Urban expansion simulation method and system based on convolutional long short-term memory network

Abstract

The present invention discloses a method and system for simulating urban expansion based on a convolutional long short-term memory network. The method comprises: obtaining a spatial variable grid and a time series urban land grid in a target area, and constructing a data set; generating map slices based on the data set to obtain training data; introducing a channel attention mechanism and a background embedding mechanism based on a convolutional long short-term memory network to construct an expansion model; training the expansion model based on the training data to obtain a trained expansion model; predicting the slice data to be tested based on the trained expansion model to obtain an urban development probability map. The system comprises: a data acquisition module, a training data construction module, a model construction module, a model training module and a prediction module. By using the present invention, an accurate simulation of future urban expansion is achieved. The present invention can be widely used in the field of model simulation.

Description

Technical Field

[0001] The present invention relates to the field of model simulation, and in particular to a method and system for simulating urban expansion based on a convolutional long short-term memory network. Background Art

[0002] Urban expansion is a direct result of urbanization. Urban expansion has multiple effects, including ecological, economic, and social effects, and has attracted attention from fields such as urban planning and ecological environment. Simulating future urban expansion is very important because it helps professionals understand the possible urban expansion in a region in the future, and thus promotes the understanding of the region's future urban development, ecological protection, and other situations.

[0003] Geographic cellular automata (CA) are widely used to simulate future urban expansion. Cellular automata are a "bottom-up", spatiotemporally discrete dynamic model. Cellular automata focus on the local behavior of individuals in complex systems from a microscopic perspective. Cellular automata use dynamic methods to construct a bottom-up model to simulate the behavior of complex systems at the macroscopic level. Geographic cellular automata use square two-dimensional grids in grid space as basic units for simulation. In multiple iterations, the future state of each cell is determined by transformation rules and neighborhood states to simulate the urban expansion process. However, this mainstream model cannot accurately capture the spatiotemporal characteristics of the urban expansion process, and thus cannot guarantee the accuracy of urban expansion simulation. Summary of the invention

[0004] In view of this, in order to solve the technical problem that the existing urban expansion simulation method cannot accurately capture the spatiotemporal characteristics of the urban expansion process, thereby resulting in low accuracy of urban expansion simulation, the present invention proposes an urban expansion simulation method based on a convolutional long short-term memory network, the method comprising the following steps:

[0005] Obtain spatial variable rasters and time series urban land rasters in the target area and construct a dataset;

[0006] Generate map slices based on the data set to obtain training data;

[0007] Based on the convolutional long short-term memory network, the channel attention mechanism and background embedding mechanism are introduced to build an expansion model;

[0008] Training the expansion model based on the training data to obtain a trained expansion model;

[0009] Based on the trained expansion model, the slice data to be tested is predicted to obtain the urban development probability map.

[0010] Among them, the spatial variable grid indicates the location characteristics of different positions in the target area, such as slope, distance from the city center, etc.; the time series urban land grid is specifically a yearly time series urban land grid, which is composed of multiple raster data, and each raster data indicates the distribution of urban land in the target area in a year.

[0011] In some embodiments, the step of collecting self-coherent holograms and constructing a training set further includes:

[0012] Determining the urban land grid increment in the slice data to be tested according to the change trend in the training data;

[0013] Based on the urban development probability map, the urban range prediction result for the corresponding year is determined according to the urban land grid increment in the slice data to be tested.

[0014] Through this preferred step, combined with the grid increment, the city range prediction result of the final corresponding year is determined, and further data of the urban development probability map can be obtained.

[0015] In some embodiments, the expansion model includes multiple convolutional long short-term memory network units, and the convolutional long short-term memory network unit includes a channel attention module, a background embedding module and a naive memory network unit.

[0016] The channel attention module combines global maximum pooling, global average pooling and shared weight multi-layer perceptron to determine the weight of each channel; the background embedding module calculates information interaction and embeds background information by introducing an independent convolutional layer; the naive memory network unit processes the input through the input gate, forget gate and output gate to determine the output at the current moment.

[0017] Through this optimization step, taking into account the data characteristics and actual needs in the field of urban expansion simulation, the channel attention mechanism and background embedding mechanism are added to the convolutional long short-term memory network, and a convolutional long short-term memory network is constructed that can capture the differences in the importance of input channels and the long-term trend of urban expansion.

