Urban development boundary identification method based on deep learning

Abstract

The invention discloses a town development boundary identification method based on deep learning, and relates to the technical field of town development, and the method comprises the steps: employing a CNN (convolutional neural network) to extract topographic features and ecological sensitive region information in a remote sensing image, employing a geographic detector to quantify the interaction relation between spatial heterogeneity and factors, and obtaining a town development boundary identification result; according to the method, key driving factors are screened out, reasonable weights are given, in addition, the extension trend of an ecological sensitive area is dynamically predicted through time sequence modeling in combination with a long and short term memory (LSTM) network, and future urban extension is simulated in combination with a neighborhood rule and an extension probability function of MCCA, so that the dynamic adaptability of the development boundary recognition method is enhanced, and the development efficiency is improved. Meanwhile, dynamic balance optimization among ecological protection, economic benefits and land utilization efficiency is realized by adopting a GAN (Generative Adversarial Network) through adversarial training of a generator and a discriminator, the rigid constraint of ecological protection and the elastic requirements of economic development are met, and the limitation of static analysis in a traditional method is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of urban development, and in particular to an urban development boundary recognition method based on deep learning. Background Art

[0002] Mountain cities have unique complex terrain and ecological sensitivity. Their terrain conditions with significant elevation differences, steep slopes, fault zones and soil and water conservation requirements make the delineation of development boundaries complicated. At the same time, the rapid urbanization process has exacerbated the contradiction between land supply and demand. The limited construction land conflicts with the rigid requirements of ecological protection. In mountain cities, the delineation of development boundaries is not only a technical issue of spatial planning, but also a balance issue between ecological security and social and economic development.

[0003] Existing methods for identifying urban development boundaries mainly rely on technical means such as geographic information system (GIS) overlay analysis, land suitability evaluation and cellular automaton simulation to define boundaries. However, such schemes are mostly based on the hypothetical framework of plain cities and have insufficient response capabilities to the terrain characteristics of mountainous cities. Mountainous cities have differences in terrain slope and altitude. Such differences will affect the suitability of construction land and easily lead to low adaptability between the development area in the planning results and the actual terrain conditions.

[0004] In order to meet the needs of mountain city development and construction, some traditional solutions introduce terrain factors and ecological constraints for adaptive adjustment, and use slope and fault zone distribution as key factors for land suitability evaluation, which reduces the development of high-risk areas to a certain extent; however, such solutions are not sufficiently responsive to dynamically changing terrain. Over time, mountain stability, vegetation coverage, and hydrological conditions will change. The traditional terrain factor superposition analysis lacks systematicity and cannot effectively capture the interaction between terrain, ecology, and socio-economy, resulting in insufficient practical applicability of development boundaries. Therefore, a deep learning-based urban development boundary identification method is urgently needed to solve such problems. Summary of the invention

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

[0006] The present invention provides a method for identifying urban development boundaries based on deep learning to solve the problems of shortcomings of existing boundary planning schemes in dynamic adaptability, multi-factor interaction analysis and geological risk response, low accuracy, insufficient sustainability and poor dynamic adaptability in practical applications.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: The embodiment of the present invention provides a method for identifying urban development boundaries based on deep learning, which comprises: Step S1, collecting development data of mountainous cities, including remote sensing images, digital elevation models (DEMs), land use status, and socio-economic indicator data; Step S2, import the development data into the geographic detector, analyze the spatial heterogeneity and interaction of different factors, evaluate the contribution rate of each factor through the q value, and screen out the key driving factors and their weights; Step S3, input the driving factors and their weights into the hybrid cellular automaton model MCCA; Step S4, using the preliminary boundary combined with the deep learning generative adversarial network GAN to simulate the dynamic balance process of ecological protection, economic benefits and land use efficiency. GAN optimizes the rigid development boundary and the elastic development boundary through the generation and discrimination mechanism, and outputs the final development boundary; Step S5, based on the development data of step S1, calibrate the MCCA of step S3 and the GAN of step S4, update the neighborhood rules, driving factor weights and expansion probability functions, and re-simulate the development boundaries to generate real-time rigid and elastic boundaries.

[0008] As a preferred solution of the urban development boundary identification method based on deep learning described in the present invention, in which: in step S1, the remote sensing image data uses a convolutional neural network to extract features to identify mountainous terrain features and ecologically sensitive areas.

