Method and system for identifying urban functional areas based on spatial pattern characteristics of geographic objects

Abstract

This invention discloses a method and system for identifying urban functional zones based on the spatial pattern characteristics of geographic objects. First, open-source big data is normalized and a series of natural and artificially constructed separating elements are used to divide urban spatial regions into research units, namely blocks. Second, the spatial distribution and semantic attributes of multimodal data are mined based on urban surface elements within the blocks, point of interest data (POIs), and medium-resolution remote sensing imagery to obtain the socioeconomic feature vector and spatial pattern feature vector of each block. Finally, based on the obtained spatial pattern feature vectors and socioeconomic feature vectors, a convolutional neural network is designed to automatically identify and classify the spatial semantics of urban functional zones. Based on open-source big data, this invention intelligently mines the multidimensional features of various geographic objects, such as urban surface elements and point of interest data (POIs), achieving high-precision, fine-grained automatic identification and classification of the spatial semantics of urban functional zones.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing application, and in particular relates to a method and system for identifying urban functional areas based on spatial pattern characteristics of geographic objects. Background Art

[0002] Urban functional zones refer to geographical combinations of landscape elements that spatially constitute the same urban function and are semantically abstracted to represent the same urban function. They are considered the basic units of urban management and planning. The spatial pattern of urban functional zones influences the efficiency of urban life and is closely related to issues such as traffic congestion, the demarcation of "sleeping zones," and urban air pollution. Urban functional zone detection is an effective method for understanding urban space and the interactions between human activities and the environment. Numerous models for extracting and analyzing urban surface features have been developed in my country, and the driving forces behind rapid urbanization and the characteristics of the resulting functional zones have been widely discussed. Traditionally, the demarcation of urban functional zones has been primarily based on field surveys and observations. With the widespread use of high-resolution satellite imagery (Landsat, SPOT, and QuickBird), remote sensing technology can be used to extract high-precision urban land use maps, effectively describe the visual characteristics of urban features, and subsequently segment them. However, existing urban functional zone identification results based on land cover in remote sensing images still have a significant application gap with actual living spaces. Furthermore, research on the dynamic changes in urban land use classification and the complex urban-suburban transition zone is far from sufficient, making the provision of high-precision, fine-grained, and real-time urban functional zone identification a significant challenge.

[0003] Remote sensing images contain a wealth of visual and spatial information about objects. Urban functional zone identification should not focus solely on object-oriented surface feature recognition. Only by capturing the spatial arrangement and layout of objects in images and selecting the spatial aggregation patterns of geographic objects as key identification factors can we effectively change the current approach of classifying urban functional scenes based solely on the visual characteristics of objects. Furthermore, with the rapid development of artificial intelligence methods such as machine learning in high-resolution remote sensing image recognition, machine learning models can effectively quantify the spatial pattern characteristics of objects, integrating image, spatial, and semantic information, making it possible to accurately survey and map, and to update and timely plan urban functional zones.

[0004] Although urban functional zone detection methods based on geographic objects in remote sensing images have achieved certain success in academia and industry, with the acceleration of global urbanization, rapid urban expansion, and growing populations, urban functional zones are constantly changing to meet people's growing socioeconomic needs in this rapid urbanization process. Current methods that focus solely on extracting ground objects and spatial details are limited to functional zone identification in single street scenes or certain specific scenarios. They cannot yet achieve intelligent detection and classification of urban land use in complex geographic scenarios, nor can they provide automatically updated urban functional zone planning maps for sustainable urban development. Therefore, providing highly timely and accurate urban functional zone planning has become a major requirement for smart city construction. It is the basis for further capturing urban human behavior patterns and an important basis for providing traffic control, energy recovery, and resource emergency response in urban management. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a typical urban functional area identification method and system based on the spatial pattern characteristics of geographic objects. Based on open source big data, it intelligently mines the multidimensional characteristics of various geographic objects such as urban surface elements and point of interest data (POI), and can achieve high-precision and fine-grained automatic recognition and classification of urban functional area spatial semantics.

[0006] The present invention provides a method for identifying urban functional areas based on spatial pattern features of geographic objects, comprising the following steps:

[0007] Step 1: pre-process the open source geospatial vector data and divide the urban spatial area into research units, i.e., blocks;

[0008] Step 1.1: Preprocess the open source geospatial vector data to meet the data usage specifications;

[0009] Step 1.2: Use ArcGIS software to merge the multi-lane line data in the open source geospatial vector data. Based on the topology check results, connect the broken lines in the multi-lane line data to simplify the road line data.

[0010] Step 1.3: Divide the urban space using the main road-level road data and river data excluding wetlands to obtain urban blocks;

[0011] Step 2: The spatial distribution attributes and location relationships of POI objects within the block are converted into a spatial semantic corpus of Block-POI. High-dimensional continuous dense vectors of different POI types within each block are obtained, and the socioeconomic characteristic vector of each block is obtained through weighted averaging.

[0012] Step 2.1: Consider the blocks in the study area as “documents” and the POIs within the blocks as “words” to construct a Block-POI spatial semantic corpus;

[0013] Step 2.2: Use the Word2Vec model to obtain high-dimensional continuous dense vectors of different POI types in each block;

[0014] Step 2.3: Improve the true value part of the log function of the TF-IDF algorithm. Use the improved TF-IDF algorithm to weight the POI word vectors, calculate the actual contribution of different POI types in the Block-POI corpus, and explore the scale and usage characteristics of POI functions in each block.

[0015] Step 3: Taking the block as a unit, calculate the various elements, neighborhood, and overall spatial distribution structure indicators of its internal surface features based on vector data. Calculate the land cover remote sensing application index and image depth semantic features formed by the aggregation of various elements within the block based on medium-resolution remote sensing imagery. The combination of the two constitutes the spatial pattern feature vector of each block.

[0016] Step 3.1: Taking the block as the unit, calculate the indicators based on the element characteristics of the geographic object itself, the indicators based on the neighborhood characteristics of the surface elements, and the indicators based on the overall distribution characteristics of the surface elements in the block based on the internal surface element vector data;

[0017] Step 3.2: Using the minimum bounding rectangle of the block segmented in step 1 as a mask, extract the corresponding surface area in the medium-resolution remote sensing image and calculate the three surface cover remote sensing application indices for each block.