[0018] In some embodiments, the expansion model is represented as follows:

[0019] p n+1 , p n+2 , ..., p n+m =ConvLSTM(l1,l2,...,l n , spaVars)

[0020] Among them, l1, l2, ..., l n represents the urban land map slices input by the model, spaVars represents the spatial variable map slices, and p n+1 , pn+2 , ..., p n+m Represents the urban land development probability slice of the corresponding range of the model output, n represents the number of input years, m represents the number of output years, and ConvLSTM represents the use of convolutional long short-term memory network for calculation.

[0021] In some embodiments, the training of the expansion model includes:

[0022] Use GPU for training and adopt the mini-batch training strategy;

[0023] With the goal of minimizing the loss function, the corresponding model parameters are saved to obtain the trained expansion model.

[0024] Through this preferred step, a mini-batch method is adopted to improve the convergence ability of the model during the training process.

[0025] In some embodiments, the step of predicting the slice data to be tested based on the trained expansion model to obtain the urban development probability map specifically includes:

[0026] Predict the slice data to be tested based on the trained expansion model to obtain predicted slice data;

[0027] The predicted slice data is the development probability of the corresponding area and year of the slice;

[0028] The predicted slice data are spliced ​​to obtain a city development probability map.

[0029] The present invention also proposes an urban expansion simulation system based on a convolutional long short-term memory network, the system comprising:

[0030] The data acquisition module is used to obtain the spatial variable grid and time series urban land grid in the target area and construct a data set;

[0031] A training data construction module generates map slices based on the data set to obtain training data;

[0032] The model building module is based on the convolutional long short-term memory network, introduces the channel attention mechanism and background embedding mechanism, and builds an expansion model;

[0033] A model training module, which trains the expansion model based on the training data to obtain a trained expansion model;

[0034] The prediction module predicts the slice data to be tested based on the trained expansion model to obtain the urban development probability map.

[0035] Based on the above scheme, the present invention provides a method and system for simulating urban expansion using a convolutional long short-term memory network, and introduces the convolutional long short-term memory network into the field of urban expansion simulation. It is proposed to use convolution in a neural network to process urban land use data for many years, instead of a simple neighborhood density estimation method, to capture the spatiotemporal dependency of the urban expansion process, and ultimately achieve a more accurate simulation of future urban expansion; further, the channel attention mechanism and background embedding mechanism are added to the convolutional long short-term memory network (ConvLSTM), to construct an improved ConvLSTM that can capture the difference in the importance of input channels and the long-term trend of urban expansion. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of the steps of a method for simulating urban expansion based on a convolutional long short-term memory network of the present invention;

[0037] Figure 2 It is a data flow diagram of an expansion model of a specific embodiment of the present invention;

[0038] Figure 3 It is a schematic diagram of the structure of a convolutional long short-term memory network unit according to a specific embodiment of the present invention;

[0039] Figure 4 is a schematic diagram of the structure of a channel attention module in a specific embodiment of the present invention;

[0040] Figure 5 It is a schematic diagram of the structure of a background embedding module and a naive convolutional long short-term memory network unit in a specific embodiment of the present invention; Figure 5 (a) is a schematic diagram of the structure of the background embedding module; Figure 5 (b) is a schematic diagram of the structure of a naive convolutional long short-term memory network unit. DETAILED DESCRIPTION

[0041] The basic content of the present invention is to address the problem that the mainstream urban expansion simulation model cannot accurately capture the spatiotemporal characteristics of the urban expansion process, and propose an urban expansion simulation method based on an improved convolutional long short-term memory network (MConvLSTM). The present invention cuts the spatial variables and time series urban land use maps within a specific range into map slices. The training set consisting of a certain length time series urban land grid map slice and a spatial variable grid slice is input into the MConvLSTM, and it is trained to output the spatial layout of urban land at multiple continuous time points within the slice range in the future period. Then the test set is input into the trained model to obtain the model prediction result. The prediction results are then spliced ​​to obtain a probability map of urban land development within a specific range for a period of time in the future. Finally, according to the demand for urban land, the urban land probability map is converted into the future urban land range to complete the simulation of future urban expansion. The present invention has good flexibility, and the input and output time series lengths and the input spatial variable types can be determined according to actual needs.