[0009] As a preferred solution of the urban development boundary identification method based on deep learning described in the present invention, the step of using a convolutional neural network to extract features and identify mountain terrain features and ecologically sensitive areas is as follows: Assume that the image data is a three-dimensional tensor , ,in, is the height of the image, is the width of the image, is the number of channels of the image, Indicates the numerical range of spatial raster data. Normalize, the normalization formula is: , in, is the normalized pixel value, is the pixel position of the corresponding image, is the spatial coordinate, is the channel index, are the minimum and maximum values ​​of the image, Perform convolution operation, the convolution formula is: , in, The convolution output feature map is at position and The value of the feature map, is the convolution kernel at position and Channel and The weight of the kernel, For the The bias of the convolution kernel, is the height and width of the convolution kernel, is the number of input channels, The convolution output is activated using the nonlinear activation function ReLU. The activation formula is: , in, is the feature map value after activation, is the spatial position, is the feature channel index, Perform pooling operation, the pooling formula is: , in, is the feature value after pooling, is the area covered by the pooling window, For classification, flatten the pooled feature map and map it to the classification output space. The mapping formula is: , in, The fully connected layer The output of a neuron, is the weight of the fully connected layer, is the flattened value of the pooling layer, is the bias, The classification formula is: , in, is the classification probability, is the total number of classification categories, For Category The exponential activation value of For the neural network to classify The activation value is based on the classification probability , extract terrain feature matrix and Ecologically Sensitive Areas Matrix , Represents the sum of exponentially scaled activations of all classes.

[0010] As a preferred solution of the urban development boundary identification method based on deep learning described in the present invention, wherein: the geographic detector is a statistical method based on spatial heterogeneity analysis, which is used to quantify the contribution rate of driving factors to target variables and the interaction between factors; The steps of importing the development data into the geographic detector, analyzing the spatial heterogeneity and interaction of different factors, evaluating the contribution rate of each factor by the q value, and screening out the key driving factors and their weights are as follows: Assume that the development data outputted in step S1 is the standardized raster data, including factors and a target variable, defined as a raster dataset ,in, For the The spatial distribution matrix of the factors is , Indicates the numerical range of spatial raster data. is the target variable matrix, are the number of rows and columns of the grid, The target variable The spatial region is divided into sub-regions, and generate a region set based on the distribution of its values. The partitioning formula is: , in, For the space area, is the total number of sub-regions, The geographic detector evaluates the spatial heterogeneity contribution rate of each factor through the q value. The evaluation formula is: , in, Factor For the target variable The contribution rate is in the range of [0,1], For sub-region The number of pixels, is the total number of pixels, For sub-region Internal target variable The variance of The target variable The global variance of , indicating factor Completely explained If the spatial distribution of , indicating factor right The distribution of has no contribution.

[0011] As a preferred solution of the urban development boundary identification method based on deep learning described in the present invention, the step of analyzing the spatial heterogeneity and interaction relationship of different factors also includes: For factors and The interaction q value is defined as: , in, Factor and The interactive contribution rate, Interaction Area The number of pixels, Interaction Area Internal target variable The variance of The types of interaction results are: if , indicating interaction enhancement, if , indicating independence, if , indicating that the interaction is weakened, according to And the interactive results , select contribution rate Factors greater than a certain threshold, and factor pairs with significant enhanced interactions , the key factor set is defined as : , in, is the key factor set, is the threshold of contribution rate and interactive contribution rate, The selected key factor set , according to its q value, the weight is normalized and calculated. The calculation formula is: , in, Factor The normalized weight of is the total contribution rate of the key factors.

[0012] As a preferred solution of the urban development boundary identification method based on deep learning described in the present invention, in which: in step S3, a spatial prediction model based on a long short-term memory network LSTM is used to dynamically model the expansion trend of ecologically sensitive areas, correct the neighborhood rules and expansion probability functions of MCCA, and generate preliminary boundaries under different scenarios.

[0013] As a preferred solution of the urban development boundary identification method based on deep learning described in the present invention, the step of dynamically modeling the expansion trend of ecologically sensitive areas using a spatial prediction model based on a long short-term memory network LSTM is as follows: Extract the ecologically sensitive area matrix from step S1 The historical sequence is divided into training set and test set: ,in, For time The ecologically sensitive area matrix at the moment, is the prediction time window, is the time step, Will The sequence is reconstructed into the input and target of the LSTM model. The reconstruction formula is: , in, is the input sequence, is the target prediction value, Use mean square error MSE as the loss function and define the loss function as: , in, is the loss value, is the sample size, is the prediction output of LSTM, Use the trained LSTM model to predict the future The expansion trend of ecologically sensitive areas in 2018 is predicted by the following formula: ,in, For the future Matrix of ecologically sensitive areas at all times; The steps of correcting the neighborhood rules and expanding the probability function of the MCCA to generate preliminary boundaries under different scenarios include: According to the dynamic prediction of ecologically sensitive areas, the neighborhood rule of MCCA is corrected. Let the neighborhood rule of MCCA be : ,in, For cells The neighborhood influence value of is the dynamically adjusted neighborhood weight, which is determined by the predicted Regulation, The expansion trend of ecologically sensitive areas is embedded in the expansion probability function of MCCA, and the embedding formula is: , in, For cells The expansion probability of are model parameters, is the predicted ecological sensitivity value, By simulating cell expansion through extended probability functions, preliminary development boundaries under different scenarios are generated .

