Urban land utilization automatic classification method and system

Abstract

The invention provides an automatic classification method and system for urban land utilization. The method comprises the following steps: making a large-range land utilization training sample based on open-source crowdsourcing geographic data; constructing a context deep learning semantic segmentation network of end-to-end uncertainty perception, and obtaining a pixel-level land utilization classification product; and segmenting a remote sensing image to be classified based on a bottom-up multi-scale segmentation algorithm, and fusing a pixel-level land utilization classification result with an object-level segmentation boundary based on mode voting to obtain an object-level land utilization classification product. The land utilization types are automatically classified based on deep learning, the classification precision is improved, and the labor cost is saved.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing data processing, and in particular relates to an automatic classification method and system for urban land use. Background Art

[0002] Urban land use refers to the way humans use and manage the earth's surface, covering various land use patterns such as agriculture, urban construction, industrial development, and recreational activities. It is the result of the interaction between human social activities and the natural environment, reflecting how humans allocate and use surface resources to meet their social, economic, and cultural needs. Land use information is essential for understanding the impact of humans on the environment, planning urban and rural development, managing natural resources, and formulating sustainable development strategies. Therefore, the classification and continuous updating of land use information is of great significance for promoting social progress and protecting the natural environment.

[0003] Traditional land use classification technology relies on manual annotation of a large number of samples, which is time-consuming and laborious. With the emergence of open source geographic data OSM, detailed semantic labels of geographic locations are provided for large-scale land use classification. However, deep learning networks are highly dependent on the accuracy of reference labels, and since OSM is provided by volunteers, its labels and ranges are inaccurate. Therefore, a deep network that can learn effective land use knowledge from noisy labels is needed. In addition, due to the large number of pooling operations of deep convolutional neural networks, their pixel-level classification results are inaccurate at the edges, the classification results are fragmented, and cannot reflect the true spatial pattern of land use. Summary of the invention

[0004] In order to overcome the problems existing in the above-mentioned existing land use classification methods, the present invention proposes an automatic urban land use classification method to solve the problem that the existing deep network cannot effectively process the noise labels from open crowd-source geographic data, as well as the problems of inaccurate edges and fragmented classification results of pixel-level classification results.

[0005] According to a first aspect of an embodiment of the present invention, a method for extracting land use types is provided, comprising:

[0006] Acquire Sentinel-2 remote sensing image data, create corresponding OSM land use type training samples, and crop and enhance the training samples;

[0007] Constructing a deep convolutional semantic segmentation network model, and training the deep learning semantic segmentation network model through land use type training samples;

[0008] A deep convolutional network semantic segmentation model with an encoder-decoder structure is established, and an uncertainty perception module is introduced to learn useful land use knowledge from noise samples.

[0009] Establish bottom-up multi-scale segmentation to extract precise boundary information from remote sensing images.

[0010] Fusion of deep learning semantic segmentation network model with multi-scale resolution segmentation to generate an end-to-end object-level land use classification model;

[0011] The remote sensing image data to be classified is input into the end-to-end land use classification model to obtain object-level land use classification products.

[0012] Preferably, the acquisition of Sentinel-2 remote sensing image data, preparation of corresponding OSM land use type training samples, and cropping and data enhancement of the training samples are specifically as follows:

[0013] Download the cloud-processed and radiometrically corrected Sentinel-2 data from Google Earth Engine, download the vector data from OSM, clean the labels, and leave the labeled vector data related to land use attributes. Then, according to the one-to-one correspondence between the labels and the land use categories, rasterize the vector data to obtain the raster sample data.

[0014] Slice the remote sensing data images and their corresponding land use samples to generate land use type training samples of standard size;

[0015] Perform image enhancement on land use sample data and remote sensing images of standard size, including rotation, cropping, contrast enhancement, etc.;

[0016] Preferably, the deep convolutional semantic segmentation network includes a feature encoder, a context feature extraction and an uncertainty-aware decoder part;

[0017] The feature encoder part adopts the pre-trained ResNet-34, which includes two convolutional layers and one maximum pooling layer;

[0018] The context feature extraction part, including dense hole convolution blocks and residual multi-core pooling blocks, is responsible for extracting context semantic information and generating higher-level feature maps.