[0042] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0043] It should be noted that, for the convenience of description, only the parts related to the relevant invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0044] It should be understood that the "system", "device", "unit" and / or "module" used in this application is a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the word can be replaced by other expressions.

[0045] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not refer to the singular, but also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements. The elements defined by the sentence "includes a..." do not exclude the existence of other identical elements in the process, method, commodity or device that includes the elements.

[0046] In the description of the embodiments of the present application, "plurality" means two or more than two. The following terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0047] In addition, flow charts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, the various steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0048] Reference Figure 1 , which is a flow chart of an optional example of the urban expansion simulation method of the convolutional long short-term memory network proposed in the present invention. The method can be applied to a computer device. The urban expansion simulation method proposed in this embodiment may include but is not limited to the following steps:

[0049] Step S1, obtaining the spatial variable grid and the time series urban land grid in the target area, and constructing a data set;

[0050] Among them, the spatial variable grid indicates the location characteristics of different locations in the target area, such as slope, distance from the city center, etc. The annual time series urban land grid is composed of multiple raster data, each of which indicates the distribution of urban land in the target area in a year.

[0051] In this embodiment, the spatial variable grid is further normalized to the maximum and minimum values ​​to generate a spatial variable grid with a value range of 0-1; the time series urban land grid is reclassified so that non-urban land has a value of 0 and urban land has a value of 1.

[0052] Step S2: Generate map slices based on the data set to obtain training data.

[0053] According to actual needs, determine the number of input years n and the number of output years m. Determine the starting year of the training set The spatial variable raster and The urban land raster of each year is determined as the training set.

[0054] According to the predetermined map grid slice size s (such as 64), all the grids in the training set are cut into map slices of size s×s. Each slice is a grid with a length and width of s pixels. Thus, several groups of slices are obtained. Each group of slices contains spatial variable grid map slices and urban land grid slices covering a square area. It should be ensured that the union of the training set clipping results completely covers the entire target area.

[0055] From the training set raster map, a certain strategy is used to extract several groups of slices as training samples. The number of extracted slice groups should be no less than 2000; if the number of training set slice groups is less than 2000, the sampling method such as flipping and rotating can be adopted to obtain at least 2000 groups of raster slices as training samples.

[0056] Step S3: Based on the convolutional long short-term memory network, a channel attention mechanism and a background embedding mechanism are introduced to construct an expansion model;

[0057] The expansion model includes multiple convolutional long short-term memory network units, and the convolutional long short-term memory network unit includes a channel attention module, a background embedding module and a naive convolutional long short-term memory network unit.

[0058] The channel attention module combines global maximum pooling, global average pooling and shared weight multi-layer perceptron to determine the weight of each channel; the background embedding module calculates information interaction and embeds background information by introducing an independent convolutional layer; the naive convolutional long short-term memory network unit processes the input through the input gate, forget gate and output gate to determine the output at the current moment.

[0059] Step S4: training the expansion model based on the training data to obtain a trained expansion model;

[0060] Input the training samples obtained in S2 into the expansion model to train the model for predicting future urban expansion. If hardware conditions permit, GPU should be used for training; the training strategy should adopt the mini-batch method to improve the convergence ability of the model during training. Trial training can be repeated many times to tune the hyperparameters such as learning rate, batch size, optimizer, batch size, etc. required for model training. After the determined hyperparameters are obtained, the final training begins. The final training should last for multiple rounds, and the model loss function of each round should be observed and recorded during the training process. When the loss function no longer decreases significantly or even starts to increase, the training can be stopped. Among all the models obtained from multiple rounds of training, the model with the smallest loss function is selected as the model training result.

[0061] Step S5: predict the slice data to be tested based on the trained expansion model to obtain a city development probability map.

[0062] The slice data to be tested is input into the trained expansion model. From the input of each set of test set raster map slices, m raster map slices of the same size of s×s can be obtained. These slices correspond to the area in Annual urban development probability.

[0063] Iteration According to the spatial position corresponding to each slice group, all urban development probability slices of the corresponding year obtained in S7 are spliced ​​to obtain the entire target area Map of urban development probabilities for each year during the year.