[0014] As a preferred solution of the urban development boundary identification method based on deep learning described in the present invention, wherein: the rigid development boundary refers to the ecological area boundary within the ecological protection red line; The flexible development boundary refers to a boundary that can be adjusted to meet development needs.

[0015] As a preferred solution of the urban development boundary identification method based on deep learning described in the present invention, wherein: the dynamic balance process of simulating ecological protection, economic benefits and land use efficiency, GAN optimizes the rigid development boundary and the elastic development boundary through the generation and discrimination mechanism, and the step of outputting the final development boundary is, The preliminary boundary generated in step S3 and its related spatial feature matrix as the input of GAN, the input feature tensor is , ,in, is the preliminary boundary matrix, is the ecological protection area matrix, is the economic benefit matrix, is the land use efficiency matrix, Indicates the numerical range of spatial raster data. Generator Receive input feature tensor , output the optimized development boundary matrix , ,in, is the generator model, are the parameters of the generator, The core of the generator is a multi-layer convolutional network, and the convolution formula is expressed as: , in, For the generator The feature map of the layer, is the convolution kernel weight, is the bias, is the activation function, is the convolution operation, Discriminator Development boundaries that receive generator output and the true boundary label , to determine whether the generated boundary meets the actual ecological protection and development needs. The judgment formula is: ,in, is the discriminator model, are the parameters of the discriminator, is the binary label matrix of the true boundary, 1 represents the developable area and 0 represents the undevelopable area; The generator and discriminator optimize the boundary generation effect through adversarial training, and the loss function includes generator loss and discriminator loss. Generator loss: , in, is the loss of the generator, Generate boundaries for the discriminator pair Rating, represents an expected value operation, Discriminator loss: , in, is the loss of the discriminator, is the true label, Total loss function: , in, is the balance factor, The parameters of the generator and discriminator are optimized by stochastic gradient descent SGD. The optimization formula is: , in, is the learning rate. After training, the optimized development boundary is generated. , output the rigid development boundary and the elastic development boundary respectively, Rigid Development Boundaries: , rigid development boundary It is the initial development boundary and ecological protection red line The intersection of Flexible development boundaries: , flexible development boundaries It is the area from the preliminary development boundary minus the rigid development boundary.

[0016] As a preferred solution of the method for identifying urban development boundaries based on deep learning described in the present invention, the steps of calibrating the MCCA of step S3 and the GAN of step S4, updating the neighborhood rules, driving factor weights and expansion probability functions, and resimulating the development boundaries to generate real-time rigid and elastic boundaries are as follows: The development data collected in real time from step S1 includes: ,in, The latest ecological protection area matrix, For the latest economic benefit matrix, is the latest land efficiency matrix, is the latest population distribution matrix, Real-time ecological protection data Using the neighborhood rule to correct the MCCA model, the correction formula is: , in, For cells The neighborhood influence value of The neighborhood weights are dynamically adjusted based on the latest ecological protection data. renew, is the state of the neighborhood cells, are the height and width of the neighborhood window, According to the latest development data , recalculate the driving factor weights , based on the analysis results of the geographic detector, recalculate the contribution rate And normalized, the formula is: , in, is the real-time contribution rate, which is recalculated based on the method in step S2. Factor The real-time weight of Neighborhood influence value and the calibrated weights Incorporating the extended probability function of MCCA, the incorporation formula is: , in, For cells The expansion probability of is a parameter that controls the sensitivity of the expansion function. Factor In the cell The value of Perform GAN model calibration, including: Generator calibration: incorporating recent data , retrain the generator Parameters , the calibration formula is: ,in, is the calibrated development boundary, Discriminator Calibration: Discriminator Parameters Optimize the real boundary through the latest data update and generate boundaries The calibration formula is: , Adversarial training: Optimizing the generator and discriminator through adversarial loss functions, ,in, are the loss functions of the generator and the discriminator respectively, is the balance factor; Re-model the development boundary.

