Multi-source data intelligent learning and fusion urban living space identification method

Abstract

The invention relates to the technical field of urban space recognition, and provides a multi-source data intelligent learning and fusion urban living space recognition method, which comprises the following steps of: obtaining a plurality of urban space remote sensing images of a target area from a plurality of sources; performing superpixel segmentation on the plurality of urban space remote sensing images to obtain superpixel images corresponding to the plurality of urban space remote sensing images; extracting low-frequency features and high-frequency features of the plurality of superpixel images; fusing the low-frequency features of the plurality of superpixel images with the low-frequency features to obtain fused low-frequency features, and fusing the high-frequency features of the plurality of superpixel images with the high-frequency features to obtain fused high-frequency features; reconstructing an image based on the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image; identifying a living space area of the fused urban space remote sensing image by using an adaptive threshold segmentation algorithm; the method has high urban living space identification precision.

Description

Technical Field

[0001] The present invention relates to the technical field of urban space identification, and more specifically, to an urban residential space identification method based on multi-source data intelligent learning and fusion. Background Art

[0002] Urban residential space identification technology is a method that uses remote sensing data, geospatial data, etc. to identify and extract residential land in urban areas. The advantage of this technology is that it can provide large-scale, timely and objective residential space information, and is widely used in urban planning, land management and other fields.

[0003] Most of the existing urban residential space identification methods are based on the zoning identification of urban functional space. In terms of urban functional area identification, traditional methods mostly rely on a large number of manual collection and field surveys of urban land use, which not only requires a huge workload, but also the identification results are greatly affected by the subjective influence of investigators. In addition, related research is also highly dependent on traditional data such as urban planning and socio-economic statistical yearbooks. These traditional data have certain value, but their update cycle is long, and the data are greatly affected by factors such as administrative boundaries and statistical caliber. The current spatial distribution of urban residential areas is crucial for the government to formulate more scientific urban planning. Therefore, under the defects of traditional data and methods, how to improve the accuracy of urban residential area identification and extraction is an important issue that needs to be solved urgently. In this context, efficient methods that rely on massive data of satellite remote sensing images have emerged for the identification of urban functional areas or land use types. However, the division of urban functional areas is essentially a reflection of its social and economic functions, while remote sensing surveys mainly rely on the spectral characteristics of objects and cannot directly detect the social and economic functions carried by objects. With the rapid development of computer technology, data that can reflect the socio-economic functions of cities are becoming increasingly abundant and more accessible. Research on identifying urban functional areas based on geospatial big data has gradually become one of the hot topics in the field of urban research.

[0004] At present, the identification of urban residential space is mostly included in the research on the zoning identification of urban functional space. In terms of data application, most of them focus on single remote sensing image data. The characteristics of these data determine that it is difficult to fully reflect the characteristics of urban residential space, resulting in low accuracy of recognition results. Even if the research is based on multi-source remote sensing image data, there are still shortcomings such as over-segmentation of remote sensing images and poor boundary processing, resulting in low recognition accuracy of urban residential space. Summary of the invention

[0005] In order to overcome the defect of low urban residential space recognition accuracy in the above-mentioned prior art, the present invention provides an urban residential space recognition method with high recognition accuracy through intelligent learning and fusion of multi-source data.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0007] Acquire several urban space remote sensing images of the target area from several sources;

[0008] Perform super-pixel segmentation on several urban space remote sensing images to obtain super-pixel images corresponding to several urban space remote sensing images;

[0009] Extract low-frequency features and high-frequency features of several super-pixel images;

[0010] The low-frequency features of several super-pixel images are fused with the low-frequency features to obtain fused low-frequency features, and the high-frequency features of several super-pixel images are fused with the high-frequency features to obtain fused high-frequency features;

[0011] Reconstructing an image based on the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image;

[0012] The residential space area of ​​the fused urban space remote sensing image is identified using an adaptive threshold segmentation algorithm.

