5 convolutional neural network is used to train land use data to estimate the growth probability of different land use types; The neighborhood effect is quantified by Gaussian function, and the parameters are calibrated using a multi-objective optimization algorithm; The urban sprawl trend is described by Gaussian function, capturing the macro-urban expansion process; The final land use type is determined through roulette wheel selection, taking into account growth probability, neighborhood effect and urban sprawl tendency.

2. The land use simulation method based on convolutional neural network according to claim 1, characterized in that: The collection of land use data in the study area includes using the annual land cover dataset of China with a resolution of 30m to collect land use data; The data preprocessing includes processing the driving factor data set, the initial land use map and the land use expansion map.

3. The land use simulation method based on convolutional neural network as claimed in claim 2, characterized in that: The improved LeNet-5 convolutional neural network is used to train land use data to estimate the growth probability of different land use types, including: A 21×21 sampling window is constructed at the input layer, which includes 14 driving factors and the initial land use map; Extract image features through multiple convolutional layers; The pooling layer is used to select the data features extracted by the convolution layer, and the function value is calculated for the data features in the pooling area. The formula is as follows: In the formula is the i-th dimension input variable of the l-th pooling layer, and They are the multiplication and addition bias terms of the j-th dimension variable in the l-th pooling layer; is the output variable of the j-th dimension variable in the l-th pooling layer; down is the pooling function; The connection layer connects all the output variables of the pooling layer; Each vector in the output layer represents the probability of growth of a land-use type at the corresponding cell.

4. The land use simulation method based on convolutional neural network as claimed in claim 3, characterized in that: The Gaussian function is used to quantify the attraction and neighborhood effect between land use types, and the neighborhood effect is defined as: in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that the initial land use type of cell i is transferred to k; V(i) is the set of all cells in the Moore neighborhood with i as the central cell; K(i') is the current land use status of any cell i' in the neighborhood; d(i,i′) is the distance between cells i and i′; W d(i,i′),K(i′,t) It represents the probability that cell i is transformed into land use k by the influence of land use K(i') on cell i' with a distance d(i,i') in the neighborhood space, and t represents the transfer time; The Gaussian function quantifies the neighborhood effect by dividing the neighborhood effect into an inertia point and a relationship curve. The inertia point represents the inherent characteristics of the land use category and the strength of the pixel that maintains the original land use type. The relationship curve represents the force affected by the spatial distribution of other land use types in the neighborhood environment, which is defined as: Where y(x) represents the influence of any cell with a different land use type from the central cell on the central cell; a, b, c are all control variables of the Gaussian function; when 0<x≤x c When x>x, there is an attractive force between the center cell and the target cell, and the neighborhood interaction related to the distance is shown by the Gaussian function; when x>x c When the neighborhood range x is exceeded c The influence of the cell on the central cell is zero, and the final neighborhood effect is defined as: in is the force exerted by land use k on the central cell i in the neighborhood space, resulting in the probability that cell i is transferred to the target land use k. When the target land use k is the land use type K(i,t) of the current cell i, the probability that cell i maintains its original land use type is d, which is only related to the land use type. When the target land use k is inconsistent with the land use type of the current cell i, the probability that cell i is transferred to the target land use k is the sum of the forces exerted on the central cell i by all cells i' with land use type k in the neighborhood space, where V(i) is the set of all cells in the Moore neighborhood with i as the center unit; a, b, c are related to the initial land type K(i,t) and the target land type k; d(i,i') is the distance between cells i and i', and d(i,i') is greater than 0 and less than the radius of the Moore neighborhood; t represents the transfer time.

5. The land use simulation method based on convolutional neural network according to claim 1, characterized in that: The urban sprawl trend is described by a Gaussian function, whose formula is: is the probability of urban sprawl of cell i in the concentric circles, r represents the radius of the concentric circles from the city center, a represents the height of the peak, b represents the peak of the peak, describing the hotspot of urban sprawl, c represents the width of the function curve, which is approximately the spatial compactness of urban sprawl, and d·t describes the variation of the Gaussian function along the horizontal axis under the assumption that the macroscopic probability of urban sprawl moves uniformly toward the periphery of the city over time.

6. A land use change simulation system as claimed in claim 5, characterized in that: The final land use allocation is determined by roulette wheel selection by comprehensively considering growth probability, neighborhood effect and urban sprawl trend. The calculation formula is: in represents the overall conversion probability of each land use type, describes the expansion probability of k on cell i, represents the neighborhood effect of land use type k, represents the probability of urban sprawl, γ represents the random perturbation mechanism based on roulette, and t represents the number of iterations.

7. A land use change simulation system, comprising the land use change simulation system according to claims 1 to 6, characterized in that: include: Data collection module to collect land use data and drivers; Data processing module, used to preprocess and normalize data; Model training module for growth probability estimation of land use types using convolutional neural networks; Neighborhood effect calibration module, used to calibrate neighborhood effect parameters through Gaussian function and multi-objective optimization algorithm; Urban sprawl trend simulation module, used to describe urban sprawl trends through Gaussian functions; The land use assignment module is used to determine the final land use type through a roulette wheel selection.

8. A land use change simulation system as claimed in claim 3, characterized in that: The data collection module includes a submodule for collecting 30m resolution China's annual land cover dataset and a submodule for collecting 14 driving factor data.

9. A land use change simulation system as claimed in claim 8, characterized in that: The model training module adopts an improved LeNet-5 convolutional neural network model, and the input layer constructs a 21×21 sampling window.

10. A land use change simulation system as claimed in claim 9, characterized in that: The land use allocation module determines the final land use type through a roulette wheel selection algorithm by comprehensively considering growth probability, neighborhood effect and urban sprawl trend.

Images