[0019] The uncertainty-aware decoder part includes two encoders, namely the main classifier and the auxiliary classifier. Compared with the main classifier, the receptive field size of the auxiliary classifier is halved. Given a 3D input sample X∈R H×W×C , the two decoders are derived from the encoder θ e Receive the same deep feature F e , and then respectively by the main decoder P prim and auxiliary decoder F aux Generate predicted class probabilities and It can be formulated as:

[0020]

[0021] To solve the label noise problem, the prediction variance is calculated to model the uncertainty. The Kullback-Leibler (KL) divergence is approximated as the variance of the noise label and the true label, which is used to measure the difference between the outputs P of the two classifiers. prim and P aux The distance between them is calculated as follows:

[0022]

[0023] Among them, p t is the ground truth class probability of the target data. In the loss function, a variance regularization term is used to correct the results learned from noisy labels as shown below:

[0024]

[0025] Where L ce represents CEloss. If the OSM label annotation is inaccurate, the two classifiers will produce different results, resulting in D kl This ensures that noisy labels are not learned first. t ) is very small, that is, the prediction results of the main classifier and the auxiliary classifier are very close, the model learning will proceed normally. By adding auxiliary classifiers for uncertainty modeling, the network can learn to distinguish between clean samples and noisy samples and assign different weights to their predictions. This helps the model reduce the impact of noisy labels and focus more on reliable information.

[0026] Preferably, the bottom-up multi-scale segmentation module is specifically:

[0027] Initialize and generate superpixels, and use simple linear iterative clustering to locally aggregate pixels into image objects of extremely small scale. The pixels inside each superpixel have similar color, brightness or texture features. The formula for local clustering to calculate the distance from a pixel to the surrounding seed points is as follows:

[0028]

[0029] d c，j，i =(l j -l i ) 2 +(a j -a i ) 2 +(b j -b i ) 2 (5)

[0030] ds，j，i =(x j -x i ) 2 +(y j -y i ) 2 (6)

[0031] In the formula, dj,i represents the distance from the jth pixel to the ith seed point in the 2S×2S neighborhood, M is used to adjust the weight between the spectral distance and the spatial distance, which is generally 10, and dj,i is the distance between the jth pixel and the ith seed point in the 2S×2S neighborhood. c，j，i , d s，j，i represents the spectral distance and spatial distance, l j , a j , b j , x j ,y j represents the spectral value and spatial position of the j-th pixel in the CIELAB color space in the 2S×2S neighborhood, l i , a i , b i , x i ,y i represents the spectral value and spatial position of the i-th pixel in the CIELAB color space. Finally, the label of the j-th pixel is assigned as label j = argmin({d j，i |i∈[0,K)});

[0032] The computational heterogeneity h between two objects based on the spectrum, texture, brightness and other characteristics of superpixels is calculated as follows:

[0033] h=ω color h color +(1-ω color )h shape (7)

[0034]

[0035] h shape =ω comlpact h compact +(1-ω compact )h smooth (9)

[0036] h smooth =n m s m -(n obj1 s obj1 +n obj2 s obj2 ) (10)

[0037] h compact =nm c m -(n obj1 c obj1 +n obj2 c obj2 ) (11)

[0038]

[0039]

[0040] In the formula, h color and h shape represents the spectral heterogeneity and shape heterogeneity between two objects, ω color represents the spectral heterogeneity weight, B represents the number of image bands, ω b Indicates the weight of each band involved in the segmentation, n m and σ b，m represents the number of pixels and spectral standard deviation of the merged object, n obj1 , σ b，obj1 , n obj1 , σ b，obj2 They represent the number of pixels and the spectral standard deviation of the two objects before merging, respectively. compact and h smooth represents the compactness heterogeneity and smoothness heterogeneity between two objects, ω compact represents the compactness heterogeneity weight, s m ,s obj1 ,s obj2 Indicates the smoothness of the merged object and the two objects before and after merging, c m , c obj1 , c obj2 Indicates the compactness of the merged object and the two objects before and after the merge, l and b respectively represent the perimeter of the object and the shortest side length of the object's minimum enclosing rectangle. In the multi-resolution segmentation algorithm, three parameters need to be set, namely the segmentation scale, the spectral heterogeneity weight w color , compactness heterogeneity weight w compact ;