[0017] The beneficial effects of the present invention are as follows: the present invention adopts a convolutional neural network (CNN) to extract terrain features and ecologically sensitive area information from remote sensing images; secondly, a geographic detector is used to quantify spatial heterogeneity and the interaction between factors, and key driving factors are screened out and given reasonable weights; in addition, a long short-term memory network (LSTM) is combined with time series modeling to dynamically predict the expansion trend of ecologically sensitive areas, and combined with the neighborhood rule and expansion probability function of MCCA to simulate future urban expansion, thereby enhancing the dynamic adaptability of the development boundary identification method; at the same time, a generative adversarial network (GAN) is used to achieve dynamic balance optimization between ecological protection, economic benefits and land use efficiency through adversarial training of generators and discriminators, thereby meeting the rigid constraints of ecological protection and the elastic requirements of economic development, and solving the limitations of static analysis in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0019] Figure 1 Flowchart of the boundary identification method for town development of the present invention. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0021] 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.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, reference Figure 1 , this embodiment provides a method for identifying urban development boundaries based on deep learning, comprising the following steps: Step S1, collecting development data of mountainous cities, including remote sensing images, digital elevation models (DEMs), land use status, and socio-economic indicator data; In step S1, convolutional neural network is used to extract features from remote sensing image data to identify mountain terrain features and ecologically sensitive areas; The steps of using convolutional neural network for feature extraction and identification of mountain terrain features and ecologically sensitive areas are as follows: Assume that the image data is a three-dimensional tensor , ,in, is the height of the image, is the width of the image, is the number of channels of the image, Indicates the numerical range of spatial raster data. Normalize, the normalization formula is: , in, is the normalized pixel value, is the pixel position of the corresponding image, is the spatial coordinate, is the channel index, are the minimum and maximum values ​​of the image, Perform convolution operation, the convolution formula is: , in, The convolution output feature map is at position and The value of the feature map, is the convolution kernel at position and Channel and The weight of the kernel, For the The bias of the convolution kernel, is the height and width of the convolution kernel, is the number of input channels, The convolution output is activated using the nonlinear activation function ReLU. The activation formula is: , in, is the feature map value after activation, is the spatial position, is the feature channel index, Perform pooling operation, the pooling formula is: , in, is the feature value after pooling, is the area covered by the pooling window, For classification, flatten the pooled feature map and map it to the classification output space. The mapping formula is: , in, The fully connected layer The output of a neuron, is the weight of the fully connected layer, is the flattened value of the pooling layer, is the bias, The classification formula is: , in, is the classification probability, is the total number of classification categories, For Category The exponential activation value of For the neural network to classify The activation value is based on the classification probability , extract terrain feature matrix and Ecologically Sensitive Areas Matrix , represents the sum of the exponential activation values ​​of all categories; Specifically, in this step, dynamic convolution kernels are used to adjust weights to improve the ability to distinguish between terrain and ecological features and output terrain and ecological matrices.

[0024] Step S2, import the development data into the geographic detector, analyze the spatial heterogeneity and interaction of different factors, evaluate the contribution rate of each factor through the q value, and screen out the key driving factors and their weights; Geographic detector is a statistical method based on spatial heterogeneity analysis, which is used to quantify the contribution of driving factors to target variables and the interactions between factors.

[0025] The steps of importing development data into geographic detectors, analyzing the spatial heterogeneity and interaction of different factors, evaluating the contribution rate of each factor by q value, and screening out key driving factors and their weights are as follows: Assume that the development data outputted in step S1 is the standardized raster data, including factors and a target variable, defined as a raster dataset ,in, For the The spatial distribution matrix of the factors is , Indicates the numerical range of spatial raster data. is the target variable matrix, are the number of rows and columns of the grid, The target variable The spatial region is divided into sub-regions, and generate a region set based on the distribution of its values. The partitioning formula is: , in, For the space area, is the total number of sub-regions, The geographic detector evaluates the spatial heterogeneity contribution rate of each factor through the q value. The evaluation formula is: , in, Factor For the target variable The contribution rate is in the range of [0,1], For sub-region The number of pixels, is the total number of pixels, For sub-region Internal target variable The variance of The target variable The global variance of , indicating factor Completely explained If the spatial distribution of , indicating factor right The distribution of has no contribution, The steps to analyze the spatial heterogeneity and interaction of different factors also include: For factors and The interaction q value is defined as: , in, Factor and The interactive contribution rate, Interaction Area The number of pixels, Interaction Area Internal target variable The variance of The types of interaction results are: if , indicating interaction enhancement, if , indicating independence, if , indicating that the interaction is weakened, according to And the interactive results , select contribution rate Factors greater than a certain threshold, and factor pairs with significant enhanced interactions , the key factor set is defined as : , in, is the key factor set, is the threshold of contribution rate and interactive contribution rate, The selected key factor set , according to its q value, the weight is normalized and calculated. The calculation formula is: , in, Factor The normalized weight of is the total contribution rate of the key factors; Specifically, this step introduces the interaction , identify the synergistic relationship between factors, and dynamically adjust the threshold , adapt to the factors selection requirements of complex terrain, and based on Automatically normalize and generate weights; calculate key driving factors and their weights.