[0018] In step 3.3, the remote sensing image of the study area is segmented into input images of a certain size, and the existing deep convolutional autoencoder model MegNet is used to extract the deep semantic features of the image;

[0019] Step 3.4: Perform Z-Score normalization on the various element, neighborhood, and overall spatial distribution structure indices calculated in step 3.1 and the land cover remote sensing application index calculated in step 3.2;

[0020] Step 4: Based on the socioeconomic feature vector obtained in step 2 and the spatial pattern feature vector obtained in step 3, a convolutional neural network is used to automatically identify and classify the spatial semantics of urban functional areas.

[0021] Moreover, the pre-processing operations in step 1.1 include: ① data spatial reference conversion; ② topological check of geographic feature data; ③ semantic attribute supplementation of data with missing key attributes of POI.

[0022] Moreover, the specific operations involved in step 2.1 are as follows: ① Calculate the Euclidean distance between all POI pairs in the block, and select the POI pair with the farthest distance as the two endpoints of the word arrangement sequence in the corpus, that is, <P start ,P end >;② Taking the block as the unit, abstract the POI into the graph node, construct the unweighted and undirected Voronoi diagram, and calculate it according to the Dijkstra algorithm <P start ,P end >The shortest path between them is calculated, and "POI words" are recorded in sequence to construct a "block document" containing POI spatial context information; ③ Following the document construction process, point pairs are constructed based on block centers, and the Euclidean distance between block point pairs in the study area is calculated to obtain the "block document"; ② and ③ constitute the Block-POI spatial semantic corpus.

[0023] Moreover, in step 2.2, the CBOW model of the Word2Vec neural network is used to train the probability vector of each type of POI in the block, and its objective function L and the probability distribution p(w t |Context(w t ))The calculation method is as follows:

[0024] Context(w t )=w t-c ,…,w t-1, w t+1 ,…,w t+c (1)

[0025]

[0026] Where w t is the current word, c is the context window size, Context(w t ) represents the current word w t The context input vector, T is the corpus size, w i is any word in the corpus, E(w t ,Context(w t )) is the energy function of the Word2Vec model, which is calculated as follows:

[0027] E(w i ,w j )=-(v(w i )·v(w j )) (3)

[0028] Where, E(w i ,w j ) represents word w i Its context word wj The vector inner product, v(w i )、v(w j ) represents the word vector.

[0029] Furthermore, in step 2.3, the TF value of the POI semantic function type is first determined, and the specific calculation method is as follows:

[0030]

[0031] Where TF i,j represents the TF value of the i-th POI type in the j-th block, n i,j is the number of times the i-th POI type appears in the j-th block, ∑ k n k,j is the total number of occurrences of each type of POI in the jth block.

[0032] To address the uneven distribution problem between classes, the improved IDF is used to determine the IDF value of the POI semantic function type. The improved calculation method is as follows:

[0033]

[0034] Where, IDF i represents the IDF value of the i-th POI type, N is the total number of blocks, N m represents the number of blocks of type m, d im represents the number of blocks in the mth type of block that contain the i-th POI type, d ic,c≠m It represents the number of blocks in other categories that contain the i-th POI type. The specific calculation formula is as follows:

[0035]

[0036] Where w i Refers to the i-th POI type, d j refers to the jth block, c m Refers to the mth type of block, <w i ∈d j >∩ <d j ∈c m > represents the value judgment condition.

[0037] To solve the problem of uneven distribution within a class, we use the parameter class frequency CF i , combined with the above calculation parameters TF i,j and IDF i , the improved TF-IDF algorithm is shown in formula (8):

[0038]

[0039] Where TF-IDF i represents the weight of the i-th POI type, d im N represents the number of blocks in the mth block in the dataset that contain the i-th POI type. m represents the number of blocks of type m, N is the total number of blocks, d ic,c≠m Indicates the number of blocks containing the i-th POI type in other categories of blocks.

[0040] To avoid IDF i When a true number or a denominator of zero appears during the calculation process, equation (8) is smoothed, as shown in equation (9):

[0041]

[0042] Based on the POI semantic type feature vector and type weight obtained in steps 2.2 and 2.3, the socioeconomic characteristics of each block are calculated as follows:

[0043]

[0044] Where, v SE (block) refers to the socioeconomic characteristic vector v of the block Social-Economic (block), v(POI_Category(i)) is the vector representation of different POI types in the block obtained in step 2.2, and weight(POI_Category(i)) is the weight of different POI types obtained by the improved TF-IDF algorithm.

[0045] Moreover, in step 3.1, the information entropy principle is combined to calculate the area distribution entropy, perimeter distribution entropy, average radius distribution entropy, minimum circumscribed rectangle direction distribution entropy, compactness distribution entropy, fractal dimension distribution entropy, and concavity distribution entropy of the building objects in the block based on the element level. The density distribution entropy of the building objects is calculated based on the neighborhood level. The overall aggregation structure characteristics of the geographical objects in the block are calculated to measure the diversity of the geometric and distribution attributes of the buildings in the block. The specific calculation method is shown in Equations (11) to (30):

[0046] Area distribution entropy:

[0047]

[0048] Where H1(d) is the area distribution entropy of the building geographic objects in the block, A i is the area of the i-th building, S is the total building area, and n is the number of building geographic objects.

[0049] Perimeter distribution entropy:

[0050]

[0051] Where H2(d) is the perimeter distribution entropy of the building geographic objects in the block, P i is the perimeter of the i-th building, L is the total perimeter of buildings, and n is the number of building geographic objects.

[0052] Mean radius distribution entropy:

[0053]

[0054] Where H3(d) is the average radius distribution entropy of building geographic objects in the block, R i is the average radius of the i-th building, R is the sum of the average radius of all buildings, and n is the number of building geographic objects; the average radius of a building object is calculated by the average distance from all boundary points on its outline to its center point:

[0055]

[0056] Where, d j represents the distance from the boundary point j to the center point, and M is the number of boundary points.

[0057] Minimum circumscribed rectangle directional distribution entropy: The main direction of the minimum circumscribed rectangle of a building is defined as the angle between its long axis and the horizontal axis. First, the main direction θ of each building object needs to be calculated to determine the number of main direction categories. Then, the number of buildings in different main direction classification intervals within the block is counted to analyze the difference in building direction distribution within the block. The direction statistics results are substituted into the following formula:

[0058]

[0059] Where H4(d) is the directional distribution entropy of the building instance object level in the building group, S i is the number of buildings in the i-th direction, N is the total number of buildings, and n4 is the number of main direction categories.