[0041] Perform bottom-up superpixel merging based on heterogeneity to obtain segmented objects;

[0042] Preferably, the deep learning semantic segmentation network model and the multi-scale resolution segmentation fusion module are specifically:

[0043] The pixel-level classification results and the object-level segmentation results are subjected to majority voting, and the number of each category in each object is counted based on the object-level segmentation.

[0044] The second aspect of the present invention is an automatic classification system for urban land use, the system comprising:

[0045] Sample production module: obtain Sentinel-2 remote sensing image data and corresponding OSM-based land use type training samples;

[0046] Deep convolutional network semantic segmentation module: construct a deep learning semantic segmentation network model, and train the deep learning semantic segmentation network model through land use type training samples;

[0047] Multi-scale segmentation module: Establish bottom-up multi-scale segmentation to extract accurate boundary information from remote sensing images.

[0048] Model fusion module: Fusion of deep learning semantic segmentation network model with multi-scale resolution segmentation to generate an end-to-end object-level land use classification model;

[0049] Classification and recognition module: inputs the remote sensing image data to be classified into the end-to-end land use classification model to obtain object-level land use classification products.

[0050] In addition, the present invention also provides a storage medium, wherein the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the method described.

[0051] A processor is used to run a program, wherein the method described is executed when the program is run.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1) The present invention constructs a process for automatically extracting land use samples, which can obtain land use samples for free based on OSM crowd-sourced geographic data without the need for manual labeling, which is time-consuming and labor-intensive;

[0054] 2) The present invention constructs a deep convolutional neural network based on uncertainty perception, and improves the accuracy and robustness of urban land use extraction in the presence of label noise by modeling the uncertainty of samples.

[0055] 3) The present invention combines the pixel-level deep semantic segmentation results with the image-based segmentation boundaries to form object-level land use classification results, thereby improving the accuracy of the land use classification results at the edges. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 A schematic diagram of the overall process of the land use type classification method provided by an embodiment of the present invention;

[0058] Figure 2 A schematic diagram of an uncertainty-aware deep convolutional semantic segmentation network provided by an embodiment of the present invention;

[0059] Figure 3 Schematic diagram of object-oriented multi-scale bottom-up segmentation;

[0060] Figure 4 Schematic diagram of the fusion of object-level segmentation and pixel-level classification based on majority voting. DETAILED DESCRIPTION

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] In one embodiment of the present invention, a method for extracting land use types is provided. Figure 1 The overall flow diagram of the land use type extraction method provided by the embodiment of the present invention includes:

[0063] S1, obtain remote sensing image data and its corresponding OSM land use data, and make land use samples and image pairs of standard size, specifically;

[0064] Download the cloud-processed and radiometrically corrected Sentinel-2 data from Google Earth Engine, download the vector data from OSM, clean the labels, and leave the labeled vector data related to land use attributes. Then, according to the one-to-one correspondence between the labels and the land use categories, rasterize the vector data to obtain the raster sample data.

[0065] Slice the remote sensing data images and their corresponding land use samples to generate land use type training samples of standard size;

[0066] Perform image enhancement on land use sample data and remote sensing images of standard size, including rotation, cropping, contrast enhancement, etc.;

[0067] S2, the deep convolutional semantic segmentation network includes feature encoder, context feature extraction and uncertainty-aware decoder parts, combined with Figure 2 , specifically;

[0068] The feature encoder part adopts the pre-trained ResNet-34, which includes two convolutional layers and one maximum pooling layer;

[0069] The context feature extraction part, including the dense hole convolution block and the residual multi-core pooling block, is responsible for extracting context semantic information and generating higher-level feature maps. The calculation formula of the dense hole convolution block is as follows:

[0070]

[0071] The input feature map x and the filter w are convolved to generate the output y. In this formula, x[i+rk]w[k] represents the sampling point of the input feature map, w[k] is the weight in the convolution kernel, and r is the atrous rate, which corresponds to the step size of the sampled input signal. Dense dilated convolution enhances the ability of feature extraction by using dilated convolutions of different scales, thereby more effectively capturing objects of different sizes in the image and their features.