[0064] Step S6, determining the urban land grid increment in the slice data to be tested according to the change trend in the training data;

[0065] Using the Markov chain model or simple extrapolation method, according to The change trend of urban land use in 2018 and 2019 determines the slice data to be tested. The urban land raster increment of the target area in each year (i.e. the target year of urban expansion simulation).

[0066] Step S7: Based on the urban development probability map, according to the urban land grid increment in the slice data to be tested, determine the urban range prediction result for the corresponding year.

[0067] Iteration Each year in the year. According to the urban land increment determined in S6, a certain number of pixels with the highest development probability in the urban development probability map of that year and which are not classified as urban land in the urban land grid of the previous year are designated as newly added urban land. This number is the urban land increment of that year. Thus, the urban range prediction result of that year is obtained. The city area should be The actual urban land grid in the year; When the city scope is calculated for the year, it should be the simulation result obtained in the previous iteration.

[0068] In some feasible embodiments, the expansion model can be expressed as follows:

[0069] p n+1 ,p n+2 ,…,p n+m =ConvLSTM(l1,l2,…,l n ,spaVars)#(1)

[0070] Among them, l1,l2,…,l n represents the urban land map slices input by the model, spaVars represents the spatial variable map slices, and p n+1 ,p n+2 ,…,p n+m Represents the urban land development probability slice of the corresponding range of the model output, n represents the number of input years, m represents the number of output years, and ConvLSTM represents the use of convolutional long short-term memory network for calculation.

[0071] The convolutional long short-term memory network is composed of n+m-1 convolutional long short-term memory network units (ConvLSTM units, hereinafter referred to as units). The first n-1 units receive one year's input data in sequence, the nth unit receives the nth year's data and outputs the development probability map slice of the n+1th year, and the last m-2 units output the development probability map slice of one year in sequence. Figure 2 shown.

[0072] For the first n units, the input of the tth unit is the spatial variable map slice spaVars and the urban land map slice l in the tth year t The input matrix X is composed of t , the hidden state h output by the t-1th unit t-1 ,c t-1 (The hidden state of the first unit is a random value.) For the next m-2 units, the input to the tth unit is a slice of the spatial variable map spaVars and h t-1 The input matrix X is composed of t , the hidden state h output by the t-1th unit t-1 , c t-1 Among them, for the last m-1 units, the h output by the tth unit is t That is p t+1 .Right now:

[0073] h t , c t =Cell(X t ,h t-1 , c t-1 )#(2)

[0074]

[0075] Each unit consists of a channel attention module, a background embedding module, and a naive convolutional long short-term memory network unit. The input data is processed by these modules to obtain the output data, which is represented by formula 4-6 (where ChA, CE, and VCell represent the processing of the three modules respectively), as shown in the attached figure. Figure 3 shown.

[0076] CX t =ChA(X t )#(4)

[0077] CX′ t , h′ t-1 =CE(CX t ,h t-1 )#(5)

[0078] h t , c t=VCell(CX′ t , h′ t-1 , c t-1 )#(6)

[0079] The channel attention module combines global maximum pooling, global average pooling, and shared weight multi-layer perceptron (MLP) to determine the input X t The weight of each channel in gets the weighted input CX t . As shown in formula 7-14, channel i Represents X t The i-th channel of , max and average represent the maximum value and average value operations respectively, and ⊙ represents the Hadamard product. Figure 4 shown.

[0080] mp i =max(channel i )#(7)

[0081] ap i =average(channel i )#(8)

[0082] V max =[mp1, mp2, ..., mp k+1 ]#(9)

[0083] V avg =[ap1, ap2, ..., ap k+1 ]#(10)

[0084] W max =σ(W3σ(W2σ(W1V max +b1)+b2)+b3)#(11)

[0085] W avg =σ(W3σ(W2σ(W1V avg +b1)+b2)+b3)#(12)

[0086] weight = σ(W max +W avg )#(13)

[0087] CX t =weight⊙X t #(14)

[0088] The background embedding module introduces two independent convolutional layers Conv h , Conv X , calculate CX tThe interaction with the information provided by the previous unit results in CX′ embedded with background information t and h′ t-1 The background embedding module can improve the module's ability to predict longer time series. It is expressed by equations 15-17, where W h,conv , b h，conv It is the convolution layer Conv h Parameters, W X,conv , b x,conv It is the convolution layer Conv X Parameters. Figure 5 (a) shown.