[0060] Compactness distribution entropy:

[0061]

[0062] Where H5(d) is the compactness distribution entropy of the building geographic objects in the block, S_Com i is the compactness of the i-th building, S_Com is the total building compactness, and n is the number of building geographic objects. The compactness of a building object is a quadratic relationship between its area and perimeter:

[0063]

[0064] Where Ai is the area of the i-th building, P i is the perimeter of the i-th building.

[0065] Fractal dimension distribution entropy:

[0066]

[0067] Where H6(d) is the fractal dimension distribution entropy of the building geographic objects in the block, S_Frac i is the fractal dimension of the i-th building, S_Frac is the total fractal dimension of buildings, and n is the number of building geographic objects.

[0068] The fractal dimension of a building object is the logarithmic relationship between its area and perimeter:

[0069]

[0070] Where A i is the area of the i-th building, P i is the perimeter of the i-th building.

[0071] Concavity distribution entropy:

[0072]

[0073] Where H7(d) is the concavity distribution entropy of the building geographic objects in the block, S_Con i is the concavity of the i-th building, S_Com is the total building concavity, and n is the number of building geographic objects. The concavity of a building object is the area ratio of the object to its convex hull:

[0074]

[0075] Where A i is the area of the i-th building, is the area of the convex hull of the building object.

[0076] Density distribution entropy:

[0077]

[0078] Where H8(d) is the density distribution entropy of the building geographic objects in the block, S_den i is the density of the i-th building, S_den is the total building density, and n is the number of building geographic objects. The density of a building object refers to the area ratio of the object to its corresponding Voronoi polygon:

[0079]

[0080] Where A i is the area of the i-th building, is the area of the Voronoi polygon corresponding to the building.

[0081] The overall aggregation structural characteristics of the geographical objects of the block are calculated, including the average height of the buildings in the block and six land use ratio indicators, as shown in Equations (24) to (30):

[0082] Average building height in the block The calculation of is as follows:

[0083]

[0084] Where h i,b is the height of the i-th building in the block, N b is the number of building objects within the block.

[0085] The six land use ratio indicators are calculated as follows:

[0086]

[0087] In the formula, ULR1 represents the first land use indicator, A i,g is the area of the i-th green space object in the block, a i,w is the area of the i-th water body in the block, A block is the area of the block, N g and N w are the number of green space and water objects within the block, respectively.

[0088]

[0089] In the formula, ULR2 represents the second land use indicator, N POI_facilities is the number of urban facility POIs in the block, is the total area of all building objects within the block.

[0090]

[0091] In the formula, ULR3 represents the third land use indicator, N POI_facilities is the number of urban facility POIs in the block, A block is the area of the block, A i,g is the area of the i-th green space object in the block, A i,w is the area of the i-th water body in the block, A i,b is the area of the i-th building object in the block, N g and N w are the number of green spaces and water bodies in the block, N bis the number of building objects within the block.

[0092]

[0093] Where, ULR4 represents the fourth land use indicator, L inner_roads is the length of the internal road object of the block, N r is the number of internal roads within the block, A i,b is the area of the i-th building object in the block.

[0094]

[0095] Where, ULR5 represents the fifth land use indicator, L inner_roads is the length of the internal road object of the block, N r is the number of internal roads within the block, A block is the area of the block, A i,g is the area of the i-th green space object in the block, A i,w is the area of the i-th water body in the block, A i,b is the area of the i-th building object in the block, N g and N w are the number of green spaces and water bodies in the block, N b is the number of building objects within the block.

[0096]

[0097] In the formula, ULR6 represents the sixth land use indicator, E entities is the sum of the boundary lengths of buildings, green spaces, water bodies, road objects, and other areas of the block, A block is the area of the block.

[0098] Moreover, the calculation methods of the three land cover remote sensing application indices in step 3.2 are shown in equations (31) to (33):

[0099]

[0100] Where NDVI is the normalized difference vegetation index, NIR represents the near infrared band, and Red represents the red band.

[0101]

[0102] Where NDBI is the normalized building index, SAVI is the soil conditioning vegetation index, SWIR represents the short infrared band, NIR represents the near infrared band, Red represents the red band, and L is the soil conditioning factor.

[0103]

[0104] Where NDWI is the normalized water index, Green represents the green wave band, and NIR represents the near infrared band.

[0105] Moreover, the convolutional neural network constructed in step 4 includes a functional area category feature extraction module and a functional area category feature classification module. The feature extraction module obtains high-dimensional functional area category features in the form of a combination of convolution layers and pooling layers, including four convolution layers and three pooling layers; wherein, each convolution layer is followed by an activation function layer, which uses pytorch's ReLU() activation function, and the pooling layer uses pytorch's MaxPool2d() function. The feature classification module processes the high-dimensional features obtained in the feature extraction module to predict functional area categories. The feature classification module consists of a flattening layer, two fully connected layers and a classification output layer; wherein, the flattening layer uses pytorch's flatten() function, the fully connected layer uses the Linear() function, the output layer uses the log_softmax() function, and the output dimension is the total number of label functional area types.

[0106] Aggregate the semantic spatial distribution feature vectors of POIs within the block obtained in step 2 and the elements, neighborhoods, overall spatial pattern feature vectors, surface cover remote sensing application index, and image depth semantic features of the geographic objects within the block obtained based on vector and image data in steps 3.1 to 3.3, and give each block a total of 530-dimensional feature vector description.

[0107] The block dataset is divided into training set, validation set and test set. The network input is the 530-dimensional feature vector description of the block, and the output is the category of urban functional area. The network optimizer uses the Adam algorithm, and the learning rate and decay rate are set to 3e respectively. -5 and 0.95; the main architecture of encoder-fully connected layer-Softmax classifier uses the encoder parameters as initial values and realizes automatic recognition and classification of urban functional areas through supervised learning of parameters.

[0108] The present invention also provides a system for identifying urban functional areas based on spatial pattern features of geographic objects, which is used to implement the above-mentioned method for identifying urban functional areas based on spatial pattern features of geographic objects.

[0109] Moreover, it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the above-mentioned urban functional area identification method based on the spatial pattern characteristics of geographic objects.

[0110] Compared with the prior art, the present invention has the following advantages:

[0111] 1) A new spatial semantic recognition framework for urban functions was established, integrating multi-source data to mine the attribute semantics, spatial patterns, and multi-dimensional geographic features of urban blocks, quantitatively describing typical urban functions and achieving the coupling of geographic space and geographic relationships.