[0072] The uncertainty-aware decoder part includes two encoders, namely the main classifier and the auxiliary classifier. Compared with the main classifier, the receptive field size of the auxiliary classifier is halved. Given a 3D input sample X∈R H×W×C , the two decoders are derived from the encoder θ e Receive the same deep feature F e , and then respectively by the main decoder P prim and auxiliary decoder F aux Generate predicted class probabilities and It can be formulated as:

[0073]

[0074] To solve the label noise problem, the prediction variance is calculated to model the uncertainty. The Kullback-Leibler (KL) divergence is approximated as the variance of the noise label and the true label, which is used to measure the difference between the outputs P of the two classifiers. prim and P aux The distance between them is calculated as follows:

[0075]

[0076] Among them, p t is the ground truth class probability of the target data. In the loss function, a variance regularization term is used to correct the results learned from noisy labels as shown below:

[0077]

[0078] Where L cerepresents CEloss. If the OSM label annotation is inaccurate, the two classifiers will produce different results, resulting in D kl This ensures that noisy labels are not learned first. t ) is very small, that is, the prediction results of the main classifier and the auxiliary classifier are very close, the model learning will proceed normally. By adding auxiliary classifiers for uncertainty modeling, the network can learn to distinguish between clean samples and noisy samples and assign different weights to their predictions. This helps the model reduce the impact of noisy labels and focus more on reliable information.

[0079] S3, the bottom-up multi-scale segmentation module is specifically:

[0080] Initialize and generate superpixels, and use simple linear iterative clustering to locally aggregate pixels into image objects of extremely small scale. The pixels inside each superpixel have similar color, brightness or texture features. The formula for local clustering to calculate the distance from a pixel to the surrounding seed points is as follows:

[0081]

[0082] d c，j，i =(l j -l i ) 2 +(a j -a i ) 2 +(b j -b i ) 2 (5)

[0083] d s，j，i =(x j -x i ) 2 +(y j -y i ) 2 (6)

[0084] In the formula, , represents the distance from the th pixel to the th seed point in the 2×2 neighborhood, which is used to adjust the weight between the spectral distance and the spatial distance, and is generally taken as 10. c，j，i , d s，j，i represents the spectral distance and spatial distance, l j , a j , b j , x j ,y j represents the spectral value and spatial position of the pixel in the CIELAB color space in a 2×2 neighborhood, l i , a i , b i , xi ,y i represents the spectral value of the th pixel in the CIELAB color space and its position in the space. Finally, the label of the th pixel is assigned as label j = argmin({d j，i |i∈[0,K)});

[0085] The computational heterogeneity h between two objects based on the spectrum, texture, brightness and other characteristics of superpixels is calculated as follows:

[0086] h=ω color h color +(1-ω color )h shape (7)

[0087]

[0088] h shape =ω compact h compact +(1-ωc ompact )h smooth (9)

[0089] h smooth =n m s m -(n obj1 s obj1 +n obj2 s obj2 ) (10)

[0090] h compact =n m c m -(n obj1 c obj1 +n obj2 c obj2 ) (11)

[0091]

[0092] In the formula, h color and h shape represents the spectral heterogeneity and shape heterogeneity between two objects, ω color represents the spectral heterogeneity weight, represents the number of bands of the image, ω b Indicates the weight of each band involved in the segmentation, n m and σ b，m represents the number of pixels and spectral standard deviation of the merged object, n obj1 , σ b，obj1 , n obj1 , σ b，obj2They represent the number of pixels and the spectral standard deviation of the two objects before merging, respectively. compact and h smooth represents the compactness heterogeneity and smoothness heterogeneity between two objects, ω compact represents the compactness heterogeneity weight, s m ,s obj1 ,s obj2 Indicates the smoothness of the merged object and the two objects before and after merging, c m , c obj1 , c obj2 Indicates the compactness of the merged object and the two objects before and after merging, and respectively represents the perimeter of the object and the shortest side length of the object's minimum enclosing rectangle. In the multi-resolution segmentation algorithm, three parameters need to be set, namely the segmentation scale, the spectral heterogeneity weight w color , compactness heterogeneity weight w compact ;