[0013] The present invention also proposes a multi-source data intelligent learning and fusion urban residential space identification system for realizing the multi-source data intelligent learning and fusion urban residential space identification method, the system comprising:

[0014] An image acquisition module is used to acquire a number of urban space remote sensing images of a target area from a number of sources;

[0015] A superpixel segmentation module is used to perform superpixel segmentation on a plurality of urban space remote sensing images to obtain superpixel images corresponding to the plurality of urban space remote sensing images;

[0016] A feature extraction module is used to extract low-frequency features and high-frequency features of several super-pixel images;

[0017] A feature fusion module is used to fuse the low-frequency features of several super-pixel images with the low-frequency features to obtain fused low-frequency features, and to fuse the high-frequency features of several super-pixel images with the high-frequency features to obtain fused high-frequency features;

[0018] An image reconstruction module, used to reconstruct an image based on the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image;

[0019] The residential space recognition module is used to recognize the residential space area of ​​the fused urban space remote sensing image by using an adaptive threshold segmentation algorithm.

[0020] The present invention also proposes a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the above-mentioned urban residential space identification method based on intelligent learning and fusion of multi-source data.

[0021] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0022] The superpixel segmentation method is used to preprocess the remote sensing image, which reduces the image dimension, eliminates the influence of abnormal pixels, greatly reduces the complexity of image processing, and improves processing efficiency. By extracting the high-frequency and low-frequency features of the superpixel image, and then through feature fusion, it makes up for the defects of the traditional method of high time cost, low accuracy, difficulty in replication and promotion; and the optimal residential area extraction range is calculated through an adaptive threshold segmentation algorithm to obtain the extraction result of the residential area more quickly and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of the first process of the urban residential space identification method of multi-source data intelligent learning and fusion proposed in Example 1;

[0024] Figure 2 A second flow chart of the urban residential space identification method of multi-source data intelligent learning and fusion proposed in Example 2;

[0025] Figure 3 Schematic diagram of three scales for superpixel segmentation of the Guangdong-Hong Kong-Macao Greater Bay Area image proposed in Example 2;

[0026] Figure 4 This is a diagram of the superpixel wavelet transform fusion process proposed in Example 2;

[0027] Figure 5 This is a schematic diagram of the accuracy verification results of different fusion data proposed in Example 2. DETAILED DESCRIPTION

[0028] The accompanying drawings are only used for illustrative purposes and should not be construed as limiting the present embodiment;

[0029] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0030] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0031] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0032] Example 1

[0033] This embodiment proposes a method for identifying urban residential spaces by intelligent learning and fusion of multi-source data. Figure 1 This is a schematic diagram of the first process of the urban residential space identification method of multi-source data intelligent learning and fusion in this embodiment;

[0034] The present embodiment proposes a method for identifying urban residential spaces by intelligent learning and fusion of multi-source data, comprising the following steps:

[0035] S1: Acquire several urban space remote sensing images of the target area from several sources;

[0036] S2: performing super-pixel segmentation on a plurality of urban space remote sensing images to obtain super-pixel images corresponding to the plurality of urban space remote sensing images;

[0037] S3: extracting low-frequency features and high-frequency features of several super-pixel images;

[0038] S4: Fusing the low-frequency features of several super-pixel images with the low-frequency features to obtain fused low-frequency features, and fusing the high-frequency features of several super-pixel images with the high-frequency features to obtain fused high-frequency features;

[0039] S5: reconstructing an image based on the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image;

[0040] S6: Using an adaptive threshold segmentation algorithm to identify the residential space area of ​​the fused urban space remote sensing image.

[0041] In the specific implementation process, the superpixel segmentation method is used to preprocess the remote sensing image, which reduces the image dimension, eliminates the influence of abnormal pixels, greatly reduces the complexity of image processing, and improves processing efficiency. By extracting the high-frequency and low-frequency features of the superpixel image, and then through feature fusion, it makes up for the defects of the traditional method of high time cost, low accuracy, difficulty in replication and promotion; and through the adaptive threshold segmentation algorithm, the optimal residential area extraction range is calculated to obtain the extraction result of the residential area more quickly and accurately.

[0042] In an optional embodiment, the sources of the urban space remote sensing image include POI data, NTL data, HRI data and LDS data.

[0043] As an example, POI data (Point of Interest) refers to points of interest in geographic spatial data, usually including basic information such as location coordinates, name, address, category, etc.; NTL data (Night Light Index, NLI) is widely used in the identification and extraction of urban built-up areas and functional areas due to its advantage of significant differences in Digital Number (DN) values ​​between built-up areas and non-built-up areas and various functional areas; HRI's full name in Chinese is high-resolution image data. LDS (LandScan) data is an early attempt to draw spatiotemporal population maps with a resolution of 30×30 arc seconds worldwide. It is the best census data in the existing population database and can better reflect the spatial distribution of the population. It is used in the identification research of various spatial structures in cities, and is also gradually used for multi-source data fusion to more accurately identify urban spaces.