[0112] 2) In response to the complexity of urban functional space and the relationships between its internal elements, a spatial metric was constructed to describe multiple geographic objects (including buildings, roads, green spaces, water bodies, and POIs) and the interactions between them, providing an effective and comprehensive method for the semantic representation of urban block space;

[0113] 3) Based on the Voronoi geometric model and the basic unit of urban blocks, a Block-POI spatial semantic corpus is established. The spatial distribution attributes and positional relationships of POI geographic objects within the block are converted into dense feature vectors using a deep learning language model to support the semantic expression of functional space. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 This is an overall flow chart of the urban functional area identification method according to an embodiment of the present invention.

[0115] Figure 2 This is a diagram of the neural network structure for identifying and classifying urban functional areas according to an embodiment of the present invention. DETAILED DESCRIPTION

[0116] The present invention provides a method and system for identifying typical urban functional areas based on the spatial pattern characteristics of geographic objects. The following uses open source data OSM data as an example and combines the accompanying drawings to further illustrate the technical solution of the present invention.

[0117] Example 1

[0118] like Figure 1 As shown, the present invention provides a typical urban functional area identification method based on the spatial pattern characteristics of geographic objects, comprising the following steps:

[0119] Step 1: Use a series of natural and man-made separating elements (such as rivers, roads, etc.) to divide the urban space area into research units, namely blocks.

[0120] Step 1.1: Preprocess the open source geospatial vector data to meet the data usage specifications.

[0121] ① Data spatial reference conversion, including conversion between WGS84, mobile Internet geodetic coordinate system and CGCS2000 coordinate system; ② Topological check of geographic feature data, such as building surface data and road linear data; ③ Semantic attribute supplementation for data with missing key attributes of POI.

[0122] In step 1.2, ArcGIS software was used to merge the OpenStreetMap (OSM) multi-lane line data. Based on the topology check results, the broken lines in the multi-lane line data were connected to simplify the OSM road line data.

[0123] In step 1.3, urban space is divided using the main road-level road data and river data excluding wetlands to obtain urban blocks.

[0124] Step 2: Use the Voronoi diagram to convert the spatial distribution attributes and location relationships of POI objects in the block into a spatial semantic corpus of Block-POI. Use the Word2Vec model to obtain high-dimensional continuous dense vectors of different POI types in each block, and obtain the socioeconomic feature vector of each block through weighted averaging.

[0125] In step 2.1, analogous to the natural language processing process, the blocks in the study area are considered as "documents" and the POIs within the blocks as "words". The spatial semantic corpus of Block-POI is constructed. The specific steps are as follows:

[0126] ① Calculate the Euclidean distance between all POI pairs in the block, and select the POI pair with the farthest distance as the two endpoints of the word arrangement sequence in the corpus, that is, <P start ,P end >;② Taking the block as the unit, abstract the POI into the graph node, construct the unweighted and undirected Voronoi diagram, and calculate it according to the Dijkstra algorithm <P start ,P end >The shortest path between them is calculated, and "POI words" are recorded in sequence to construct a "block document" containing POI spatial context information; ③ Following the document construction process, point pairs are constructed based on block centers, and the Euclidean distance between block point pairs in the study area is calculated to obtain the "block document"; ② and ③ constitute the Block-POI spatial semantic corpus.

[0127] In step 2.2, the Word2Vec model is used to obtain high-dimensional continuous dense vectors of different POI types in each block.

[0128] Word2Vec neural network is a distributed text representation method, which is divided into two models: Skip-gram and CBOW. Since the present invention aims to predict the function type through the contextual spatial distribution information, that is, "predicting the central word through the surrounding words", the CBOW model is selected for training to obtain the probability vector of each type of POI in the block. Its objective function L and the probability distribution p(w t |Context(w t ))The calculation method is as follows:

[0129] Context(w t )=w t-c ,…,w t-1, w t+1 ,…,w t+c (1)

[0130]

[0131] Where w t is the current word, c is the context window size, which is set to 5 in this embodiment. t ) represents the current word w t The context input vector, T is the corpus size, w i is any word in the corpus, e(w t ,Context(w t )) is the energy function of the Word2Vec model, which is calculated as follows:

[0132] E(w i ,w j )=-(v(w i )·v(w j )) (3)

[0133] Where, E(w i ,w j ) represents word w i Its context word w j The vector inner product, v(w i )、v(w j ) represents the word vector.

[0134] In this embodiment, the output word vector dimension parameter of the CBOW model is set to 256, and the number of iterations parameter is set to 20.

[0135] In step 2.3, an improved TF-IDF algorithm is introduced to weight the POI word vectors, calculate the actual contribution of different POI types in the Block-POI corpus, and explore the scale and usage characteristics of POI functions in each block.

[0136] Since POI geographic objects are unevenly distributed due to the influence of economic activities, geographical environment, etc., an improved TF-IDF algorithm is introduced to weight POI word vectors, calculate the actual contribution of different POI types in the Block-POI corpus, and explore the scale and usage characteristics of POI functions in each block.

[0137] First, determine the TF value of the POI semantic function type. The specific calculation method is as follows:

[0138]

[0139] Where TF i,j represents the TF value of the i-th POI type in the j-th block, n i,j is the number of times the i-th POI type appears in the i-th block, ∑ k n k,j is the total number of occurrences of each type of POI in the jth block.

[0140] To address the uneven distribution problem between classes, the improved IDF is used to determine the IDF value of the POI semantic function type. The improved calculation method is as follows:

[0141]

[0142] Where, IDF i represents the IDF value of the i-th POI type, N is the total number of blocks, N m represents the number of blocks of type m, d im represents the number of blocks in the mth type of block that contain the i-th POI type, d ic,c≠m It represents the number of blocks in other categories that contain the i-th POI type. The specific calculation formula is as follows:

[0143]

[0144] Where w i Refers to the i-th POI type, d j refers to the jth block, c m Refers to the mth type of block, <w i ∈d j >∩ <d j ∈c m > represents the value judgment condition.

[0145] To solve the problem of uneven distribution within a class, we use the parameter class frequency CF i , combined with the above calculation parameters TF i,j and IDF i , the improved TF-IDF algorithm is shown in formula (8):

[0146]

[0147] Where TF-IDF i represents the weight of the i-th POI type, d im N represents the number of blocks in the mth block in the dataset that contain the i-th POI type. m represents the number of blocks of type m, N is the total number of blocks, d ic,c≠m Indicates the number of blocks containing the i-th POI type in other categories of blocks.

