An ecological constraint calibration method, terminal device, and medium based on CNN and GCN

By employing an ecological soft constraint calibration method based on CNN and GCN, the subjective problem of ecological constraint rule design in urban cellular automata modeling is solved. This method enables accurate capture and automated acquisition of ecological factors across multiple spatiotemporal scales, thereby improving the accuracy of urban expansion simulation and the adaptability of ecological constraints.

CN123116757APending Publication Date: 2026-11-17JIMEI UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510617541.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing urban cellular automata modeling, there are challenges in accurately capturing and automatically acquiring the impact of ecological constraint mechanisms on urban development across multiple scales and time periods. In particular, the role of ecological factors in urban expansion is not fully utilized, and the design of ecological constraint rules is too subjective.

Method used

An ecological soft constraint calibration method based on CNN and GCN is adopted. By constructing an ecological soft constraint model, high-dimensional features of ecological space are mined using convolutional neural networks and improved graph convolutional neural networks. Combined with a probability coupling module, the ecological soft constraint probability is automatically obtained. Then, it is combined with an urban cellular automata model to predict the development probability of urban land use.

Benefits of technology

It improved the accuracy of urban expansion simulation, enhanced FoM and Recall-based Efficacy indices, and improved the adaptability and accuracy of ecological constraints to urban development.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116757A_ABST
    Figure CN123116757A_ABST
Patent Text Reader

Abstract

This invention relates to an ecological constraint calibration method, terminal device, and medium based on CNN and GCN. The method includes: constructing an ecological soft constraint model; predicting the ecological soft constraint probability of each pixel in the study area being developed into urban land using the ecological soft constraint model; the ecological soft constraint model includes a convolutional neural network module, an improved graph convolutional neural network module, and a probability coupling module, used to perform pixel-level and superpixel-level model training respectively, and finally coupling the ecological soft constraint probabilities mined by them; based on the terrain factors, ecological factors, and driving factors of the study area, obtaining the cell conversion probability of each pixel being developed into urban land using an urban cellular automata model; the product of the cell conversion probability and the ecological soft constraint probability is taken as the final probability of each pixel in the study area being developed into urban land. This invention achieves more accurate evolutionary results compared to existing models.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of urban evolution, and in particular to an ecological constraint calibration method, terminal device and medium based on CNN and GCN. Background Technology

[0002] Against the backdrop of global climate change and rapid urbanization, the complex coupling and feedback processes between urban expansion and natural ecological space continue to attract widespread attention. Urban residents are highly dependent on various key ecosystem services provided by natural ecosystems, such as water conservation, soil retention, water purification, and biodiversity maintenance. However, urban sprawl driven by the land market is constantly exacerbating resource and environmental pressures, leading to prominent problems such as habitat loss, water pollution, the urban heat island effect, and soil erosion. Therefore, urban development urgently needs to preserve key ecological resources in its spatial layout to ensure the sustainable supply of ecosystem services; at the same time, it should minimize disturbance to natural ecosystems.

[0003] In this process, the interactions and mobility of ecosystem services are highly correlated with urban sprawl. Urban expansion significantly alters or even disrupts the synergistic and trade-off relationships of ecosystem services, thereby affecting ecosystem stability. Simultaneously, the dynamic transfer of ecosystem service flows across spatiotemporal scales determines the intensity and scope of ecological constraints faced by urban development. The complex interactions and trade-offs between social systems and ecosystems urgently require support and guidance from effective spatial planning tools. Currently, many urbanized areas face the dual challenges of scarce land resources and fragile ecological environments; therefore, accurately predicting the spatiotemporal patterns of urban growth under the constraints of ecosystem service importance is crucial.

[0004] Accurately reproducing historical changes in land use and predicting its future evolution is one of the core issues in exploring urban expansion patterns. With the rapid development of Geographic Information Systems (GIS), cellular automata (CA) have been deeply integrated and have shown great potential in expressing the spatiotemporal processes of land use change. Existing cellular automata modeling research mainly focuses on exploring how factors such as transformation rules, neighborhood effects, and cellular configuration influence the evolution of urban spatial morphology. In recent years, urban cellular automata have made significant progress in revealing the intrinsic laws of urban growth, especially in the internal differences in land demand and the neighborhood influence of driving factors. Urban development is essentially a complex nonlinear system, and its bottom-up evolution process exhibits spatial characteristics of intertwined spontaneous and self-organized growth. However, existing research often overlooks the role of some system-endogenous potential constraints in urban expansion, particularly the influence of complex ecological and environmental factors on the formation process of these constraints in both depth and breadth.

[0005] Hard constraint mechanisms play a crucial role in urban cellular automata modeling, particularly in restricting land use type transformation. These mechanisms employ a Boolean constraint strategy, controlling whether the cell state changes by setting the constraint value to 0 or 1. For example, the SLEUTH model incorporates an exclusion layer in the input layer to define areas where urbanization is impossible. In water bodies, high-quality farmland, ecological red line zones, and urban planning control zones, hard constraint mechanisms have a direct and effective impact on urban evolution. However, due to the significant spatial heterogeneity of ecological factors, hard constraint mechanisms cannot effectively prevent ecological losses caused by urban expansion outside protected areas. How to adjust urban development policies in this context remains a key issue for planners.