[0044] In an optional embodiment, when performing superpixel segmentation on several urban space remote sensing images, SLIC image superpixel segmentation is performed on each urban space remote sensing image with a preset number of segmentation blocks to obtain a superpixel image corresponding to each urban space remote sensing image, wherein one urban space remote sensing image corresponds to one superpixel image, each superpixel image is composed of several superpixel blocks of the same size, and the total number of superpixel blocks in each superpixel image is the preset number of segmentation blocks.

[0045] As an exemplary description, the preset number of segmentation blocks is set according to actual needs. In this application, each super-pixel image is divided into 200, 400 and 800 super-pixel blocks for research.

[0046] In an optional embodiment, when extracting low-frequency features and high-frequency features of several super-pixel images, the super-pixel images are decomposed using wavelet transform based on discrete cosine transform to obtain the lowest frequency sub-band number and the highest frequency sub-band coefficient of each super-pixel image, where the lowest frequency sub-band number is the low-frequency feature, and the highest frequency sub-band coefficient is the high-frequency feature.

[0047] In an optional embodiment, when the low-frequency features of several super-pixel images are fused with the low-frequency features, the low-frequency features are fused using a coefficient weighting strategy to obtain the fused low-frequency features;

[0048] The expression of the coefficient weighting strategy includes:

[0049]

[0050] α1+α2+…+α n =1

[0051] Where n represents the total number of superpixel images, represents the result of fusing the low-frequency features of n super-pixel images with the low-frequency features, α n represents the scale factor of the nth superpixel image, Represents the low-frequency features of the nth superpixel image.

[0052] In an optional embodiment, when the high-frequency features of a plurality of super-pixel images are fused with the high-frequency features, the high-frequency features are fused using an absolute value maximum method to obtain the fused high-frequency features;

[0053] The expression of the absolute value maximum method includes:

[0054]

[0055] Where n represents the total number of superpixel images, It represents the result of fusing the high-frequency features of n super-pixel images with the high-frequency features. Represents the high-frequency features of the nth superpixel image.

[0056] In an optional embodiment, when reconstructing an image based on the fused low-frequency features and the fused high-frequency features, a two-dimensional Mallat algorithm is used to reconstruct the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image.

[0057] In an optional embodiment, when the residential space area of ​​the fused urban space remote sensing image is identified by using an adaptive threshold segmentation algorithm, the fused urban space remote sensing image is divided into a residential space area and a non-residential space area by using a maximum inter-class variance method;

[0058] Here, let the probability that the pixels of the fused urban space remote sensing image are assigned to the residential space area be P A (t), let the probability that the pixels of the fused urban space remote sensing image are assigned to the non-residential space area be P B (t), then P A (t) and P B The population variance σ of (t) 2 The expressions include:

[0059] σ 2 =P A (t)(m A (t)-m G ) 2 +P B (t)(m B (t)-m G ) 2

[0060] P A(t)+P B (t) = 1

[0061]

[0062]

[0063] Where t represents the segmentation threshold, P i represents the i-th pixel of the fused urban space remote sensing image, m A (t) represents the average pixel grayscale value of pixels belonging to the residential area, m B (t) represents the average pixel grayscale value of pixels in the non-residential space area, m G Represents the average gray value of the pixels of the entire fused urban space remote sensing image;

[0064] When dividing the fused urban space remote sensing image into residential space areas and non-residential space areas, we use the command P A (t) and P B The population variance σ of (t) 2 The maximum segmentation threshold t is used for segmentation.

[0065] Example 2

[0066] This embodiment is based on the urban residential space recognition method of multi-source data intelligent learning and fusion proposed in Embodiment 1. Figure 2 A second flow chart of the urban residential space identification method of multi-source data intelligent learning and fusion proposed in this embodiment; Figure 3 Schematic diagram of three scales for superpixel segmentation of Guangdong-Hong Kong-Macao Greater Bay Area images proposed in this embodiment; Figure 4 This is a diagram of the superpixel wavelet transform fusion process proposed in this embodiment; Figure 5 This is a schematic diagram of the accuracy verification results of different fusion data proposed in this embodiment, such as Figures 2 to 5 As shown, this embodiment proposes the following specific implementation examples:

[0067] The urban residential space recognition method of multi-source data intelligent learning and fusion comprises the following steps:

[0068] S1: First, SLIC image super-segmentation is used to extract and cluster the features of urban space remote sensing images, and super-pixel blocks with segmentation scales of 800, 400, and 200 are obtained in turn.