[0089] h t-1,conv =σ(W h,conv *h t-1 +b h,conv )#(15)

[0090] CX′ t =h t-1,conv ⊙CX t #(16)

[0091] h′ t-1 =σ(W X,conv *CX′ t +b X,conv )⊙h t-1 #(17)

[0092] The naive convolutional long short-term memory network unit uses three gate control units, input gate, forget gate and output gate, to process the input and determine the output at the current moment, that is, the development probability slice at the next moment. The forget gate determines how much information from time t-1 should be forgotten, and the input gate determines CX t The output gate determines the output information. The gate operation is expressed as Equation 18-23 and the attached Figure 5 As shown in (b), σ represents the sigmoid function, * represents the convolution operation, and W and b are the parameters of the convolution.

[0093] i t =σ(W hi *h′ t-1 +W Xi *CX′ t +b i )#(18)

[0094] g t =tanh(W hg *h′ t-1 +W Xg *CX′ t +b g )#(19)

[0095] f t =σ(W hf *h′ t-1 +W Xf *CX′ t +b f )#(20)

[0096] o t =σ(W ho *h′ t-1 +W Xo *CX′ t +b o )#(twenty one)

[0097] c t =f t ⊙c t-1 +i t ⊙g t #(twenty two)

[0098] h t =o t ⊙tanh(c t )#(twenty three)

[0099] The present invention also provides the following examples:

[0100] The Guangdong-Hong Kong-Macao Greater Bay Area is taken as the research object. The Guangdong-Hong Kong-Macao Greater Bay Area covers an area of ​​about 56,000 square kilometers and has a population of more than 86 million in 2022. It is an important economic center in South my country. The data used in this study include: Global Time Series Impervious Surface Area Dataset (GISA Dataset), slope spatial variables, distance to township centers spatial variables, and distance to county centers spatial variables, as shown in the attached Figure 5 As shown. All data have a spatial resolution of 30 meters, the slope is in degrees, and the distance is in meters. This study uses the data from 2000 to 2011 as the training set, of which 2000 to 2005 is the input and 2006 to 2011 is the output; the data from 2006 to 2017 is the test set, of which 2006 to 2011 is the input and 2012 to 2017 is the output and accuracy verification year. Based on the above data, the urban range simulation results from 2012 to 2017 are obtained. Comparisons are made with the FLUS model, RF-CA model, and LSTM-CA model. Results verification, based on the real urban range of the Guangdong-Hong Kong-Macao Greater Bay Area from 2012 to 2017, the hit rate (FoM) of the simulation results is calculated. The results and comparisons are shown in Appendix 1.

[0101] Table 1 Comparison of the simulation results FoM of the Greater Bay Area obtained by this invention with those obtained by other models

[0102]

[0103] A city expansion simulation system based on convolutional long short-term memory network, comprising:

[0104] The data acquisition module is used to obtain the spatial variable grid and time series urban land grid in the target area and construct a data set;

[0105] A training data construction module generates map slices based on the data set to obtain training data;

[0106] The model building module is based on the convolutional long short-term memory network, introduces the channel attention mechanism and background embedding mechanism, and builds an expansion model;

[0107] A model training module, which trains the expansion model based on the training data to obtain a trained expansion model;

[0108] The prediction module predicts the slice data to be tested based on the trained expansion model to obtain the urban development probability map.

[0109] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0110] A city expansion simulation device based on convolutional long short-term memory network:

[0111] at least one processor;

[0112] at least one memory for storing at least one program;

[0113] When the at least one program is executed by the at least one processor, the at least one processor implements the urban expansion simulation method based on the convolutional long short-term memory network as described above.

[0114] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0115] A storage medium stores processor-executable instructions, which are used to implement the above-mentioned urban expansion simulation method based on convolutional long short-term memory network when executed by the processor.