[0006] Some studies have constructed soft constraint rules by setting a range of constraint values ​​(e.g., [0,1]), making changes in cell states more flexible. When the soft constraint value is close to 0, the cell state hardly changes; when it is close to 1, it fully supports cell development. Soft constraint values ​​can be continuous or segmented, thus providing more room for adjustment in land use change simulations. In recent years, soft constraint mechanisms have gradually incorporated ecological factors, enhancing the application of cellular automata in urban growth simulations. Although soft constraint mechanisms are more adaptive, accurately capturing the impact of ecological factors on urban development at multiple scales in time and space remains a challenge in constructing ecological soft constraint rules.

[0007] Currently, many studies have applied hard or soft constraint mechanisms to conduct scenario simulations from an ecological protection perspective. However, urban cellular automata modeling based on ecological constraints largely focuses on projecting future development scenarios. Because ecological constraint mechanisms often have pre-set goals, their design process is heavily influenced by human factors, and the acquisition of many ecological constraints is subjective or arbitrary. Looking back at the history of urban development, its spatiotemporal evolution trajectory is actually a dynamic process of continuous coordination between the supply of specific natural ecological resources and the needs of urban residents. This indicates that specific ecological constraint mechanisms are inherent in the historical development of cities. Therefore, how to automatically acquire these ecological constraint rules has become one of the key scientific problems in urban cellular automata modeling. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes an ecological constraint calibration method, terminal device, and medium based on CNN and GCN.

[0009] The specific plan is as follows:

[0010] An ecological constraint calibration method based on CNN and GCN includes the following steps:

[0011] S1: Construct an ecological soft constraint model to predict the probability of each pixel in the study area being developed into urban land.

[0012] The ecological soft constraint model includes a convolutional neural network module, an improved graph convolutional neural network module, and a probability coupling module;

[0013] The model input data is set as the topographic and ecological factors of the area to be studied;

[0014] After receiving the input data from the model, the convolutional neural network module outputs the first predicted probability of each pixel being developed into urban land.

[0015] The improved graph convolutional neural network module first divides the model input data into several superpixels using a superpixel generation algorithm, and then inputs the superpixels as graph vertices into the traditional graph convolutional neural network. Next, a decoder is added before the linear layer of the traditional graph convolutional neural network to convert the superpixel-scale features output by the convolutional layer into pixel-scale features. Finally, the output of the traditional graph convolutional neural network is used as the second prediction probability that each pixel is developed into urban land.

[0016] The probability coupling module is used to couple the first prediction probability and the second prediction probability into an ecological soft constraint probability;

[0017] S2: Based on the topographic factors, ecological factors and driving factors of the area under study, the probability of cell conversion of each pixel being developed into urban land is obtained through the urban cellular automata model;

[0018] S3: The product of the probability of cell transformation and the probability of ecological soft constraint is the final probability that each pixel in the area under study will be developed into urban land.

[0019] Furthermore, topographic factors include elevation and slope, while ecological factors include: importance of water conservation services, importance of soil and water conservation services, importance of biodiversity conservation services, soil fertility, and NDVI.

[0020] Furthermore, the driving factors are distance-based variables, including: distance from the city center, distance from the town center, distance from the highway, distance from the railway, distance from the main road, distance from the train station, and distance from the coastline.

[0021] Furthermore, the formula for calculating the probability of ecological soft constraints is as follows:

[0022] P ESC =P CNN ×(P SGCN ) γ

[0023] Among them, P ESC P represents the probability of ecological soft constraints. CNN Let P represent the first predicted probability. SGCNγ represents the second prediction probability, and γ represents the transformation exponent.

[0024] Furthermore, when a superpixel is used as a graph vertex, the features of the vertex are characterized by the average value of all pixels contained in the corresponding superpixel.

[0025] An ecological constraint calibration terminal device based on CNN and GCN includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the method described in the embodiments of the present invention.

[0026] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above in the embodiments of the present invention.

[0027] This invention employs the above-mentioned technical solution, achieving more accurate evolutionary results compared to the widely used ANN-CA model. The model used in this invention (ANN-SGCNN-CA) achieves improvements of 5.60% and 7.0% in FoM and Recall-based Efficacy metrics, respectively. Furthermore, compared to the ANN-CNN-CA model that calibrates ecological soft constraint rules through convolutional neural networks, the model used in this invention achieves increases of 2.37% and 2.46% in FoM and Recall-based Efficacy, respectively. Attached Figure Description

[0028] Figure 1 The diagram shown is a flowchart of a method according to an embodiment of the present invention.

[0029] Figure 2 The diagram shown is a schematic representation of the overall framework of the model in this embodiment. Detailed Implementation

[0030] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention.

[0031] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0032] Example 1:

[0033] This invention provides an ecological constraint calibration method based on CNN and GCN, such as... Figure 1 As shown, the main steps of the method are as follows:

[0034] S1: Construct an ecological soft constraint model to predict the probability of each pixel in the study area being developed into urban land.