[0093] Perform bottom-up superpixel merging based on heterogeneity to obtain segmented objects;

[0094] S4, the deep learning semantic segmentation network model and the multi-scale resolution segmentation fusion module are specifically:

[0095] The pixel-level classification results and the object-level segmentation results are subjected to majority voting, and the number of each category in each object is counted on the basis of the object-level segmentation.

[0096] Corresponding to the above method embodiment, the present invention also discloses an object-oriented deep convolutional neural land use classification system, the system comprising:

[0097] Sample production module: obtain Sentinel-2 remote sensing image data and corresponding OSM-based land use type training samples;

[0098] Deep convolutional network semantic segmentation module: construct a deep learning semantic segmentation network model, and train the deep learning semantic segmentation network model through land use type training samples;

[0099] Multi-scale segmentation module: Establish bottom-up multi-scale segmentation to extract accurate boundary information from remote sensing images.

[0100] Model fusion module: Fusion of deep learning semantic segmentation network model with multi-scale resolution segmentation to generate an end-to-end object-level land use classification model;

[0101] Classification and recognition module: inputs the remote sensing image data to be classified into the end-to-end land use classification model to obtain object-level land use classification products.

[0102] In addition, the present invention also provides a storage medium, wherein the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the method described.

[0103] A processor is used to run a program, wherein the method described is executed when the program is run.

[0104] A person of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various media that can store program codes.

[0105] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic classification method for urban land use, characterized in that: The following steps are involved: (1) Based on OpenStreetMap (OSM) open source geographic data, sample data related to land use are obtained, and land use type training samples and corresponding Sentinel-2 image data are downloaded; (2) Construct a deep learning semantic segmentation network model and train the deep learning semantic segmentation network model through land use type training samples; (3) Segment the image based on a bottom-up multi-scale segmentation algorithm to generate object-level segmentation results; (4) The pixel-level land use classification results obtained by the deep learning semantic segmentation network are integrated with the object-level segmentation results, and the object-level land use classification results are obtained using the majority voting mechanism.

2. The automatic classification method for urban land use according to claim 1, characterized in that: Step (1) specifically includes the following sub-steps: (1.1) Based on the city boundary data, the urban built-up area is obtained for clipping, and then the OSM vector data and the corresponding Sentinel-2 image slices within the clipping range are obtained. The OSM vector data is vector-rasterized to obtain the corresponding sample labels with land use attributes; (1.2) Slice the remote sensing image data and its corresponding land use sample labels to generate standard-scale land use type training samples.

3. The automatic classification method for urban land use according to claim 1, characterized in that: The network model in step (2) includes an encoding part, a context extraction module and a decoding part; In the coding part, the uncertainty perception module is introduced to improve the accuracy and robustness of urban land use classification in the presence of label noise; The uncertainty perception module is composed of two different decoders in the feature decoder, namely the main classifier and the auxiliary classifier; Compared with the main classifier, the receptive field size of the auxiliary classifier is halved; Given a 3D input sample X∈R H ×W×C , the two decoders are derived from the encoder θ e Receive the same deep feature F e , and then respectively by the main decoder P prim and auxiliary decoder F aux Generate predicted class probabilities and It is formally described as: To solve the label noise problem, the prediction variance is calculated to model the uncertainty; the Kullback-Leibler divergence is approximated as the variance of the noise label and the true label, which is used to measure the difference between the outputs P of the two classifiers. prim and P aux The distance between them is calculated as follows: Among them, p t is the ground truth class probability of the target data; in the loss function, a variance regularization term is used to correct the result of learning from noisy labels, as shown below: Where L ce represents CEloss; if the OSM label annotation is inaccurate, the two classifiers will produce different results, resulting in D kl The value increases.