[0148] To avoid IDF i When a true number or a denominator of zero appears during the calculation process, equation (8) is smoothed, as shown in equation (9):

[0149]

[0150] Based on the POI semantic type feature vector and type weight obtained in steps 2.2 and 2.3, the socioeconomic characteristics of each block are calculated as follows:

[0151]

[0152] Where, v SE (block) refers to the socioeconomic characteristic vector v of the block Social-Economic (block), v(POI_Category(i)) is the vector representation of different POI types in the block obtained in step 2.2, and weight(POI_Category(i)) is the weight of different POI types obtained by the improved TF-IDF algorithm.

[0153] Step 3: Taking blocks as units, calculate the various elements, neighborhood, and overall spatial distribution structure indicators of their internal surface features (including buildings, roads, green spaces, and water bodies) based on vector data. Calculate the land cover remote sensing application index and image depth semantic features formed by the aggregation of various elements within the block based on medium-resolution remote sensing images. The combination of the two constitutes the spatial pattern feature vector of each block.

[0154] Step 3.1, taking the block as the unit, calculate the indicators based on the element characteristics of the geographic object itself, the indicators based on the neighborhood characteristics of the surface elements, and the indicators based on the overall distribution characteristics of the surface elements in the block based on the vector data of the internal surface elements (including buildings, roads, green spaces, and water bodies).

[0155] Combining the principle of information entropy, the area distribution entropy, perimeter distribution entropy, mean radius distribution entropy, minimum circumscribed rectangle direction distribution entropy, compactness distribution entropy, fractal dimension distribution entropy, and concavity distribution entropy of building objects in the block are calculated based on the element level. The density distribution entropy of building objects is calculated based on the neighborhood level. The overall aggregation structure characteristics of the geographical objects in the block are calculated to measure the diversity of the geometric and distribution attributes of the buildings in the block. The specific calculation methods are shown in Equations (11) to (30):

[0156] Area distribution entropy:

[0157]

[0158] Where H1(d) is the area distribution entropy of the building geographic objects in the block, A iis the area of the i-th building, S is the total building area, and n is the number of building geographic objects.

[0159] Perimeter distribution entropy:

[0160]

[0161] Where H2(d) is the perimeter distribution entropy of the building geographic objects in the block, P i is the perimeter of the i-th building, L is the total perimeter of buildings, and n is the number of building geographic objects.

[0162] Mean radius distribution entropy:

[0163]

[0164] Where H3(d) is the average radius distribution entropy of building geographic objects in the block, R i is the average radius of the i-th building, R is the sum of the average radius of all buildings, and n is the number of building geographic objects; the average radius of a building object is calculated by the average distance from all boundary points on its outline to its center point:

[0165]

[0166] Where, d j represents the distance from the boundary point j to the center point, and M is the number of boundary points.

[0167] Minimum circumscribed rectangle directional distribution entropy: The main direction of the minimum circumscribed rectangle of a building is defined as the angle between its long axis and the horizontal axis. First, the main direction θ of each building object needs to be calculated to determine the number of main direction categories. Then, the number of buildings in different main direction classification intervals within the block is counted to analyze the difference in building direction distribution within the block. The direction statistics results are substituted into the following formula:

[0168]

[0169] Where H4(d) is the directional distribution entropy of the building instance object level in the building group, S i is the number of buildings in the i-th direction, N is the total number of buildings, and n4 is the number of main direction categories.

[0170] Compactness distribution entropy:

[0171]

[0172] Where H5(d) is the compactness distribution entropy of the building geographic objects in the block, S_Com iis the compactness of the i-th building, S_Com is the total building compactness, and n is the number of building geographic objects. The compactness of a building object is a quadratic relationship between its area and perimeter:

[0173]

[0174] Where A i is the area of the i-th building, P i is the perimeter of the i-th building.

[0175] Fractal dimension distribution entropy:

[0176]

[0177] Where H6(d) is the fractal dimension distribution entropy of the building geographic objects in the block, S_Frac i is the fractal dimension of the i-th building, S_Frac is the total fractal dimension of buildings, and n is the number of building geographic objects.

[0178] The fractal dimension of a building object is the logarithmic relationship between its area and perimeter:

[0179]

[0180] Where A i is the area of the i-th building, P i is the perimeter of the i-th building.

[0181] Concavity distribution entropy:

[0182]

[0183] Where H7(d) is the concavity distribution entropy of the building geographic objects in the block, S_Con i is the concavity of the i-th building, S_Con is the total building concavity, and n is the number of building geographic objects. The concavity of a building object is the area ratio of the object to its convex hull:

[0184]

[0185] Where A i is the area of the i-th building, is the area of the convex hull of the building object.

[0186] Density distribution entropy:

[0187]

[0188] Where H8(d) is the density distribution entropy of the building geographic objects in the block, S_deni is the density of the i-th building, S_den is the total building density, and n is the number of building geographic objects. The density of a building object refers to the area ratio of the object to its corresponding Voronoi polygon:

[0189]

[0190] Where A i is the area of the i-th building, is the area of the Voronoi polygon corresponding to the building.

[0191] The overall aggregation structural characteristics of the geographical objects of the block are calculated, including the average height of the buildings in the block and six land use ratio indicators, as shown in Equations (24) to (30):

[0192] Average building height in the block The calculation of is as follows:

[0193]

[0194] Where h i,b is the height of the i-th building in the block, N b is the number of building objects within the block.

[0195] The six land use ratio indicators are calculated as follows:

[0196]

[0197] In the formula, ULR1 represents the first land use indicator, A i,g is the area of the i-th green space object in the block, A i,w is the area of the i-th water body in the block, A block is the area of the block, N g and N w are the number of green space and water objects within the block, respectively.

[0198]

[0199] In the formula, ULR2 represents the second land use indicator, N POI_facilities is the number of urban facility POIs in the block, is the total area of all building objects within the block.

[0200]

[0201] In the formula, ULR3 represents the third land use indicator, N POI_facilities is the number of urban facility POIs in the block, A block is the area of the block, Ai,g is the area of the i-th green space object in the block, A i,w is the area of the i-th water body in the block, A i,b is the area of the i-th building object in the block, N g and N w are the number of green spaces and water bodies in the block, N b is the number of building objects within the block.