[0116] The contents of the above method embodiments are all applicable to the present storage medium embodiments. The functions specifically implemented by the present storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0117] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for simulating urban expansion based on convolutional long short-term memory network, characterized in that: The following steps are involved: Obtain spatial variable rasters and time series urban land rasters in the target area and construct a dataset; Generate map slices based on the data set to obtain training data; Based on the convolutional long short-term memory network, the channel attention mechanism and background embedding mechanism are introduced to build an expansion model; Training the expansion model based on the training data to obtain a trained expansion model; Based on the trained expansion model, the slice data to be tested is predicted to obtain the urban development probability map; The spatial variable grid indicates the location characteristics of different positions in the target area, and the time series urban land grid indicates the distribution of urban land in the target area at the corresponding time; The expansion model includes a plurality of convolutional long short-term memory network units, and the convolutional long short-term memory network unit includes a channel attention module, a background embedding module and a naive memory network unit; The channel attention module combines global maximum pooling, global average pooling and shared weight multi-layer perceptron to determine the weight of each channel; the background embedding module calculates information interaction and embeds background information by introducing an independent convolutional layer; the naive memory network unit processes the input through the input gate, forget gate and output gate to determine the output at the current moment, that is, the development probability slice at the next moment; The expansion model is expressed as follows: p n+1 ,p n+2 ,…,p n+m =ConvLSTM(l1,l2,…,l n ,spaVars) Among them, l1, l2, ..., l n represents the urban land map slices input by the model, spaVars represents the spatial variable map slices, and p n+1 , p n+2 , ..., p n+m It represents the probability slice of urban land development in the corresponding range of the model output, n represents the number of input years, m represents the number of output years, and ConvLSTM represents the use of convolutional long short-term memory network for calculation; The step of predicting the slice data to be tested based on the trained expansion model to obtain the urban development probability map specifically includes: Predict the slice data to be tested based on the trained expansion model to obtain predicted slice data; The predicted slice data is the development probability of the corresponding area and year of the slice; splicing the predicted slice data to obtain a city development probability map; The convolutional long short-term memory network is composed of n+m-1 convolutional long short-term memory network units. The first n-1 convolutional long short-term memory network units receive one year's input data in sequence, the nth convolutional long short-term memory network unit receives the nth year's data and outputs the development probability map slice of the n+1th year, and the last m-2 units output the development probability map slice of one year in sequence; For the first n units, the input of the t-th convolutional long short-term memory network unit is a slice of the spatial variable map and a slice of the urban land use map l in the t-th year. t The input matrix X is composed of t , the hidden state h output by the t-1th convolutional long short-term memory network unit t-1 , c t-1 ; For the last m-2 units, the input of the tth convolutional LSTM network unit is composed of the spatial variable map slice and h t-1 The input matrix X is composed of t , the hidden state h output by the t-1th convolutional long short-term memory network unit t-1 , c t-1 ; Among them, for the last m-1 units, the h output by the tth unit is t That is p t+1 ;Right now: h t ,c t =Cell(X t ,h t-1 ,c t-1 ) 2. The urban expansion simulation method based on convolutional long short-term memory network according to claim 1 is characterized in that: Also includes: Determining the urban land grid increment in the slice data to be tested according to the change trend in the training data; Based on the urban development probability map, the urban range prediction result for the corresponding year is determined according to the urban land grid increment in the slice data to be tested.

3. The urban expansion simulation method based on convolutional long short-term memory network according to claim 1 is characterized in that: The training of the expansion model includes: Use GPU for training and adopt the mini-batch training strategy; With the goal of minimizing the loss function, the corresponding model parameters are saved to obtain the trained expansion model.

4. A city expansion simulation system based on convolutional long short-term memory network, characterized in that: The method for simulating urban expansion based on a convolutional long short-term memory network as claimed in claim 1 comprises: The data acquisition module is used to obtain the spatial variable grid and time series urban land grid in the target area and construct a data set; A training data construction module generates map slices based on the data set to obtain training data; The model building module is based on the convolutional long short-term memory network, introduces the channel attention mechanism and background embedding mechanism, and builds an expansion model; A model training module, which trains the expansion model based on the training data to obtain a trained expansion model; The prediction module predicts the slice data to be tested based on the trained expansion model to obtain the urban development probability map.

Images