[0026] Step S3, input the driving factors and their weights into the hybrid cellular automaton model MCCA; In step S3, a spatial prediction model based on long short-term memory network (LSTM) is used to dynamically model the expansion trend of ecologically sensitive areas, calibrate the neighborhood rule and expansion probability function of MCCA, and generate preliminary boundaries under different scenarios; The steps of dynamically modeling the expansion trend of ecologically sensitive areas using a spatial prediction model based on long short-term memory network (LSTM) are as follows: Extract the ecologically sensitive area matrix from step S1 The historical sequence is divided into training set and test set: ,in, For time The ecologically sensitive area matrix at the moment, is the prediction time window, is the time step, Will The sequence is reconstructed into the input and target of the LSTM model. The reconstruction formula is: , in, is the input sequence, is the target prediction value, Use mean square error MSE as the loss function and define the loss function as: , in, is the loss value, is the sample size, is the prediction output of LSTM, Use the trained LSTM model to predict the future The expansion trend of ecologically sensitive areas in 2018 is predicted by the following formula: ,in, For the future Matrix of ecologically sensitive areas at all times; The steps to calibrate the neighborhood rule and expand the probability function of MCCA and generate preliminary boundaries under different scenarios include: According to the dynamic prediction of ecologically sensitive areas, the neighborhood rule of MCCA is corrected. Let the neighborhood rule of MCCA be : ,in, For cells The neighborhood influence value of is the dynamically adjusted neighborhood weight, which is determined by the predicted Regulation, The expansion trend of ecologically sensitive areas is embedded in the expansion probability function of MCCA, and the embedding formula is: , in, For cells The expansion probability of are model parameters, is the predicted ecological sensitivity value, By simulating cell expansion through extended probability functions, preliminary development boundaries under different scenarios are generated ; Specifically, LSTM is used to capture the expansion trend of ecologically sensitive areas and provide time series forecasts, and then the expansion trend is directly integrated into the neighborhood rules and expansion probability functions of MCCA to enhance the model's responsiveness to future changes and generate adaptable preliminary boundaries according to different forecast scenarios.

[0027] Step S4, using the preliminary boundary combined with the deep learning generative adversarial network GAN to simulate the dynamic balance process of ecological protection, economic benefits and land use efficiency. GAN optimizes the rigid development boundary and the elastic development boundary through the generation and discrimination mechanism, and outputs the final development boundary; Rigid development boundaries refer to the boundaries of ecological regions within the ecological protection red line; A flexible development boundary refers to a boundary that can be adjusted to meet development needs; Simulating the dynamic balance process of ecological protection, economic benefits and land use efficiency, GAN optimizes the rigid development boundary and the elastic development boundary through the generation and discrimination mechanism, and the steps of outputting the final development boundary are as follows: The preliminary boundary generated in step S3 and its related spatial feature matrix as the input of GAN, the input feature tensor is , ,in, is the preliminary boundary matrix, is the ecological protection area matrix, is the economic benefit matrix, is the land use efficiency matrix, Indicates the numerical range of spatial raster data. Generator Receive input feature tensor , output the optimized development boundary matrix , ,in, is the generator model, are the parameters of the generator, The core of the generator is a multi-layer convolutional network, and the convolution formula is expressed as: , in, For the generator The feature map of the layer, is the convolution kernel weight, is the bias, is the activation function, is the convolution operation, Discriminator Development boundaries that receive generator output and the true boundary label , to determine whether the generated boundary meets the actual ecological protection and development needs. The judgment formula is: ,in, is the discriminator model, are the parameters of the discriminator, is the binary label matrix of the true boundary, 1 represents the developable area and 0 represents the undevelopable area; The generator and discriminator optimize the boundary generation effect through adversarial training. The loss function includes generator loss and discriminator loss. Generator loss: , in, is the loss of the generator, Generate boundaries for the discriminator pair Rating, represents an expected value operation, Discriminator loss: , in, is the loss of the discriminator, is the true label, Total loss function: , in, is the balance factor, The parameters of the generator and discriminator are optimized by stochastic gradient descent SGD. The optimization formula is: , in, is the learning rate. After training, the optimized development boundary is generated. , output the rigid development boundary and the elastic development boundary respectively, Rigid Development Boundaries: , rigid development boundary It is the initial development boundary and ecological protection red line The intersection of Flexible development boundaries: , flexible development boundaries It is the area from the preliminary development boundary minus the rigid development boundary; Specifically, through GAN's comprehensive simulation of ecological protection, economic benefits and land use efficiency, a dynamic balance of the development boundary is achieved. Here, the adversarial training of the generator and the discriminator improves the scientificity and reliability of boundary demarcation.