[0069] S2: Then, the discrete cosine transform is used to perform wavelet transform on the data from different sources to obtain high-frequency and low-frequency features at three scales.

[0070] S3: Next, based on superpixel wavelet fusion, high-frequency and low-frequency features of data from different sources are fused from the perspective of superpixels of different scales.

[0071] S4: Finally, the OSTU algorithm is used to calculate the optimal threshold and extract the urban residential space.

[0072] Furthermore, SLIC image super-segmentation is used to extract and cluster the color, brightness, texture and other features of urban space remote sensing images, and the segmentation coefficients are set to 200, 400, and 800 respectively, and super-pixel segmentation images of these three scales are output.

[0073] Furthermore, discrete cosine transform is used to decompose the image into four sub-bands: low-frequency-low-frequency (LL), low-frequency-high-frequency (LH), high-frequency-low-frequency (HL) and high-frequency-high-frequency (HH). A more detailed multi-scale expression is obtained through iteration for subsequent feature extraction, thereby identifying urban residential areas in the image.

[0074] Furthermore, based on superpixel wavelet fusion, feature fusion of data from different sources is performed from the perspective of superpixels of different scales. The absolute value maximum method is used to fusion the high-frequency features, and the coefficient weighting strategy is used to fusion the low-frequency features.

[0075] Furthermore, the OSTU algorithm is used to divide the grayscale histogram of the residential area into two parts: foreground and background, so that the variance between the two parts is maximized and the optimal threshold is obtained for extracting urban residential areas.

[0076] In this embodiment, superpixel segmentation and discrete cosine wavelet transform are combined, the spatial distribution law of high-resolution images is fully considered, the spatial information of urban residential areas is embedded in the wavelet transform, and a superpixel wavelet fusion method is designed, which provides a scientific and efficient way for urban residential space extraction; combined with the OSTU algorithm, the grayscale histogram of the residential area image is divided into foreground and background parts based on the image grayscale value distribution characteristics, so that the variance between the two parts is maximized, that is, the separability is maximized, and the main information of the residential area is obtained on the binary image, thereby improving the accuracy of urban residential area extraction.

[0077] SLIC image super-segmentation refers to a pixel area with certain visual significance composed of adjacent pixels with similar color, brightness, texture and other characteristics, that is, from small to large, using this "large" "new" pixel as the basic unit of other image processing algorithms. The clustered area retains the original effective features of the image, and the boundary information in the original image is generally not destroyed. As a preprocessing step for image segmentation, the image processed by the superpixel algorithm uses a small amount of information instead of a large number of pixels to express image features, reducing the image dimension, eliminating the influence of abnormal pixels, greatly reducing the complexity of image processing, and improving processing efficiency.

[0078] Commonly used superpixel algorithms include graph-based algorithms, gradient ascent-based methods, and Simple Linear Iterative Clustering (SLIC) algorithms. Different superpixel algorithms have their own advantages and disadvantages. Compared with other superpixel segmentation methods, the SLIC algorithm has the advantages of fast running speed, higher memory efficiency, good contour preservation of generated superpixels, and fewer parameters to be set. This study uses the SLIC algorithm for superpixel segmentation.

[0079] SLIC, or simple linear iterative clustering, is an improvement on the K-means algorithm. It converts the color image into a 5-dimensional feature vector in CIELAB color space and XY coordinates, namely [L, A, B, X, Y], and then constructs a distance metric for the 5-dimensional feature vector and performs local clustering. The key point in the SLIC algorithm is the calculation of each superpixel center point and the clustering of pixels. The steps for calculating the superpixel center point are as follows:

[0080] 1) Assuming the image size is M×N and the number of superpixels is k, the image is evenly divided into k superpixel blocks, so the length and width of each superpixel block can be defined as S. The center point is (S / 2,S / 2);

[0081] 2) The center point of each superpixel block may be at a noise point or pixel mutation point. In order to reduce this probability, the gradient is calculated by differential method. The point with the smallest gradient value is the new center point. The differential gradient calculation formula is as follows:

[0082] G(x,y)=d x (i,j)+d y (i,j)

[0083] d x (i,j)=I(i+1,j)-I(i,j)

[0084] d y (i,j)=I(i,j+1)-I(i,j)

[0085] Among them, G(x,y) is the gradient value of the pixel at position (i,j), and I(i,j) is the value of the pixel at position (i,j).