[0035] S2: Based on the topographic factors, ecological factors and driving factors of the area under study, the probability of cell conversion of each pixel being developed into urban land is obtained through the urban cellular automata model;

[0036] S3: The product of the probability of cell transformation and the probability of ecological soft constraint is the final probability that each pixel in the area under study will be developed into urban land.

[0037] The specific details involved in this method are described below.

[0038] As mentioned earlier, the historical development of cities has always been constrained by specific natural resources and ecological spaces. This embodiment attempts to propose the following two assumptions:

[0039] Hypothesis 1: The spatial expansion of urbanized areas is subject to specific ecological constraints;

[0040] Hypothesis 2: Ecological factors have a constraining effect on the development of the central pixel within a certain spatial range.

[0041] To verify the above hypotheses, this embodiment attempts to construct a soft constraint mechanism for urban cellular automata based on deep learning. First, in accordance with relevant guidelines, an evaluation index for the importance of ecosystem services is constructed based on the quantitative index method. Second, a high-dimensional feature representation of ecological space is mined using the CNN algorithm. Then, superpixels are segmented using the SLIC algorithm, and ecological space constraint information at the superpixel scale is obtained using the GCN algorithm. Finally, the ecological soft constraint probabilities obtained by CNN and GCN are coupled as constraint rules for urban cellular automata to conduct historical simulations and future scenario predictions.

[0042] 1. Area to be studied

[0043] Quanzhou City's area lies between 117°25′ and 119°05′E, and 24°30′ and 25°56′N. As one of the three major central cities in Fujian Province, its subordinate counties of Jinjiang, Nan'an, Hui'an, Shishi, and Anxi are among China's top 100 contributors to GDP. It has a large population, rapidly developing industries, and a rapid expansion of urban land for residential, industrial, and commercial use throughout the city. The climate is warm and humid with abundant rainfall year-round. The terrain is mountainous, with higher elevations in the northeast and lower elevations in the southeast, and mountains and hills account for four-fifths of the total land area. Numerous streams and rich vegetation cover provide excellent ecosystem services for biodiversity, water conservation, and soil and water conservation. With rapid economic development and frequent population movement, urban land has spread across the coastal plain and is scattered along river valleys and basins. The tense relationship between humans and land and the severe situation of industrial pollution control may trigger regional or local natural disasters and ecological risks at any time, making Quanzhou an ideal place to explore the ecological constraints in urban development.

[0044] 2. Data Sources and Data Processing

[0045] To express the cellular state, driving factors, and constraint mechanisms of the urban cellular automata, this embodiment selects land use, topography, location, transportation, and ecological factors as input data (Table 1). Four land use maps from 2000 to 2022 were obtained by classifying Landsat images using the GEE platform and a random forest classifier. The classification results cover types such as cropland, forest, shrub, grassland, water, barren, impervious, and wetland, with a resolution of 30m. To characterize the spatiotemporal evolution of urban spatial morphology, the acquired land use was further reclassified into three categories: urban, non-urban, and water. The digital elevation model (DEM) was derived from SRTM data jointly measured by NASA and the National Institute of Mapping (NIMA). Data on administrative centers, roads at all levels, railway stations, and coastlines were obtained from transportation and cadastral data released by the Quanzhou Municipal Bureau of Natural Resources.

[0046] Ecological factors mainly consist of ecosystem service importance, soil fertility, and NDVI. Temperature and precipitation, net primary productivity (NPP) of vegetation, and soil texture are the main GIS data used to assess the importance of ecosystem services. Among them, temperature and precipitation data are from the China Surface Annual Data Set (1981-2010) of the China Meteorological Data Network, NPP is from NASA's MOD17 data product, soil texture data is from the China Soil Data Set collected and compiled by the Northwest Institute of Eco-Environment and Resources of the Chinese Academy of Sciences, soil fertility data is from the Soil Science Data Center constructed by the Nanjing Institute of Soil Science of the Chinese Academy of Sciences, and NDVI is from the "2000-2022 China 30-meter Annual Maximum NDVI Data Set" released by the National Ecological Science Data Center.

[0047] The data collected in this embodiment includes three categories: topographic factors, ecological factors, and driving factors. The data input to the ecological soft constraint model only includes topographic and ecological factors, while the data input to the urban cellular automata model includes all three categories. Specifically, this embodiment includes 2 topographic factors, 7 driving factors, and 5 ecological factors, as shown in Table 1. Topographic factors consist of elevation and slope. The slope factor is obtained by inputting elevation into the ArcGIS platform for slope analysis. Proximity to the administrative center and transportation infrastructure is the main driving factor influencing whether a cell's state changes from non-urban to urban. This embodiment mainly considers seven distance-based variables: distance from the city center, distance from the town center, distance from the highway, distance from the railway, distance from the main road, distance from the train station, and distance from the coastline. Proximity to the coastline was chosen because Quanzhou City is promoting the strategic decision of "building a high-quality marine economy." Through ecosystem service importance assessment, three ecological factors were obtained: the importance of water conservation services, the importance of soil and water conservation services, and the importance of biodiversity maintenance services. To characterize the ecological adaptability of soil, soil fertility was calculated by comprehensively considering total nitrogen, total phosphorus, total potassium, and total organic matter. To eliminate the influence of dimensions and reduce model errors, the obtained indicators such as ecosystem service importance and soil fertility were classified into 5 levels using the natural breakpoint method on the ArcGIS platform and then standardized. Other spatial variables, such as topography, proximity, and NDVI, were directly standardized.