[0028] Step S5, based on the development data of step S1, calibrate the MCCA of step S3 and the GAN of step S4, update the neighborhood rules, driving factor weights and expansion probability functions, and re-simulate the development boundaries to generate real-time rigid and elastic boundaries; The steps of calibrating the MCCA of step S3 and the GAN of step S4, updating the neighborhood rules, driving factor weights and expansion probability functions, and resimulating the development boundary to generate real-time rigid and elastic boundaries are as follows: The development data collected in real time from step S1 includes: ,in, The latest ecological protection area matrix, For the latest economic benefit matrix, is the latest land efficiency matrix, is the latest population distribution matrix, Real-time ecological protection data Using the neighborhood rule to correct the MCCA model, the correction formula is: , in, For cells The neighborhood influence value of The neighborhood weights are dynamically adjusted based on the latest ecological protection data. renew, is the state of the neighborhood cells, are the height and width of the neighborhood window, According to the latest development data , recalculate the driving factor weights , based on the analysis results of the geographic detector, recalculate the contribution rate And normalized, the formula is: , in, is the real-time contribution rate, which is recalculated based on the method in step S2. Factor The real-time weight of Neighborhood influence value and the calibrated weights Incorporating the extended probability function of MCCA, the incorporation formula is: , in, For cells The expansion probability of is a parameter that controls the sensitivity of the expansion function. Factor In the cell The value of Perform GAN model calibration, including: Generator calibration: incorporating recent data , retrain the generator Parameters , the calibration formula is: ,in, is the calibrated development boundary, Discriminator Calibration: Discriminator Parameters Optimize the real boundary through the latest data update and generate boundaries The calibration formula is: , Adversarial training: Optimizing the generator and discriminator through adversarial loss functions, ,in, are the loss functions of the generator and the discriminator respectively, is the balance factor; Re-simulate the development boundary: Specifically, the MCCA and GAN parameters are dynamically adjusted through real-time data to enhance the adaptability and real-time performance of the model. The scientific nature of the boundary is further improved by utilizing GAN's comprehensive optimization capabilities for ecological protection, economic benefits, and land use efficiency. Through the latest data-driven calibration, the development boundary is dynamically updated to adapt to complex environmental changes.

[0029] 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 may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying urban development boundaries based on deep learning, characterized in that: include, Step S1, collecting development data of mountainous cities, including remote sensing images, digital elevation models (DEMs), land use status, and socio-economic indicator data; Step S2, import the development data into the geographic detector, analyze the spatial heterogeneity and interaction of different factors, evaluate the contribution rate of each factor through the q value, and screen out the key driving factors and their weights; Step S3, input the driving factors and their weights into the hybrid cellular automaton model MCCA; Step S4, using the preliminary boundary combined with the deep learning generative adversarial network GAN to simulate the dynamic balance process of ecological protection, economic benefits and land use efficiency. GAN optimizes the rigid development boundary and the elastic development boundary through the generation and discrimination mechanism, and outputs the final development boundary; Step S5, based on the development data of step S1, calibrate the MCCA of step S3 and the GAN of step S4, update the neighborhood rules, driving factor weights and expansion probability functions, and re-simulate the development boundaries to generate real-time rigid and elastic boundaries.

2. The method for identifying urban development boundaries based on deep learning according to claim 1, characterized in that: In step S1, the remote sensing image data is subjected to feature extraction using a convolutional neural network to identify mountain terrain features and ecologically sensitive areas.

