CN115631412A - Remote sensing image building extraction method based on coordinate attention and data correlation upsampling - Google Patents

Abstract

The invention relates to a remote sensing image building extraction method based on coordinate attention and data correlation upsampling, which comprises the following steps: acquiring remote sensing data; data preprocessing and data enhancement; constructing a building extraction network model, namely a CAD-UNet network model, comprising an encoder, a coordinate attention CA module and a data-related up-sampling DUp module; training and evaluating a model; building automation extraction: and after data preprocessing is carried out on the new remote sensing image to be extracted, inputting the trained CAD-UNet network model, and outputting a predicted image by the CAD-UNet network model to obtain a building extraction result. The network designed by the invention gradually extracts the deep features of the building, performs feature fusion and then gradually upsamples the features to the size of the input resolution, is more friendly to the building extraction task, and obviously improves the building extraction precision; the position information and the boundary information of the building can be effectively captured, so that the extracted building has smoother boundaries and complete outlines.

Description

Building Extraction from Remote Sensing Images Based on Coordinate Attention and Data Correlation Upsampling method

technical field

The invention relates to the technical field of image processing, in particular to a method for extracting buildings from remote sensing images based on coordinate attention and data correlation upsampling.

Background technique

Buildings are an indispensable activity place in people's daily life, and also an important part of the process of urban construction and development. The main task of building extraction is to identify and extract building areas from remote sensing images. Building extraction is of great significance for smart city construction, traffic management, population estimation, and land use monitoring, etc. With the rapid development of remote sensing technology, remote sensing images began to transition from low resolution to high resolution, forming a development trend characterized by high spatial resolution, high spectral resolution, and high temporal resolution. The features and information of high-resolution remote sensing images continue to increase, and the noise and interference information also increase accordingly, which brings new challenges to building extraction. How to accurately extract buildings from high-resolution remote sensing images has become a research hotspot and difficult.

Traditional building extraction methods are usually based on prior knowledge and manual features, and then clustering and other algorithms are used to extract buildings, mainly including methods based on building features and auxiliary information. Most of these methods use the features of the building's shape, texture, and other methods to extract buildings based on auxiliary information. The realization principle is relatively simple, but there are problems such as low recognition rate and many errors, and the process is time-consuming and laborious. Large limitations, performance is also greatly limited. Specifically in:

First, less attention is paid to the location information of buildings; the location information of buildings is extremely important for building extraction tasks. Buildings are usually regularly distributed in a remote sensing image, and there are usually shadows such as trees, so focus on building The location information of the object can obtain the accurate location of the building in an image and avoid misclassification. The existing technology does not fully consider the location information of the building, especially for buildings with shadow occlusion and complex adhesion. Insufficient attention is paid to location information, so misclassification is easy to occur;

Second, the boundary of building extraction is rough and fuzzy; buildings are mostly rectangular and usually have regular boundaries. Boundary information is an important feature that cannot be ignored in the task of building extraction. When extracting buildings, if the edge information is ignored, it is easy to It leads to problems such as rough and blurred boundaries, chaotic boundaries, and holes. When extracting buildings in the existing technology, only conventional feature extraction is performed, and the edge features of buildings cannot be fully excavated, resulting in poor extraction results and rough boundaries. problems such as fuzzy;

Third, there is an imbalance between positive and negative samples; building extraction is a binary classification task, which mainly distinguishes buildings and backgrounds. However, under normal circumstances, there are more background pixels in remote sensing images than building pixels, which will weaken the model's ability to extract buildings during training. The existing technology fails to fully consider the problem of the imbalance between positive and negative samples, which leads to low accuracy of building extraction and weak generalization ability of the model.

Contents of the invention

The purpose of the present invention is to provide a coordinate-based attention that effectively solves the problem of misclassification caused by insufficient attention to building location information, optimizes the boundary extraction effect of buildings, alleviates the problem of unbalanced positive and negative samples, and improves the generalization ability of the network. Force- and data-dependent upsampling for building extraction from remote sensing images.