[0048] Table 1

[0049]

[0050] 3. Methods

[0051] The technical roadmap of the model proposed in this embodiment is as follows: Figure 2As shown in the diagram, this framework consists of two parts: the construction of ecological soft constraint rules and the simulation of urban cellular automata. The former first constructs an ecological factor dataset (including topographic and ecological factors) based on an assessment of ecosystem services. Then, it mines the ecological constraint probabilities at the pixel and superpixel scales using CNN and GCN respectively, ultimately coupling them into an ecological soft constraint probability map. The latter integrates driving factors, neighborhood functions, roulette wheel, and ecological soft constraints, using the FLUS model to simulate the urban morphological evolution and future scenarios from 2006 to 2022.

[0052] The ecological factors constructed in this embodiment cover the importance of water conservation services, soil and water conservation services, biodiversity conservation services, soil fertility, and NDVI. Ecosystem service importance indicators are assessed using the NPP quantitative indicator method. NPP refers to the energy fixed by plants, minus the energy consumed by their respiration, that remains available for plant growth and reproduction.

[0053] (1) Importance of water conservation services

[0054] Water conservation services refer to the processes and capabilities of organisms and their environment in retaining water at specific temporal and spatial scales, in conjunction with hydrological processes. Specifically, water conservation is achieved through the interaction of ecosystems such as forests and grasslands with water in several ways: first, by intercepting rainfall through the canopy; second, by utilizing their unique structures to facilitate water infiltration; third, by accumulating water in the soil layer; and fourth, by regulating the water cycle through evapotranspiration. For humans, water conservation plays a crucial role in replenishing groundwater, mitigating surface runoff, reducing seasonal fluctuations in river flow, and ensuring water quality. The Water Conservation Service Capacity Index (WCI) is evaluated using the following formula:

[0055] WCI = NPP m ×F sic ×F p ×(1-F s (1)

[0056] Among them, NPP m F represents the average net primary productivity of vegetation over many years. sic F is the soil seepage factor. p F is the multi-year average precipitation factor. s This is the slope factor.

[0057] (2) Importance of Soil and Water Conservation Services

[0058] Soil and water conservation refers to the process by which ecosystems such as forests and grasslands, through their structure and processes, promote the ecological balance of water, soil, and vegetation, achieving a positive succession state of interdependence and self-renewal between water and soil. Specifically, soil and water conservation services focus on the process by which ecosystems reduce soil loss caused by water erosion, primarily manifested in the ability of vegetation roots to stabilize the soil and reduce the continuous loss of surface soil due to precipitation. If k is a soil erodibility factor, the Soil and Water Conservation Service Capacity Index (SCI) can be calculated using the following formula:

[0059] SCI = NPP m ×(k-1)×(1-F s (2)

[0060] (3) Importance of biodiversity conservation services

[0061] Biodiversity conservation refers to the role of ecosystems in maintaining diversity at the genetic, species, and ecosystem levels. It is primarily reflected in the ecosystem's capacity to generate and maintain biodiversity and its environment, regulate climate, control pests, disperse seeds and pollen, mitigate drought and flood disasters, and purify the soil environment. The biodiversity conservation service capacity index is calculated using the following formula:

[0062] BMI = NPP m ×F p ×F t ×(1-F a (3)

[0063] Among them, F t F represents the multi-year average temperature. a For altitude factors.

[0064] 4. Construction of an ecological soft constraint model

[0065] The key to constructing an ecological soft constraint model is to automatically obtain the probability of constraining the development of a cell by leveraging ecological factors and the classification state of the cell. If the ecological constraint dataset (ECD) is represented in the form of raster data (each raster is a pixel; in this embodiment, the size of a raster is set to 30*30), then it can be represented as E∈R h×w×c Where h, w, and c represent the number of rows, columns, and layers in the raster dataset ECD (5+2=7 in this example), respectively. Further, using... This represents the supervised dataset, where k = 1…M represents the cell index after flattening the spatial grid, M represents the total number of cells, and U(c k ) indicates in c k =Cell E(c) at the spatial location (i,j) kThe labels corresponding to each pixel are defined. Therefore, the defined supervised dataset can be used to train the ecological soft constraint model and predict whether each pixel will develop into urban land. To construct this supervised dataset (i.e., the training set) for training the ecological soft constraint model, this embodiment overlays land use classification results from 2006 and 2022 to detect whether the land use status of each pixel within the region has changed. If a pixel changes from non-urban to urban, its label value is set to 1; otherwise, it is 0. The labeled data is used as the training set for the ecological soft constraint model.

[0066] The ecological soft constraint model proposed in this embodiment consists of three modules: CNN (Convolutional Neural Network), SGCN (Improved Graph Convolutional Neural Network), and probabilistic coupling. These modules are used to perform pixel-level and superpixel-level model training and finally couple the ecological soft constraint probabilities they mine.