[0086] The steps of pixel clustering are as follows:

[0087] 1) Assign a cluster label to each pixel in the neighborhood around each center point. The neighborhood range is 2S×2S, which is twice the expected superpixel size:

[0088] 2) Calculate the color distance and spatial distance between the pixel point and the center point within the search range. The formula is as follows:

[0089]

[0090] Among them, d c is the color distance between the center point of the jth superpixel block and the i-th pixel in the corresponding search range, d s is the corresponding spatial distance, D is the final distance, and m is the possible maximum value of the LAB spatial distance, ranging from [1,40], and generally 10;

[0091] 3) Each pixel corresponds to multiple superpixel block centers, so there are multiple distances. The center point corresponding to the minimum value is taken as the center of the pixel;

[0092] 4) After completing one iteration, recalculate the center point coordinates of each superpixel block and perform 10 iterations again. In the segmentation process of residential area images, the best value of k is obtained through multiple experiments.

[0093] Compared with Fourier transform, discrete cosine transform and wavelet transform are more flexible and have stronger adaptability. They can decompose images into different resolutions, making it easier to process images later.

[0094] The wavelet transform generates the corresponding scale and displacement function through a mother wavelet ψ(t), which is defined as follows:

[0095]

[0096] Among them, ψ a,b (t) is the wavelet basis function, which represents the conjugate of the wavelet basis with scale a and time shift b. a and b represent the scale factor and the shift factor respectively, and t represents the range in the time domain.

[0097] The wavelet transform coefficients W(a,b) of the signal f(t) can be defined as follows:

[0098] W(a,b)=∫ R ψ a,b (t)f(t)dt

[0099] The inverse wavelet transform of f(t) can be defined as follows:

[0100]

[0101] Where: C ψ is the constant of the wavelet function;

[0102] Continuous wavelet transform requires that a and b are continuous variables and f(t) is also a continuous function. However, image is a digital signal and it is difficult to meet this condition, resulting in high redundancy of wavelet coefficients.

[0103] In order to solve the shortcomings of continuous wavelet transform, discrete cosine transform appears. Compared with Fourier transform, a key advantage of discrete cosine transform is that it can capture both frequency and position information (time position) at the same time. Restrict a and b to a certain range and perform discrete processing on them, specifically:

[0104] 1) Discretization of a, specifically: processing a into a power series, that is, a>0, the corresponding wavelet function is j=0,1,2,…。

[0105] 2) Uniform discretization of b, so that ψ a,b (t) becomes:

[0106]

[0107] The discrete cosine transform is defined as:

[0108] W(a j ,kb)=∫ψ ajkb (t)f(t)dt,j=0,1,2,…,k.

[0109] Based on SLIC superpixel segmentation and discrete cosine transform theory, this paper designs a multi-source data fusion method of superpixel wavelet transform (SWT). Specifically, firstly, the SLIC algorithm is used to perform superpixel segmentation on the high-resolution remote sensing image of the Guangdong-Hong Kong-Macao Greater Bay Area (GBA area), and 800, 400, and 200 superpixel blocks ( Figure 3 ). Then, the discrete cosine transform is used to perform wavelet transform on the data from different sources to obtain high-frequency and low-frequency features at three scales. Next, the high-frequency and low-frequency features are fused based on the classification of data from different sources based on superpixel perspectives at different scales. Among them, the absolute value maximum method is used for high-frequency features, and the coefficient weighting strategy is used for low-frequency features ( Figure 3 ). The definition of the fusion process is as follows:

[0110] In discrete cosine transform, Mallat algorithm is one of the most commonly used decomposition algorithms. After the image is decomposed by Mallat algorithm, it can be expressed as:

[0111]

[0112] Where: H r and H c Represents a high-pass filter; G r and Gc represents a low-pass filter; r and c represent the rows and columns of the image; C j+1 Indicates the low-frequency part, which can also be expressed as LL; They represent the edge details of the image in the x, y, and xy directions, that is, the high-frequency parts, and they can be represented by LH, HL, and HH respectively.