In order to achieve the above object, the present invention adopts the following technical solutions: a method for extracting buildings from remote sensing images based on coordinate attention and data correlation upsampling, the method includes the steps in the following order:

(1) Obtain remote sensing data: download WHU building dataset and Massachusetts building dataset;

(2) Data preprocessing and data enhancement: the preprocessing refers to cropping the large images in the data set, and performing data enhancement on the cropped remote sensing images and label images; The ratio of 8:1:1 is divided into training set, verification set and test set;

(3) Constructing the CAD-UNet network model: Based on the UNet network, improve and construct the building extraction network model including the encoder, the coordinate attention CA module, and the data-dependent upsampling DUp module, namely the CAD-UNet network model;

(4) Model training and evaluation: Based on the joint loss function of the training set data BCE Loss binary classification cross entropy loss and Focal Loss focal loss, the CAD-UNet network model is trained, and the test set is used to evaluate CAD-UNet after training Building extraction accuracy and effect of network model;

(5) Automatic extraction of buildings: after data preprocessing of the new remote sensing images to be extracted, the trained CAD-UNet network model is input, and the CAD-UNet network model outputs predicted images to obtain building extraction results.

Described step (2) specifically comprises the following steps:

(2a) By means of sliding window, crop the remote sensing large image and label image in the Massachusetts building dataset, cut it into a 512×512 image, and put the WHU building dataset and the Massachusetts building dataset in the label image into the building The pixel value of the background is marked as 1, and the pixel value of the background is marked as 0;

(2b) Carry out data enhancement to the remote sensing images and label images in the WHU building dataset and the cropped Massachusetts building dataset remote sensing images and label images to expand the amount of data. The data enhancement includes:

Horizontal flip: Use the image processing library OpenCV to flip the remote sensing image and the label image horizontally;

Vertical flip: Use the image processing library OpenCV to flip the remote sensing image and the label image vertically;

Horizontal and vertical flip: Use the image processing library OpenCV to flip the remote sensing image and the label image first horizontally and then vertically;

Shift, zoom, random crop and add noise: perform operations such as shift, zoom, random crop and add noise on remote sensing images and label images respectively;

(2c) divide the remote sensing image and label image after data enhancement into training set, verification set and test set according to the ratio of 8:1:1 respectively, and the training set is used to directly participate in the training of CAD-UNet network model, and carry out Feature extraction; the verification set is used to adjust the hyperparameters of the CAD-UNet network model; the test set is used to test the accuracy and extraction effect of the CAD-UNet network model after the training is completed.

Described step (3) specifically comprises the following steps:

(3a) Replace the UNet network encoder: replace the UNet network encoder with the VGG16 network module. The VGG16 network module is composed of the VGG16 network without the last pooling layer and the fully connected layer. The VGG16 network module passes multiple convolutions and four times Maximum pooling is used for down-sampling, extracting building features in remote sensing images, and outputting feature maps of four different scales;

(3b) Construct the coordinate attention CA module: embed the coordinate attention CA module into the skip connection of the UNet network obtained through step (3a);

The coordinate attention CA module captures long-range dependencies and retains location information along two spatial directions, and encodes the feature maps separately to form two feature maps, which are direction-aware and position-sensitive, respectively, and any intermediate tensor X=[x ₁ , x ₂ , x ₃ ..., x _C ]∈R ^C×H×W as input, and output a tensor of the same length Y=[y ₁ , y ₂ , y ₃ ..., y _C ] , specifically for the input X, use a pooling kernel of size (H, 1) and (1, W) to encode each channel along the horizontal and vertical directions, and the expression of the cth channel with a height of H as follows:

In the formula, H represents the height of the image, W represents the width of the image, C represents the total number of channels of the image, c represents the c-th channel, X _c represents the image of the c-th channel, i represents the abscissa of the image, and R represents the middle sheet volume collection;

Similarly, the output of the cth channel of width W is expressed as follows:

Formulas (1), (2) are two transformations of feature aggregation, which are aggregated along two spatial directions respectively, and return two-direction perceptual attention maps; the coordinate attention CA module generates two feature layers before cascading, Then share a 1×1 convolution operation to transform, as shown in formula (3):

f=δ(F ₁ ([Z ^H , Z ^W ])) (3)