[0067] (1) CNN module

[0068] Convolutional Neural Networks (CNNs) are a type of feedforward neural network that incorporates convolutional computations. Convolutional operations enable their artificial neurons to perceive surrounding units within the coverage area of ​​the convolutional kernel. The structure of a CNN mainly consists of an input layer, convolutional layers, and fully connected layers, and also includes associated weights, activation functions, and pooling layers. The convolutional layer is the core of the CNN; assuming C... l This is the feature map output by the l-th convolutional layer of the CNN, and its calculation formula is as follows:

[0069] C l =f(W l *BN(C l-1 )+b l (4)

[0070] Among them, C l-1 This is the feature map of layer l-1. BN applies C before the convolution operation. l-1 Perform normalization operation, W l and b l ...

[0071] P CNN =Softmax(FC(Dropout(C l ))) (5)

[0072] Dropout pairs C with specific probabilities. lThe pixels in the feature map are randomly discarded. FC stands for fully connected operation. Softmax is a normalized exponential function that predicts the probability of each category from the input features.

[0073] This embodiment slices the ecological factor dataset E, and uses the position c of the center pixel of the slice in the grid space. t The corresponding U(c) t The labels are used as the input data for the CNN model. After obtaining all slices, random sampling is performed based on their label categories. The structure of the CNN convolutional subnetwork consists of 2 convolutional layers, 3 dropout layers, and 2 fully connected layers. The kernel size of the two convolutional layers is 3×3, with 32 and 48 kernels respectively. The dropout rates of the three dropout layers are 0.25, 0.25, and 0.50, respectively. The number of neurons in the two fully connected layers is 64 and 2, respectively. The activation function used in the convolutional layers is the Rectified Linear Unit (ReLU). The use of this activation function can accelerate the learning process while mitigating gradient explosion or vanishing.

[0074] (2) SGCN module

[0075] Graph Convolutional Neural Networks (GCNs) are artificial neural networks that extract spatial features from irregular topological graphs through convolution operations. The construction of the graph structure is crucial for implementing the GCN algorithm. This embodiment utilizes the Simple Linear Iterative Clustering (SLIC) algorithm to segment the cellular grid into several superpixels, each representing a graph vertex. Superpixel segmentation helps express the heterogeneity of ecological constraints at the regional scale and improves the computational efficiency of the GCN algorithm. The SLIC (superpixel generation) algorithm starts from a cell in the grid space, uses the ak-means algorithm to search for cells with similar ecological features, and finally forms clustered patches. The segmented superpixels have similar internal features, compact structures, and are adjacent to each other, thus reflecting the spatial heterogeneity of ecological constraints at the regional scale. Compared to pixel-based GCNs, superpixel segmentation significantly reduces the number of vertices in the GCN, effectively accelerating its training process.

[0076] Specifically, this embodiment uses an undirected graph. Define a graph structure, where, Let υ represent the set of vertices, and ε represent the set of edges. By controlling the average size of the superpixels, the number of superpixels obtained by cutting the ECD is: Assume the superpixel set is This represents the q-th superpixel. S represents q The t-th pixel in N qS represents q Number of pixels in The set of vertices in the graph can then be represented as follows:

[0077]

[0078] The feature of vertex V is characterized by the average value of the pixels in E it contains. i Let V be the eigenvector of the i-th vertex, V be the matrix form of vertex υ, and the superscript T is the transpose of the matrix.

[0079] Establishing the correspondence between the original input and superpixels is fundamental to calculating superpixel feature values. The output of the SLIC algorithm is similar to a label classification result, where pixels belonging to the same superpixel have the same label value. Therefore, an association matrix Z∈R can be constructed. hw×L This expresses the mapping relationship between pixels and superpixels. If the i-th pixel belongs to the j-th superpixel, then Z... i,j =1, otherwise Z i,j =0, then the input E of the ecological factor dataset can be converted into graph structure data V by the following matrix multiplication.

[0080]

[0081] Where Encode(·;Z) is the encoder that encodes the ECD into graph vertices. It is the column-normalized form of the correlation matrix Z, that is, Sum(Z :,j ) is the sum of all elements in the j-th column of matrix Z. Flatten(·) means flattening each ecological factor in ECD in the spatial dimension. After the transformation by formula (7), the ECD dataset is equivalent to encoding the result of superpixel cutting into a graph structure that can be processed by GCN.

[0082] The proposed SGCN predicts the classification probabilities of pixels from the original input and computes a loss scalar to train the model parameters. The ultimate goal of the SGCN module is to predict the ecological constraint probabilities, which can be decomposed into pixels at the superpixel level. Therefore, a decoder from the superpixel map to the original ecological space needs to be designed in the SGCN network. This decoder transforms the features of each superpixel into pixel-level features before executing the Linear(·) function (i.e., before the linear layer and after the last convolutional layer). The decoder Decode(·; Z) is defined as follows:

[0083] E′=Decode(V;Z)=Reshape(ZV) (8)

[0084] Here, E′ represents the pixel-level feature. Since D∈Rhw×L V∈R L×output_dim Then ZV∈R hw×output_dim The `Reshape(·)` function reshapes the spatial dimensions of the result to (h, w, output_dim). That is, E′∈R h×w×output_dim `output_dim` is the ecological feature dimension output by the graph convolutional sub-network. Thus, the feature representations of the graph vertices are transformed back into the ecological space by the decoder.