4. The automatic classification method for urban land use according to claim 1, characterized in that: Step (4) specifically includes: Initialize and generate superpixels, and use simple linear iterative clustering to locally aggregate pixels into image objects with extremely small scales. The pixels inside each superpixel have similar color, brightness or texture features. The formula for local clustering to calculate the distance from a pixel to the surrounding seed points is as follows: d cj，i =(l j -l i ) 2 +(a j -a i ) 2 +(b j -b i ) 2 (5) d s，j，i =(x j -x i ) 2 +(y j -y i ) 2 (6) Where, d j,i represents the distance from the jth pixel to the ith seed point in the 2S×2S neighborhood. M is used to adjust the weight between the spectral distance and the spatial distance and takes 10. c，j，i , d s，j，i represents the spectral distance and spatial distance, l j , a j , b j , x j ,y j represents the spectral value and spatial position of the j-th pixel in the CIELAB color space in the 2S×2S neighborhood, l i , a i , b i , x i ,y i represents the spectral value and spatial position of the i-th pixel in the CIELAB color space; finally, the label of the j-th pixel is assigned as label j = argmin({d j，i ||i∈[0,K}); The computational heterogeneity h between two objects based on the spectrum, texture, brightness and other characteristics of superpixels is calculated as follows: h=ω color h color +(1-ω color )h shape (7) h shape =ω compact h hcompact +(1-ω compact )h smooth (9) h smooth =n m s m -(n obj1 s obj1 +n obj2 s obj2 ) (10) h compact =n m c m -(n obj1 c obj1 +n obj2 c obj2 ) (11) In the formula, h color and h shape represents the spectral heterogeneity and shape heterogeneity between two objects, ω color represents the spectral heterogeneity weight, B represents the number of image bands, ω b Indicates the weight of each band involved in segmentation, s m and σ b，m represents the number of pixels and spectral standard deviation of the merged object, n obj1 , σ b，obj1 , n obj1 , σ b，obj2 They represent the number of pixels and the spectral standard deviation of the two objects before merging, respectively. compact and h smooth represents the compactness heterogeneity and smoothness heterogeneity between two objects, ω compact represents the compactness heterogeneity weight, s m ,s obj1 ,s obj2 Indicates the smoothness of the merged object and the two objects before and after merging, c m , c obj1 , c obj2 Indicates the compactness of the merged object and the two objects before and after the merge, l and b respectively represent the perimeter of the object and the shortest side length of the object's minimum enclosing rectangle; in the multi-resolution segmentation algorithm, three parameters are set, namely, the segmentation scale, the spectral heterogeneity weight w color , compactness heterogeneity weight w compact ; Bottom-up superpixel merging is performed based on heterogeneity, and finally object-level boundaries are obtained.

5. The automatic classification method for urban land use according to claims 3 and 4, characterized in that: The pixel-level classification results and the object-level segmentation results are subjected to majority voting to count the number of categories in each object, and the category with the largest number of pixels is the category of the object.

6. Urban land use automatic classification system, characterized by: The system comprises: Sample production module: obtain Sentinel-2 remote sensing image data and corresponding OSM-based land use type training samples; Deep convolutional network semantic segmentation module: construct a deep learning semantic segmentation network model, and train the deep learning semantic segmentation network model through land use type training samples; Multi-scale segmentation module: Establish bottom-up multi-scale segmentation to extract accurate boundary information from remote sensing images; Model fusion module: Fusion of deep learning semantic segmentation network model with multi-scale resolution segmentation to generate an end-to-end object-level land use classification model; Classification and recognition module: inputs the remote sensing image data to be classified into the end-to-end land use classification model to obtain object-level land use classification products.

7. A storage medium, characterized in that: The storage medium includes a stored program, wherein when the program is executed, the device where the storage medium is located is controlled to execute the method according to any one of claims 1 to 5.

8. A processor, characterized in that: The processor is used to run a program, wherein the program executes the method according to any one of claims 1 to 5 when running.

Images