[0202]

[0203] Where, ULR4 represents the fourth land use indicator, L inner_roads is the length of the internal road object of the block, N r is the number of internal roads within the block, A i,b is the area of the i-th building object in the block.

[0204]

[0205] Where, ULR5 represents the fifth land use indicator, L inner_roads is the length of the internal road object of the block, N r is the number of internal roads within the block, A block is the area of the block, A i,g is the area of the i-th green space object in the block, A i,w is the area of the i-th water body in the block, A i,b is the area of the i-th building object in the block, N g and N w are the number of green spaces and water bodies in the block, N b is the number of building objects within the block.

[0206]

[0207] In the formula, ULR6 represents the sixth land use indicator, E entities is the sum of the boundary lengths of buildings, green spaces, water bodies, road objects, and other areas of the block, A block is the area of the block.

[0208] In step 3.2, the minimum bounding rectangle of the block segmented in step 1 is used as a mask to extract the corresponding surface range in the medium-resolution remote sensing image and calculate the three surface cover remote sensing application indices for each block, as shown in equations (31) to (33):

[0209]

[0210] Where NDVI is the normalized difference vegetation index, NIR represents the near infrared band, and Red represents the red band.

[0211]

[0212] Wherein, NDBI is the normalized building index, SAVI is the soil conditioning vegetation index, SWIE represents the short infrared band, NIR represents the near infrared band, Red represents the red band, and L is the soil conditioning factor, which is set to 0.5 in this embodiment.

[0213]

[0214] Where NDWI is the normalized water index, Green represents the green wave band, and NIR represents the near infrared band.

[0215] In step 3.3, the remote sensing image of the study area is segmented into input images of size 512×512, and the existing deep convolutional autoencoder model MegNet is used to learn image features.

[0216] The autoencoder model consists of two parts: an encoder and a decoder. The input image also serves as the model's label image. The model uses the similarity estimate between the decoded output and the original input image as the loss function for iterative optimization. As a result, the trained MegNet autoencoder has excellent image feature representation capabilities and can be used as a feature extractor to obtain deep semantic features of each block in the study area. These high-dimensional deep semantic features have a dimension of 256 (the channel parameter value of the encoder's last convolutional layer).

[0217] In step 3.4, Z-score normalization is performed on the various elements, neighborhood, and overall spatial distribution structure indicators calculated in step 3.1 and the land cover remote sensing application index calculated in step 3.2.

[0218] To prevent the metric's dimension from affecting subsequent analysis, the calculated metrics were normalized using the Z-score, ensuring invariance to translation, scaling, and rotation. Their mean and standard deviation were calculated for all instances in the MegNet autoencoder training data set. Since the convolutional network model outputs predicted probabilities, with probability values in the range [0, 1], the deep semantic features extracted by the MegNet encoder do not require normalization.

[0219] Step 4: Based on the socioeconomic feature vector obtained in step 2 and the spatial pattern feature vector obtained in step 3, a convolutional neural network is constructed to automatically identify and classify the spatial semantics of urban functional areas.

[0220] The concat function based on PyTorch aggregates the semantic spatial distribution feature vector of POIs within the block obtained in step 2 and the elements, neighborhoods, overall spatial pattern feature vectors, surface cover remote sensing application index, and image depth semantic features of the geographic objects within the block obtained based on vector and image data in steps 3.1 to 3.3, giving each block a total of 530-dimensional (256-dim + 8-dim + 7-dim + 3-dim + 256-dim = 530-dim) feature vector description.

[0221] The constructed convolutional neural network includes a functional area category feature extraction module and a functional area category feature classification module, and its structure is as follows: Figure 2 As shown in the figure. The feature extraction module uses a combination of convolutional layers and pooling layers to obtain high-dimensional functional area category features, including four convolutional layers (Conv layer1, Conv layer2, Conv layer3, Conv layer4) and three pooling layers (Pooling layer1, Pooling layer2, Pooling layer3). Each convolutional layer is followed by an activation function layer, which uses PyTorch's ReLU() activation function, and the pooling layer uses PyTorch's MaxPool2d() function. The feature classification module processes the high-dimensional features obtained in the feature extraction module to predict functional area categories. The feature classification module consists of a flattening layer, two fully connected layers (FC layer1, FC layer2), and a classification output layer. The flattening layer uses PyTorch's flatten() function, the fully connected layer uses the Linear() function, and the output layer uses the log_softmax() function. The output dimension is the total number of labeled functional area types.

[0222] The block dataset is divided into training set, validation set and test set in a ratio of 6:2:2. The network input is the 530-dimensional feature vector description of the block, and the output is the category of urban functional area. The network optimizer uses the Adam algorithm, and the learning rate and decay rate are 3×10 -5 The main architecture of encoder-fully connected layer-Softmax classifier uses the encoder parameters as initial values and realizes automatic recognition and classification of urban functional areas through supervised learning of parameters.

[0223] Example 2

[0224] Based on the same inventive concept, the present invention also provides an urban functional area identification system based on the spatial pattern characteristics of geographic objects, including a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the above-mentioned urban functional area identification method based on the spatial pattern characteristics of geographic objects.