[0085] To represent the adjacency relationships between vertices in a graph space, it is necessary to determine whether the boundary cells of two superpixels are adjacent. Based on this, an adjacency matrix A∈R can be constructed. L×L A i,j =1 indicates superpixel S i and S j Adjacent. That is, at least two pixels in their boundary cells are directly adjacent in the horizontal, vertical, or diagonal direction. Conversely, A i,j =0 indicates S i and S j They are not adjacent. Therefore, the adjacency matrix A defines an unweighted graph space consisting of superpixels.

[0086] Inspired by traditional deep learning techniques such as convolutional neural networks, graph convolutional networks generate an embedding representation of each central vertex by fusing features from itself and its neighboring vertices during convolution. This embodiment employs a graph convolutional network based on spectral domain theory to extract graph node features. Spectral domain graph convolutional networks introduce convolution kernels from the perspective of graph signal processing to define graph convolution operations, completing pioneering work on spectral domain graph convolutional networks based on spectral graph theories such as Fourier transform. Due to the high computational cost of this network and the lack of spatial localization in the convolution kernels, researchers proposed ChebyNet, which has lower parameter complexity and spatial localization in its convolution kernels. It parameterizes the convolution kernels using Chebyshev polynomials. Subsequently, researchers simplified the definition of ChebyNet using a first-order approximation as follows:

[0087]

[0088] In the formula, s and y are the input and output vectors defined on graph G, respectively. Let I be the learnable parameters, I be the identity matrix, and D be the degree matrix of graph G. D is a diagonal matrix, where the diagonal elements in the i-th row represent the number of neighboring vertices of the i-th vertex in graph G, i.e., ... The eigenvalues ​​range from [0,2]. Repeatedly performing the convolution operation of formula (9) may cause gradient explosion or vanishing. To address this problem, researchers have designed a transformation technique, namely, in, and, This transformation is essentially a renormalization process. Meanwhile, A+I implies that the convolution operation considers neighboring vertices, including itself. To automatically learn differentiated edge weights of neighboring vertices based on vertex features at the τth convolutional layer, a replacement is designed. Adjacency weight matrix A τ as follows:

[0089]

[0090] Where H∈R L×c H is the matrix form of the features of graph vertex V, and is the network input of the SGCN module. τ-1 It is the (τ-1)th graph convolutional layer of the model. These are learnable parameters. The operation ⊙ represents element-wise multiplication between corresponding positions of matrices. The Sigmoid(·) function compresses the calculation result to the range (0,1). Therefore, the graph convolutional network defined by formula (9) can be expressed as:

[0091]

[0092] Among them, D τ It is based on A τ The degree matrix is ​​calculated, where δ represents the activation function, such as LeakyReLU(·), W τ Let H represent the learnable parameters of the τth layer. Further, based on H... τ The following is a predictable probability diagram of ecological constraints:

[0093] P SGCN =Softmax(Linear(Decode(H τ ;Z))) (12)

[0094] Where, assume H τ It is the last convolutional layer of the graph convolutional network, Decode(H) τ Z) transforms the convolutional layer from superpixel to pixel scale, Linear(·) performs a linear transformation of the acquired pixel-scale features, and Softmax(·) further compresses the pixel-scale feature dimension to the classification dimension, making the prediction probability of each category within the range of (0,1).

[0095] (3) Probabilistic Coupling Module: Coupling of ecological soft constraint probabilities at the pixel and superpixel scales.

[0096] SGCN extracts convolutional layer H τ The features of each pixel within the superpixel scale are identical. Therefore, although it ultimately outputs pixel-scale ecological constraint probabilities, the constraint probability value for each pixel belonging to the same superpixel is the same. That is, SGCN essentially obtains the superpixel-scale ecological constraint probabilities. To couple the pixel-scale and superpixel-scale ecological constraints predicted by the CNN and SGCN networks, the final soft ecological constraint probability surface is obtained using the following formula:

[0097] P ESC =P CNN ×(P SGCN ) γ (13)

[0098] Among them, P ESC It represents the constraint probability at the pixel scale within the ecological space. γ performs an exponential transformation of the constraint probability of heterogeneous ecological spaces. This value is set according to the specific application. ESC It will be integrated as a constraint mechanism into the city cellular automata model.

[0099] 5. Simulation of Spatiotemporal Evolution of Urban Morphology Based on Urban Cellular Automata (CA) Model

[0100] Macro-level land demand and its micro-level spatial allocation are two key aspects of urban CA models in simulating the spatiotemporal evolution of land use. Regarding macro-level total land demand, a Markov model is used to estimate the land use transition probability during 2006-2022 and predict the area changes of various land uses in Quanzhou City in 2038. At each step of the Markov model, the system can maintain its current state or transition from one state to another based on the transition probability. Since land use change exhibits significant Markov properties, Markov transition probabilities are well-suited for predicting future changes in land use area. Simultaneously, an improved FLUS model is used to realize the micro-level spatial allocation of land use at specific locations. The FLUS model, based on cellular automata, designs an adaptive and competitive mechanism to handle the complex competition and interactions of future land use.