3. A method for identifying urban development boundaries based on deep learning as claimed in claim 2, characterized in that: The steps of using a convolutional neural network to extract features and identify mountain terrain features and ecologically sensitive areas are as follows: Assume that the image data is a three-dimensional tensor , ,in, is the height of the image, is the width of the image, is the number of channels of the image, Represents the numerical range of spatial raster data. Normalize, the normalization formula is: ， in, is the normalized pixel value, is the pixel position of the corresponding image, is the spatial coordinate, is the channel index, are the minimum and maximum values ​​of the image, Perform convolution operation, the convolution formula is: ， in, The convolution output feature map is at position and The value of the feature map, is the convolution kernel at position and Channel and The weight of the kernel, For the The bias of the convolution kernel, is the height and width of the convolution kernel, is the number of input channels, The convolution output is activated using the nonlinear activation function ReLU. The activation formula is: ， in, is the feature map value after activation, is the spatial position, is the feature channel index, Perform pooling operation, the pooling formula is: ， in, is the feature value after pooling, is the area covered by the pooling window, For classification, flatten the pooled feature map and map it to the classification output space. The mapping formula is: ， in, The fully connected layer The output of a neuron, is the weight of the fully connected layer, is the flattened value of the pooling layer, is the bias, The classification formula is: ， in, is the classification probability, is the total number of classification categories, For Category The exponential activation value of For the neural network to classify The activation value is based on the classification probability , extract terrain feature matrix and Ecologically Sensitive Areas Matrix , Represents the sum of exponentially scaled activations of all classes.

4. A method for identifying urban development boundaries based on deep learning as claimed in claim 3, characterized in that: The steps of importing the development data into the geographic detector, analyzing the spatial heterogeneity and interaction of different factors, evaluating the contribution rate of each factor by the q value, and screening out the key driving factors and their weights are as follows: Assume that the development data outputted in step S1 is the standardized raster data, including factors and a target variable, defined as a raster dataset ,in, For the The spatial distribution matrix of the factors is , Represents the numerical range of spatial raster data. is the target variable matrix, are the number of rows and columns of the grid, The target variable The spatial region is divided into sub-regions, and generate a region set based on the distribution of its values. The partitioning formula is: ， in, For the space area, is the total number of sub-regions, The geographic detector evaluates the spatial heterogeneity contribution rate of each factor through the q value. The evaluation formula is: ， in, Factor For the target variable The contribution rate is in the range of [0,1], For sub-region The number of pixels, is the total number of pixels, For sub-region Internal target variable The variance of The target variable The global variance of , indicating factor Completely explained If the spatial distribution of , indicating factor right The distribution of has no contribution.

5. The method for identifying urban development boundaries based on deep learning according to claim 4, characterized in that: The step of analyzing the spatial heterogeneity and interaction of different factors also includes: For factors and The interaction q value is defined as: ， in, Factor and The interactive contribution rate, Interaction Area The number of pixels, Interaction Area Internal target variable The variance of The types of interaction results are: if , indicating interaction enhancement, if , indicating independence, if , indicating that the interaction is weakened, according to And the interactive results , select contribution rate Factors greater than a certain threshold, and factor pairs with significant enhanced interactions , the key factor set is defined as : ， in, is the key factor set, is the threshold of contribution rate and interactive contribution rate, The selected key factor set , according to its q value, the weight is normalized and calculated. The calculation formula is: ， in, Factor The normalized weight of is the total contribution rate of the key factors.

6. A method for identifying urban development boundaries based on deep learning as claimed in claim 5, characterized in that: In step S3, a spatial prediction model based on long short-term memory network (LSTM) is used to dynamically model the expansion trend of ecologically sensitive areas, calibrate the neighborhood rule and expansion probability function of MCCA, and generate preliminary boundaries under different scenarios.

7. A method for identifying urban development boundaries based on deep learning as claimed in claim 6, characterized in that: The steps of dynamically modeling the expansion trend of ecologically sensitive areas using a spatial prediction model based on a long short-term memory network (LSTM) are as follows: Extract the ecologically sensitive area matrix from step S1 The historical sequence is divided into training set and test set: ,in, For time The ecologically sensitive area matrix at the moment, is the prediction time window, is the time step, Will The sequence is reconstructed into the input and target of the LSTM model. The reconstruction formula is: ， in, is the input sequence, is the target prediction value, Use mean square error MSE as the loss function and define the loss function as: ， in, is the loss value, is the sample size, is the prediction output of LSTM, Use the trained LSTM model to predict the future The expansion trend of ecologically sensitive areas in 2018 is predicted by the following formula: ,in, For the future Matrix of ecologically sensitive areas at all times; The steps of correcting the neighborhood rules and expanding the probability function of the MCCA to generate preliminary boundaries under different scenarios include: According to the dynamic prediction of ecologically sensitive areas, the neighborhood rule of MCCA is corrected. Let the neighborhood rule of MCCA be : ,in, For cells The neighborhood influence value of is the dynamically adjusted neighborhood weight, which is determined by the predicted Regulation, The expansion trend of ecologically sensitive areas is embedded in the expansion probability function of MCCA, and the embedding formula is: ， in, For cells The expansion probability of are model parameters, is the predicted ecological sensitivity value, By simulating cell expansion through extended probability functions, preliminary development boundaries under different scenarios are generated .