[0225] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0226] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A method for identifying urban functional areas based on spatial pattern characteristics of geographic objects, characterized in that: The following steps are involved: Step 1: pre-process the open source geospatial vector data and divide the urban spatial area into research units, i.e., blocks; Step 2: The spatial distribution attributes and location relationships of POI objects within the block are converted into a spatial semantic corpus of Block-POI. High-dimensional continuous dense vectors of different POI types within each block are obtained, and the socioeconomic characteristic vector of each block is obtained through weighted averaging. Step 2.1: Consider the blocks in the study area as "documents" and the POIs within the blocks as "words" to construct a Block-POI spatial semantic corpus. The specific operation is as follows: ① Calculate the Euclidean distance between all POI pairs in the block, and select the POI pair with the farthest distance as the two endpoints of the word arrangement sequence in the corpus, that is, <P start ,P end >;② Taking the block as the unit, abstract the POI into the graph node, construct the unweighted and undirected Voronoi diagram, and calculate it according to the Dijkstra algorithm <P start ,P end >, and record "POI words" in sequence to construct a "block document" containing POI spatial context information; ③ Following the document construction process, point pairs are constructed based on block centers, and the Euclidean distance between block point pairs in the study area is calculated to obtain the "block document"; ② and ③ constitute the Block-POI spatial semantic corpus; Step 2.2: Use the Word2Vec model to obtain high-dimensional continuous dense vectors of different POI types in each block; The CBOW model of the Word2Vec neural network is used to train the probability vector of each type of POI in the block. Its objective function L and the probability distribution p(w t |Context(w t ))The calculation method is as follows: Context(w t )=w t-c ,…,w t-1, w t+1 ,…,w t+c (1) Where w t is the current word, c is the context window size, Context(w t ) represents the current word w t The context input vector, T is the corpus size, w i is any word in the corpus, E(w t ,Context(w t )) is the energy function of the Word2Vec model, which is calculated as follows: E(w i ,w j )=-(v(w i )·v(w j )) (3) Where, E(w i ,w j ) represents word w i Its context word w j The vector inner product, v(w i )、v(w j ) represents word vector; Step 2.3: Improve the true value part of the log function of the TF-IDF algorithm. Use the improved TF-IDF algorithm to weight the POI word vectors, calculate the actual contribution of different POI types in the Block-POI corpus, and explore the scale and usage characteristics of POI functions in each block. Step 3: Taking the block as a unit, calculate the various elements, neighborhood, and overall spatial distribution structure indicators of its internal surface features based on vector data. Calculate the land cover remote sensing application index and image depth semantic features formed by the aggregation of various elements within the block based on medium-resolution remote sensing imagery. The combination of the two constitutes the spatial pattern feature vector of each block. Step 4: Based on the socioeconomic feature vector obtained in step 2 and the spatial pattern feature vector obtained in step 3, a convolutional neural network is used to automatically identify and classify the spatial semantics of urban functional areas.

2. The method for identifying urban functional areas based on spatial pattern characteristics of geographic objects according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1: Preprocess the open source geospatial vector data to meet the data usage specifications; Preprocessing operations include: ① data spatial reference conversion; ② topological check of geographic feature data; ③ semantic attribute supplementation for data with missing key attributes of POI; Step 1.2: Use ArcGIS software to merge the multi-lane line data in the open source geospatial vector data. Based on the topology check results, connect the broken lines in the multi-lane line data to simplify the road line data. In step 1.3, urban space is divided using the main road-level road data and river data excluding wetlands to obtain urban blocks.

3. The method for identifying urban functional areas based on spatial pattern characteristics of geographic objects according to claim 1, characterized in that: In step 2.3, the TF value of the POI semantic function type is first determined. The specific calculation method is as follows: Where TF i,j represents the TF value of the i-th POI type in the j-th block, n i,j is the number of times the i-th POI type appears in the j-th block, ∑ k n k,j is the total number of occurrences of each type of POI in the jth block; To address the uneven distribution problem between classes, the improved IDF is used to determine the IDF value of the POI semantic function type. The improved calculation method is as follows: Where, IDF i represents the IDF value of the i-th POI type, N is the total number of blocks, N m represents the number of blocks of type m, d im represents the number of blocks in the mth type of block that contain the i-th POI type, d ic,c≠m It represents the number of blocks in other categories that contain the i-th POI type. The specific calculation formula is as follows: Where w i Refers to the i-th POI type, d j refers to the jth block, c m Refers to the mth type of block, <w i ∈d j >∩ <d j ∈c m >Represents the value judgment condition; To solve the problem of uneven distribution within a class, we use the parameter class frequency CF i , combined with the above calculation parameters TF i,j and IDF i , the improved TF-IDF algorithm is shown in formula (8): Where TF-IDF i represents the weight of the i-th POI type, d im N represents the number of blocks in the mth block in the dataset that contain the i-th POI type. m represents the number of blocks of type m, N is the total number of blocks, d ic,c≠m represents the number of blocks in other categories that contain the i-th POI type; To avoid IDF i When a true number or a denominator of zero appears during the calculation process, equation (8) is smoothed, as shown in equation (9): Based on the POI semantic type feature vector and type weight obtained in steps 2.2 and 2.3, the socioeconomic characteristics of each block are calculated as follows: Where, v SE (block) refers to the socioeconomic characteristic vector v of the block Social-Economic (block), v(POI_Category(i)) is the vector representation of different POI types in the block obtained in step 2.2, and weight(POI_Category(i)) is the weight of different POI types obtained by the improved TF-IDF algorithm.

4. The method for identifying urban functional areas based on spatial pattern characteristics of geographic objects according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1: Taking the block as the unit, calculate the indicators based on the element characteristics of the geographic object itself, the indicators based on the neighborhood characteristics of the surface elements, and the indicators based on the overall distribution characteristics of the surface elements in the block based on the internal surface element vector data; Step 3.2: Using the minimum bounding rectangle of the block segmented in step 1 as a mask, extract the corresponding surface area in the medium-resolution remote sensing image and calculate the three surface cover remote sensing application indices for each block. In step 3.3, the remote sensing image of the study area is segmented into input images of a certain size, and the existing deep convolutional autoencoder model MegNet is used to extract the deep semantic features of the image; In step 3.4, Z-score normalization is performed on the various elements, neighborhood, and overall spatial distribution structure indicators calculated in step 3.1 and the land cover remote sensing application index calculated in step 3.

2.