[0101] Mining transformation rules is crucial for accurate simulation of urban spatial morphology (CA) at the micro-spatial level. The FLUS model utilizes a multilayer feedforward neural network (BP-ANN) algorithm to calibrate the probability of different land use classifications at specific cell locations. Therefore, ANN-CA is chosen as the benchmark model for simulating and predicting urban spatial morphology evolution in this embodiment. In recent years, the random forest algorithm has demonstrated excellent performance in mining the contributions of natural, social, and economic driving factors. Therefore, this embodiment selects RF-CA as one of the reference models for comparing simulation results. The main objective of this embodiment is to couple CNN and SGCN algorithms to automatically calibrate the ecological soft constraint mechanism of urban expansion. Traditionally, the analytic hierarchy process (AHP), as a weighted linear combination algorithm, is a classic method for aggregating multiple ecological factors. Therefore, we define ANN-CA, ANN-AHP-CA, and ANN-SGCNN-CA as urban CA models without ecological constraints, with AHP ecological constraints, and with ecological soft constraints, respectively, and conduct comparative analysis of their simulation results.

[0102] Artificial Neural Networks (ANNs) mimic human brain neurons to create machine learning network models that fit complex systems with large inputs and outputs. ANNs have proven effective in uncovering the intrinsic patterns of nonlinear geographical phenomena. The ANN-CA (Artificial Neural Network + Cellular Automata) model comprises an input layer, multiple hidden layers, and an output layer to fit the relationship between driving factors and land use change. ANNs calculate the suitability probability of different land use types distributed in cellular space using the following formula:

[0103]

[0104] Where p(i,r,t) represents the probability of land use type r occurring in pixel i at training time t, and w j,r For adaptive weights calibrated during training, net j (i,t) is the signal received by neuron j in the hidden layer from all input neurons in cell i at time t.

[0105] To express the complex neighborhood interactions between land use types within a specific spatial range, a neighborhood function is defined based on an N×N Moore neighborhood window as follows:

[0106]

[0107] in, Let the land use type r at iteration time t be the influence on the neighborhood of cell i. w is the total number of cells with land use status r in the N×N Moore neighborhood during the (t-1)th iteration. r It represents the neighborhood effect weights for different land use types.

[0108] Furthermore, Let represent the adaptive inertia coefficient of land type r at iteration time t. This coefficient is adaptively adjusted in each iteration by calculating the difference between the current area of ​​each land type and their target demand. Considering factors such as land use suitability, neighborhood function, and inertia coefficient, the probability of cell transformation is calculated using the following formula.

[0109]

[0110] in, P is the overall probability that cell i transforms into land class r at iteration time t. i,k C represents the suitability of land type k for development in cell i. k→r This represents the transformation cost from the original land class k to the target land class r.

[0111] In this embodiment, the product of the probability of ecological soft constraint and the probability of cell transformation is used as the final probability that each pixel in the area under study will be developed into urban land, so as to express the limitation of ecological soft constraint on urban expansion. The specific formula is as follows:

[0112]

[0113] in, Let represent the final probability that cell i is developed into urban land u at iteration time t. This represents the overall probability (i.e., the probability of cell transformation occurring) of cell i transforming into urban land use u at iteration time t.

[0114] 6. Evaluation Indicators (FoM / FoM urban)

[0115] OA and Kappa are widely used pixel-level accuracy evaluation methods in existing technologies. Pixel-level accuracy verification uses a pixel-by-pixel comparison between simulated and actual maps to evaluate the model's simulation performance. By overlaying the simulation results with the actual urban morphology and performing point-to-point comparisons, a confusion matrix can be generated, and the OA and Kappa accuracy indices are calculated based on this matrix. To further evaluate the differences between the simulation results and land use change, it is beneficial to add an initial land use map during the overlay analysis. This embodiment uses three indices—OA, Kappa, and FoM—to evaluate the simulation performance of the proposed model from different perspectives. The formula for calculating the FoM index is as follows:

[0116]

[0117] Among them, A, B, C, and D are four indicators generated by overlaying the initial year, the final year, and the land use map predicted by the model. A is the number of pixels that actually changed during the simulation period but remained unchanged in the prediction result; B is the number of pixels that actually changed during the simulation period and whose prediction result also changed to the correct category; C is the number of pixels whose actual changes during the simulation period and the change category in the prediction result are inconsistent; and D is the number of pixels that did not actually change during the simulation period but whose land use change was predicted.

[0118] Imbalanced class distribution is prevalent in image segmentation. Traditional accuracy metrics are incomparable or distorted in evaluating the effectiveness of AI-based image segmentation algorithms. Based on extensive research into multi-source data, multi-class schemes, and multi-model methods, researchers have proposed a more efficient evaluation metric: the Map-Level Image Classification Efficiency Index (MICE) to assess the performance of image segmentation algorithms. Drawing on this efficiency-based evaluation method, a recall-based efficiency index for newly added urban land use is further used to evaluate the simulation performance of the proposed model in urban growth dynamics. The calculation formula is as follows:

[0119]

[0120] Where h is the number of correctly predicted pixels in the simulated new urban land, n′ is the total number of pixels in the newly developed urban land during the simulation period, and n is the number of non-urban pixels in the initial land use map, i.e., the number of pixels that may potentially be converted into urban land.