8. A method for identifying urban development boundaries based on deep learning as claimed in claim 7, characterized in that: The rigid development boundary refers to the ecological area boundary within the ecological protection red line; The flexible development boundary refers to a boundary that can be adjusted to meet development needs.

9. A method for identifying urban development boundaries based on deep learning as claimed in claim 8, characterized in that: The dynamic balance process of simulating ecological protection, economic benefits and land use efficiency is described. GAN optimizes the rigid development boundary and the elastic development boundary through the generation and discrimination mechanism, and the steps of outputting the final development boundary are as follows: The preliminary boundary generated in step S3 and its related spatial feature matrix as the input of GAN, the input feature tensor is , ,in, is the preliminary boundary matrix, is the ecological protection area matrix, is the economic benefit matrix, is the land use efficiency matrix, Represents the numerical range of spatial raster data. Generator Receive input feature tensor , output the optimized development boundary matrix , ,in, is the generator model, are the parameters of the generator, The core of the generator is a multi-layer convolutional network, and the convolution formula is expressed as: ， in, For the generator The feature map of the layer, is the convolution kernel weight, is the bias, is the activation function, is the convolution operation, Discriminator Development boundaries that receive generator output and the true boundary label , determine whether the generated boundary meets the actual ecological protection and development needs, the judgment formula is: ,in, is the discriminator model, are the parameters of the discriminator, is the binary label matrix of the true boundary, 1 represents the developable area and 0 represents the undevelopable area; The generator and discriminator optimize the boundary generation effect through adversarial training, and the loss function includes generator loss and discriminator loss. Generator loss: ， in, is the loss of the generator, Generate boundaries for the discriminator pair Ratings, represents an expected value operation, Discriminator loss: ， in, is the loss of the discriminator, is the true label, Total loss function: , in, is the balance factor, The parameters of the generator and discriminator are optimized by stochastic gradient descent SGD. The optimization formula is: ， in, is the learning rate. After training, the optimized development boundary is generated. , output the rigid development boundary and the elastic development boundary respectively, Rigid Development Boundaries: , rigid development boundary It is the initial development boundary and ecological protection red line The intersection of Flexible development boundaries: , flexible development boundaries It is the area from the preliminary development boundary minus the rigid development boundary.

10. A method for identifying urban development boundaries based on deep learning as claimed in claim 9, characterized in that: The steps of calibrating the MCCA of step S3 and the GAN of step S4, updating the neighborhood rules, driving factor weights and expansion probability functions, resimulating the development boundary, and generating real-time rigid and elastic boundaries are as follows: The development data collected in real time from step S1 includes: ,in, The latest ecological protection area matrix, For the latest economic benefit matrix, is the latest land efficiency matrix, is the latest population distribution matrix, Real-time ecological protection data Using the neighborhood rule to correct the MCCA model, the correction formula is: ， in, For cells The neighborhood influence value of The neighborhood weights are dynamically adjusted based on the latest ecological protection data. renew, is the state of the neighborhood cells, are the height and width of the neighborhood window, According to the latest development data , recalculate the driving factor weights , based on the analysis results of the geographic detector, recalculate the contribution rate And normalized, the formula is: ， in, is the real-time contribution rate, which is recalculated based on the method in step S2. Factor The real-time weight of Neighborhood influence value and the calibrated weights Incorporating the extended probability function of MCCA, the incorporation formula is: ， in, For cells The expansion probability of is a parameter that controls the sensitivity of the expansion function. Factor In the cell The value of Perform GAN model calibration, including: Generator calibration: incorporating recent data , retrain the generator Parameters , the calibration formula is: ,in, is the calibrated development boundary, Discriminator Calibration: Discriminator Parameters Optimize the real boundary through the latest data update and generate boundaries The calibration formula is: , Adversarial training: Optimizing the generator and discriminator through adversarial loss functions, ,in, are the loss functions of the generator and the discriminator respectively, is the balance factor; Re-model the development boundary.

Images