5. The method for identifying urban functional areas based on spatial pattern characteristics of geographic objects according to claim 4, characterized in that: In step 3.1, based on the principle of information entropy, the area distribution entropy, perimeter distribution entropy, mean radius distribution entropy, minimum circumscribed rectangle direction distribution entropy, compactness distribution entropy, fractal dimension distribution entropy, and concavity distribution entropy of the building objects in the block are calculated based on the element level. The density distribution entropy of the building objects is calculated based on the neighborhood level to measure the diversity of the geometric and distribution attributes of the buildings in the block. The specific calculation method is shown in Equations (11) to (30): Area distribution entropy: Where H1(d) is the area distribution entropy of the building geographic objects in the block, A i is the area of the i-th building, S is the total building area, and n is the number of building geographic objects; Perimeter distribution entropy: Where H2(d) is the perimeter distribution entropy of the building geographic objects in the block, P i is the perimeter of the i-th building, L is the total perimeter of buildings, and n is the number of building geographic objects; Mean radius distribution entropy: Where H3(d) is the average radius distribution entropy of building geographic objects in the block, R i is the average radius of the i-th building, R is the sum of the average radius of all buildings, and n is the number of building geographic objects; the average radius of a building object is calculated by the average distance from all boundary points on its outline to its center point: Where, d j represents the distance from the boundary point j to the center point, and M is the number of boundary points; Minimum circumscribed rectangle directional distribution entropy: The main direction of the minimum circumscribed rectangle of a building is defined as the angle between its long axis and the horizontal axis. First, the main direction θ of each building object needs to be calculated to determine the number of main direction categories. Then, the number of buildings in different main direction classification intervals within the block is counted to analyze the difference in building direction distribution within the block. The direction statistics results are substituted into the following formula: Where H4(d) is the directional distribution entropy of the building instance object level in the building group, S i is the number of buildings in the i-th direction, N is the total number of buildings, and n4 is the number of main direction categories; Compactness distribution entropy: Where H5(d) is the compactness distribution entropy of the building geographic objects in the block, S_Com i is the compactness of the i-th building, S_Com is the total building compactness, and n is the number of building geographic objects; where the compactness of a building object is a quadratic relationship between its area and perimeter: Where A i is the area of the i-th building, P i is the perimeter of the i-th building; Fractal dimension distribution entropy: Where H6(d) is the fractal dimension distribution entropy of the building geographic objects in the block, S_Frac i is the fractal dimension of the i-th building, S_Frac is the total fractal dimension of buildings, and n is the number of building geographic objects; where the fractal dimension of a building object is the logarithmic relationship between its area and perimeter: Where A i is the area of the i-th building, P i is the perimeter of the i-th building; Concavity distribution entropy: Where H7(d) is the concavity distribution entropy of the building geographic objects in the block, S_Con i is the concavity of the i-th building, S_Con is the total building concavity, and n is the number of building geographic objects; where the concavity of a building object is the area ratio of the object to its convex hull: Where A i is the area of the i-th building, A chi is the area of the convex hull of the building object; Density distribution entropy: Where H8(d) is the density distribution entropy of the building geographic objects in the block, S_den i is the density of the i-th building, S_den is the total building density, and n is the number of building geographic objects; the density of a building object refers to the area ratio of the object to its corresponding Voronoi polygon: Where A i is the area of the i-th building, A voronoi_polugoni is the area of the Voronoi polygon corresponding to the building; The overall aggregation structural characteristics of the geographical objects of the block are calculated, including the average height of the buildings in the block and six land use ratio indicators, as shown in Equations (24) to (30): Average building height in the block The calculation is as follows: Where h i,b is the height of the i-th building in the block, N b is the number of building objects in the block; The six land use ratio indicators are calculated as follows: In the formula, ULR1 represents the first land use indicator, A i,g is the area of the i-th green space object in the block, A i,w is the area of the i-th water body in the block, A block is the area of the block, N g and N w are the number of green space and water objects within the block, respectively; In the formula, ULR2 represents the second land use indicator, N POI_facilities is the number of urban facility POIs in the block, is the total area of all building objects within the block; In the formula, ULR3 represents the third land use indicator, N POI_facilities is the number of urban facility POIs in the block, A block is the area of the block, A i,g is the area of the i-th green space object in the block, A i,w is the area of the i-th water body in the block, A i,b is the area of the i-th building object in the block, N g and N w are the number of green spaces and water bodies in the block, N b is the number of building objects in the block; Where, ULR4 represents the fourth land use indicator, L inner_roads is the length of the internal road object of the block, N r is the number of internal roads within the block, A i,b is the area of the i-th building object in the block; Where, ULR5 represents the fifth land use indicator, L inner_roads is the length of the internal road object of the block, N r is the number of internal roads within the block, A block is the area of the block, A i,g is the area of the i-th green space object in the block, A i,w is the area of the i-th water body in the block, A i,b is the area of the i-th building object in the block, N g and N w are the number of green spaces and water bodies in the block, N b is the number of building objects in the block; In the formula, ULR6 represents the sixth land use indicator, E entities is the sum of the boundary lengths of buildings, green spaces, water bodies, road objects, and other areas of the block, A block is the area of the block.

6. The method for identifying urban functional areas based on spatial pattern characteristics of geographic objects according to claim 4, characterized in that: The calculation methods of the three land cover remote sensing application indices in step 3.2 are shown in equations (31) to (33): Where NDVI is the normalized vegetation index, NIR represents the near infrared band, and Red represents the red band; Where NDBI is the normalized building index, SAVI is the soil conditioning vegetation index, SWIR represents the short infrared band, NIr represents the near infrared band, Red represents the red band, and L is the soil conditioning factor. Where NDWI is the normalized water index, Green represents the green wave band, and NIR represents the near infrared band.

7. The method for identifying urban functional areas based on spatial pattern characteristics of geographic objects according to claim 1, characterized in that: The convolutional neural network constructed in step 4 includes a functional area category feature extraction module and a functional area category feature classification module. The feature extraction module obtains high-dimensional functional area category features in the form of a combination of convolutional layers and pooling layers, including four convolutional layers and three pooling layers; each convolutional layer is followed by an activation function layer, which uses the ReLU() activation function of pytorch, and the pooling layer uses the MaxPool2d() function of pytorch; the feature classification module processes the high-dimensional features obtained in the feature extraction module to predict the functional area category. The feature classification module consists of a flattening layer, two fully connected layers and a classification output layer; the flattening layer uses the flatten() function of pytorch , the fully connected layer uses the Linear() function, the output layer uses the log_softmax() function, and the output dimension is the total number of label functional area types; the semantic spatial distribution feature vector of POIs in the block obtained in step 2 is aggregated with the elements, neighborhood, overall spatial pattern feature vector, surface cover remote sensing application index and image depth semantic features of the geographic objects in the block obtained based on vector and image data in steps 3.1 to 3.3, and a total of 530-dimensional feature vector description is given to each block; the block dataset is divided into training set, validation set and test set, the network input is the 530-dimensional feature vector description of the block, and the output is the category of the urban functional area; the network optimizer uses the Adam algorithm, and the learning rate and decay rate are set to 3e respectively -5 and 0.95; the main architecture of encoder-fully connected layer-Softmax classifier uses the encoder parameters as initial values and realizes automatic recognition and classification of urban functional areas through supervised learning of parameters.

8. A system for identifying urban functional areas based on spatial pattern characteristics of geographic objects, characterized by: It includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the urban functional area identification method based on the spatial pattern characteristics of geographic objects as described in any one of claims 1 to 7.

Images