[0121] 7. Conclusion

[0122] To more clearly depict the constraints of complex natural ecological environments and their ecosystem services on urban development, this embodiment develops a method for obtaining ecological constraint rules for urban cellular automata models using deep learning technology. The main contributions are as follows:

[0123] (1) The design of ecological soft constraints is integrated into the calibration simulation process of urban cellular automata, and its parameters are dynamically corrected based on ecological factors such as the importance of ecosystem services.

[0124] (2) A pixel-level feature extraction module for ecological soft constraints was designed based on a convolutional neural network (CNN). Convolution operations, activation functions, and dropout operations effectively emphasize the influence of the neighborhood ecological space on whether the central pixel develops.

[0125] (3) A superpixel-level extraction module for ecological features, SGCN, was designed using the SLIC algorithm and Graph Convolutional Neural Network (GCN). SGCN models medium- and long-range relationships in the graph topology data structure based on superpixel segmentation. Simultaneously, SGCN automatically updates edge weights and fuses ecological constraint information from neighboring vertices at each graph convolutional layer. By integrating the high-dimensional ecological features extracted by CNN and SGCN modules, the proposed ANN-SGCNN-CA model effectively improves the performance of urban expansion evolution simulation.

[0126] Example 2:

[0127] The present invention also provides an ecological constraint calibration terminal device based on CNN and GCN, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the method embodiment described above in Embodiment 1 of the present invention.

[0128] Furthermore, as an executable solution, the CNN / GCN-based ecological constraint calibration terminal device can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The CNN / GCN-based ecological constraint calibration terminal device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above-described structure of the CNN / GCN-based ecological constraint calibration terminal device is merely an example and does not constitute a limitation on the CNN / GCN-based ecological constraint calibration terminal device. It may include more or fewer components, or combine certain components, or different components. For example, the CNN / GCN-based ecological constraint calibration terminal device may also include input / output devices, network access devices, buses, etc., and this embodiment of the invention does not limit this.

[0129] Furthermore, as an executable solution, the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices. The general-purpose processor can be a microprocessor or any conventional processor. This processor serves as the control center of the CNN / GCN-based ecological constraint calibration terminal device, connecting all parts of the device via various interfaces and lines.

[0130] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the CNN / GCN-based ecological constraint calibration terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application required for a function; the data storage area may store data created based on the use of the mobile phone. In addition, the memory may include high-speed random access memory and non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0131] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in the embodiments of the present invention.

[0132] If the modules / units integrated in the CNN / GCN-based ecological constraint calibration terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), and a software distribution medium, etc.

[0133] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. An ecological constraint calibration method based on CNN, GCN, characterized in that, Includes the following steps: S1: Construct an ecological soft constraint model to predict the probability of each pixel in the study area being developed into urban land. The ecological soft constraint model includes a convolutional neural network module, an improved graph convolutional neural network module, and a probability coupling module; The model input data is set as the topographic and ecological factors of the area to be studied; After receiving the input data from the model, the convolutional neural network module outputs the first predicted probability that each pixel will be developed into urban land. The improved graph convolutional neural network module first divides the model input data into several superpixels using a superpixel generation algorithm, and then inputs the superpixels as graph vertices into the traditional graph convolutional neural network. Next, a decoder is added before the linear layer of the traditional graph convolutional neural network to convert the superpixel-scale features output by the convolutional layer into pixel-scale features. Finally, the output of the traditional graph convolutional neural network is used as the second prediction probability that each pixel is developed into urban land. The probability coupling module is used to couple the first prediction probability and the second prediction probability into an ecological soft constraint probability; S2: Based on the topographic factors, ecological factors and driving factors of the area under study, the probability of cell conversion of each pixel being developed into urban land is obtained through the urban cellular automata model; S3: The product of the probability of cell transformation and the probability of ecological soft constraint is the final probability that each pixel in the area under study will be developed into urban land.

2. The CNN, GCN-based ecological constraint calibration method according to claim 1, characterized in that: Topographic factors include elevation and slope, while ecological factors include: importance of water conservation services, importance of soil and water conservation services, importance of biodiversity conservation services, soil fertility, and NDVI.

3. The CNN, GCN-based ecological constraint calibration method according to claim 1, characterized in that: The driving factors are distance-based variables, including: distance from the city center, distance from the town center, distance from the highway, distance from the railway, distance from the main road, distance from the train station, and distance from the coastline.

4. The CNN, GCN-based ecological constraint calibration method according to claim 1, characterized in that: The formula for calculating the probability of ecological soft constraints is: P ESC = P CNN × (P SGCN ) γ Among them, P ESC P represents the probability of ecological soft constraints. CNN Let P represent the first predicted probability. SGCN γ represents the second prediction probability, and γ represents the transformation exponent.

5. The CNN, GCN-based ecological constraint calibration method according to claim 1, characterized in that: When a superpixel is used as a graph vertex, the features of the vertex are characterized by the average value of all pixels contained in the corresponding superpixel.

6. An ecological constraint calibration terminal device based on CNN, GCN, characterized by: It includes a processor, a memory, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 5.

7. A computer readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 5.