[0113] The two-dimensional Mallat algorithm for image reconstruction is:

[0114]

[0115] Here, the two original images to be fused are A and B, and the fused image is F. The steps of image fusion are as follows: First, the two-dimensional Mallat algorithm of wavelet transform is used to decompose the two images A and B to obtain the low-frequency sub-band and high-frequency band coefficients of the image. Then, different rules are used to fuse the sub-band coefficients of different levels of the image. Finally, the two-dimensional Mallat algorithm is used to reconstruct the fused wavelet coefficients to obtain the fused image F.

[0116] In the image fusion process of wavelet transform, the design of fusion rules plays an important role, because the more favorable the wavelet coefficient is to the image fusion result, the more information it provides for the wavelet coefficient. There are two current image fusion rules: pixel fusion rule and window fusion rule. Among them, the pixel fusion rule is very simple, but the fusion result is not ideal; the window fusion rule takes into account the correlation of adjacent pixels, which can obtain rich detail information and better image vision. For this reason, this paper chooses the window fusion rule to realize image fusion.

[0117] Suppose that after images A and B are decomposed by N layers of wavelet transform, their lowest frequency subband numbers are and and Respectively represent their highest frequency subband coefficients, (x, y) is the coordinate of the coefficient, and Respectively represent the coefficients of image fusion, then for the lowest frequency subband coefficient and The fusion rules are as follows:

[0118]

[0119] In the formula, α1+α2=1.

[0120] For the highest frequency subband coefficient and You can get:

[0121]

[0122] The OTSU algorithm, also known as the maximum inter-class variance method (Otsu algorithm), is an adaptive threshold segmentation algorithm based on grayscale images, proposed by Japanese scholar Hideyuki Otsu in 1979. The algorithm divides the grayscale histogram of the residential area image into foreground and background parts based on the grayscale value distribution characteristics of the image. Its purpose is to maximize the variance between the two parts, that is, maximize the separation, so that the main information of the residential area can be displayed on the binary image.

[0123] First, standardize an image and map its pixel values ​​to 0-255, then count the number of pixels of each gray value. Let n i is the total number of pixels with grayscale i in the image, then we can get n0,n1,…,n i ,…,n 255 The probability that pixel i in the image is: And there is

[0124] Assuming the threshold t is known, the variance between class A and class B can be determined.

[0125] For class A, the probability that a pixel is assigned to class A is: The average pixel gray value of class A pixels is Similarly, the probability that a pixel is assigned to class B is The average pixel grayscale value of class B pixels is And the average gray value of the pixels of the entire image is

[0126] Then we can get:

[0127] P A (t)+P B (t) = 1

[0128] P A (t)×m A (t)+P B (t)×m B (t) = m G

[0129] And the overall variance of class A and class B is:

[0130] σ 2 =P A (t)(m A (t)-m G ) 2 +P B (t)(m B (t)-m G ) 2

[0131] Therefore, the inter-class variance is expressed as a function of the segmentation threshold t. It is only necessary to solve t so that the variance σ 2 The maximum value is sufficient. To do this, we can let t traverse [0-255] and find the value that makes σ 2 The largest t.

[0132] Figure 5 This is a schematic diagram of the accuracy verification results of different fused data proposed in this embodiment. The POI data, night light data and LandScan data of the Guangdong-Hong Kong-Macao Greater Bay Area in 2020 are fused through superpixel wavelet transform to obtain the POI_NTL, POI_LDS, and POI_NTL_LDS fused data sets, respectively. The OTSU algorithm is used to automatically calculate the classification threshold of the grayscale images of the three fused data sets. After the image is standardized, the optimal threshold is determined according to the calculation results, and the probability distribution map of the residential area in the Guangdong-Hong Kong-Macao Greater Bay Area is extracted according to the optimal threshold. Then, referring to the Google earth remote sensing image of the same phase, 3,000 verification points are manually marked within the residential area, and combined with relevant urban planning drawings and urban cadastral maps to determine whether the verification points are residential areas. Finally, the overall accuracy, precision, recall rate and F1 score are used in the quantitative accuracy evaluation to verify the extraction results of the three different fused data. The results are as follows Figure 5 As shown in the figure, the overall accuracy of residential areas extracted by NLT-POI data is 0.8152, the precision is 0.9228, the recall is 0.6973, and the F1 score is 0.7943; the overall accuracy of residential areas extracted by LDS-POI data is 0.7770, the precision is 0.8188, the recall is 0.7245, and the F1 score is 0.7688; the overall accuracy of residential areas extracted by POI_NTL_LDS data is 0.9040, the precision is 0.9563, the recall is 0.8514, and the F1 score is 0.9008. The results show that the residential areas extracted by POI_NTL_LDS data have higher accuracy.

[0133] In view of the problems that urban residential space identification relies on traditional data, has a large workload and low identification efficiency, the present invention proposes a method for urban residential space identification based on superpixel wavelet transform. The technical solution of the present invention has the following advantages: the present invention uses SLIC image super-segmentation to pre-process remote sensing images; uses discrete cosine transform to perform superpixel segmentation of low-frequency information and high-frequency information of the image in different methods, and realizes feature fusion of different data through superpixel wavelet transform, which makes up for the defects of traditional methods such as high time cost, low accuracy, and difficulty in replication and promotion; and obtains the best residential area extraction range through OTSU adaptive threshold calculation. This method can obtain the extraction results of residential areas more quickly and accurately, which is of great significance for the application of multi-source data in urban space research.

[0134] Example 3

[0135] This embodiment proposes a system for identifying urban residential spaces through intelligent learning and fusion of multi-source data, which is used to implement a method for identifying urban residential spaces through intelligent learning and fusion of multi-source data proposed in Embodiment 1.

[0136] The urban residential space recognition system for multi-source data intelligent learning and fusion includes:

[0137] An image acquisition module is used to acquire a number of urban space remote sensing images of a target area from a number of sources;

[0138] A superpixel segmentation module is used to perform superpixel segmentation on a plurality of urban space remote sensing images to obtain superpixel images corresponding to the plurality of urban space remote sensing images;

[0139] A feature extraction module is used to extract low-frequency features and high-frequency features of several super-pixel images;

[0140] A feature fusion module is used to fuse the low-frequency features of several super-pixel images with the low-frequency features to obtain fused low-frequency features, and to fuse the high-frequency features of several super-pixel images with the high-frequency features to obtain fused high-frequency features;

[0141] An image reconstruction module, used to reconstruct an image based on the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image;

[0142] The residential space recognition module is used to recognize the residential space area of ​​the fused urban space remote sensing image by using an adaptive threshold segmentation algorithm.

[0143] This embodiment proposes a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the urban residential space identification method through intelligent learning and fusion of multi-source data described in Example 1.

[0144] It can be understood that the urban residential space identification system and computer equipment for multi-source data intelligent learning and fusion in this embodiment improves the method of embodiment 1. The optional items in the above embodiment 1 are also applicable to this embodiment, so they will not be repeated here.

[0145] The same or similar reference numerals correspond to the same or similar components;

[0146] The terms used to describe the positional relationship in the drawings are only used for illustrative purposes and should not be construed as limiting the present embodiment;

[0147] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for identifying urban residential space by intelligent learning and fusion of multi-source data, characterized in that: The following steps are involved: Acquire several urban space remote sensing images of the target area from several sources; Perform super-pixel segmentation on several urban space remote sensing images to obtain super-pixel images corresponding to several urban space remote sensing images; Extract low-frequency features and high-frequency features of several super-pixel images; The low-frequency features of several super-pixel images are fused with the low-frequency features to obtain fused low-frequency features, and the high-frequency features of several super-pixel images are fused with the high-frequency features to obtain fused high-frequency features; Reconstructing an image based on the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image; The residential space area of ​​the fused urban space remote sensing image is identified using an adaptive threshold segmentation algorithm.

2. The urban residential space identification method based on multi-source data intelligent learning and fusion according to claim 1 is characterized in that: The sources of the urban space remote sensing images include POI data, NTL data, HRI data and LDS data.

3. The urban residential space identification method based on multi-source data intelligent learning and fusion according to claim 1 is characterized in that: When performing superpixel segmentation on several urban space remote sensing images, SLIC image superpixel segmentation is performed on each urban space remote sensing image with a preset number of segmentation blocks to obtain a superpixel image corresponding to each urban space remote sensing image, wherein one urban space remote sensing image corresponds to one superpixel image, each superpixel image is composed of several superpixel blocks of the same size, and the total number of superpixel blocks in each superpixel image is the preset number of segmentation blocks.

4. The urban residential space identification method based on multi-source data intelligent learning and fusion according to claim 1 is characterized in that: When extracting the low-frequency features and high-frequency features of several super-pixel images, the super-pixel images are decomposed using wavelet transform based on discrete cosine transform to obtain the lowest frequency sub-band number and the highest frequency sub-band coefficient of each super-pixel image, where the lowest frequency sub-band number is the low-frequency feature, and the highest frequency sub-band coefficient is the high-frequency feature.

5. The urban residential space identification method based on multi-source data intelligent learning and fusion according to claim 1 is characterized in that: When the low-frequency features of several super-pixel images are fused with the low-frequency features, the low-frequency features are fused using a coefficient weighting strategy to obtain fused low-frequency features; The expression of the coefficient weighting strategy includes: α1+α2+…+α n ＝1 Where n represents the total number of superpixel images, represents the result of fusing the low-frequency features of n super-pixel images with the low-frequency features, α n represents the scale factor of the nth superpixel image, Represents the low-frequency features of the nth superpixel image.

6. The urban residential space identification method based on multi-source data intelligent learning and fusion according to claim 1 is characterized in that: When the high-frequency features of several super-pixel images are fused with the high-frequency features, the high-frequency features are fused using the absolute value maximum method to obtain the fused high-frequency features; The expression of the absolute value maximum method includes: Where n represents the total number of superpixel images, It represents the result of fusing the high-frequency features of n super-pixel images with the high-frequency features. Represents the high-frequency features of the nth superpixel image.

7. The urban residential space identification method based on multi-source data intelligent learning and fusion according to claim 1 is characterized in that: When reconstructing an image based on the fused low-frequency features and the fused high-frequency features, a two-dimensional Mallat algorithm is used to reconstruct the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image.

8. The urban residential space identification method according to any one of claims 1 to 7, characterized in that: When the residential space area of ​​the fused urban space remote sensing image is identified by using an adaptive threshold segmentation algorithm, the fused urban space remote sensing image is divided into a residential space area and a non-residential space area by using a maximum inter-class variance method; Here, let the probability that the pixels of the fused urban space remote sensing image are assigned to the residential space area be P A (t), let the probability that the pixels of the fused urban space remote sensing image are assigned to the non-residential space area be P B (t), then P A (t) and P B The population variance σ of (t) 2 The expressions include: σ 2 =P A (t)(m A (t)-m G ) 2 +P B (t)(m B (t)-m G ) 2 P A (t)+P B (t)=1 Where t represents the segmentation threshold, P i represents the i-th pixel of the fused urban space remote sensing image, m A (t) represents the average pixel grayscale value of pixels belonging to the residential area, m B (t) represents the average pixel grayscale value of pixels in the non-residential space area, m G Represents the average gray value of the pixels of the entire fused urban space remote sensing image; When dividing the fused urban space remote sensing image into residential space areas and non-residential space areas, we use the command P A (t) and P B The population variance σ of (t) 2 The maximum segmentation threshold t is used for segmentation.

9. A system for identifying urban residential space by intelligent learning and fusion of multi-source data, used to implement the method for identifying urban residential space by intelligent learning and fusion of multi-source data as claimed in any one of claims 1 to 8, characterized in that: include: An image acquisition module is used to acquire a number of urban space remote sensing images of a target area from a number of sources; A superpixel segmentation module is used to perform superpixel segmentation on a plurality of urban space remote sensing images to obtain superpixel images corresponding to the plurality of urban space remote sensing images; A feature extraction module is used to extract low-frequency features and high-frequency features of several super-pixel images; A feature fusion module is used to fuse the low-frequency features of several super-pixel images with the low-frequency features to obtain fused low-frequency features, and to fuse the high-frequency features of several super-pixel images with the high-frequency features to obtain fused high-frequency features; An image reconstruction module, used to reconstruct an image based on the fused low-frequency features and the fused high-frequency features to obtain a fused urban space remote sensing image; The residential space recognition module is used to recognize the residential space area of ​​the fused urban space remote sensing image by using an adaptive threshold segmentation algorithm.

10. A computer device comprising a memory and a processor, wherein the memory stores computer-readable instructions, characterized in that: When the computer-readable instructions are executed by the processor, the processor executes the steps of the urban residential space identification method based on multi-source data intelligent learning and fusion as claimed in any one of claims 1 to 8.

Images