In the formula, δ is a nonlinear activation function, and f is an intermediate feature map, which is the result of feature encoding of spatial information in both horizontal and vertical directions; then f is decomposed into two separate Tensor, f ^H ∈ R ^C/r×H and f ^W ∈ R ^C/r×W , r is used to control the reduction ratio of the number of channels, and then use two other 1×1 convolutions to transform F _H and F _W transforms f ^H and f ^W into two tensors g ^H and g ^W with the same number of feature layers, respectively:

g ^H ＝σ(F _H (f ^H )) (4)

g ^W ＝σ(F _H (f ^W ))(7)

In the formula, F _H and F _W are two 1×1 convolution transformations, f ^H and f ^W are two separate tensors obtained by decomposing f, g ^H and g ^W are convolution transformations and activation functions The obtained tensor, σ is the sigmoid activation function. During the transformation process, use the reduction ratio r to reduce the number of channels of f, and then expand the output g ^H and g ^W , respectively as attention weights; coordinate attention CA The final output of the module is shown in formula (6):

(3c) Construct a data-dependent upsampling DUp module: combine the convolutional layer and the data-dependent upsampling module to construct a data-dependent upsampling DUp module, and use it to extract the boundary information of high-resolution buildings. For the four different scale features of the input Figure, first pass through a 3×3 convolutional layer to reduce the number of channels of the feature map; then perform data-related upsampling to directly restore the feature map to a size of 512×512, and perform point-by-point processing on the four feature maps obtained after upsampling After addition and fusion, output from the data correlation upsampling DUp module;

(4) Obtain the CAD-UNet network model.

Described step (4) specifically comprises the following steps:

(4a) Construct a joint loss function: Construct a joint loss function that combines BCE Loss binary cross-entropy loss and Focal Loss focal loss. The formulas of BCE Loss binary cross-entropy loss and Focal Loss focal loss are as follows:

BL(p _t , target)=-ω*(target*ln(p _t )+(1-target)*ln(1-p _t )) (7)

In the formula, p _t is the predicted value of CAD-UNet network model, target is the label value, and ω is the weight value;

FL(p _t )＝-α(1-p _t ) ^γ log(p _t ) (8)

In the formula, p _t is the prediction value of the CAD-UNet network model, α is a balance parameter, used to balance the proportion of positive and negative samples, and the value range is (0, 1]; γ is a focusing parameter, used to reduce the number of easy-to-classify samples Loss, value range [0, +∞);

The joint loss function is shown in Equation (9):

Loss＝BL+FL (9)

(4b) Parameter setting: set ω=1, α=0.5, γ=2;

(4c) Training strategy: use the pre-training weight of the VGG16 network during training, adopt the frozen training method, freeze the parameters of the backbone network for the first 100 epochs for training, and train normally for the next 100 epochs, and train a total of 200 epochs for each experiment;

(4d) Evaluation of model accuracy: The accuracy of evaluation indicators Precision and IoU are used to evaluate the accuracy. The calculation formulas of evaluation indicators are shown in formulas (10) and (11):

In the formula, TP means that the true value is positive, and the model judges it as positive; FP means that the true value is negative, and the model judges it as positive; FN means that the true value is positive, and the model judges it as negative.

Described step (5) specifically comprises the following steps:

(5a) After performing data preprocessing on the new remote sensing image to be extracted, the size is adjusted to 512×512;

(5b) Input the adjusted image into the trained CAD-UNet network model, output the predicted image through the CAD-UNet network model, and obtain the building extraction result. The CAD-UNet network model predicts that the pixel value of the building is 255 , the CAD-UNet network model predicts that the pixel value of the background is 0, so the white part in the predicted image is the building area, and the black part is the background area.

It can be seen from the above technical solution that the beneficial effects of the present invention are as follows: First, the building extraction accuracy is high. Compared with other methods, the network designed by the present invention gradually extracts the deep features of the buildings, and then gradually upsamples to The input resolution size is more friendly to the building extraction task, and the accuracy of building extraction is significantly improved; second, it has better building boundary extraction effect, the coordinate attention CA module added and constructed by the present invention, and the data correlation The sampling DUp module can effectively capture the location information and boundary information of the building, so that the extracted building can have a smoother boundary and a complete outline; the third, the network parameter amount is small, and it is easy to train. The coordinates used in the present invention pay attention to Force is a kind of plug-and-play lightweight attention, and the CAD-UNet network model of the present invention has fewer channels than the original UNet model, which reduces the complexity of the network, because the method of the present invention has fewer parameters , easy to train.

Description of drawings

Fig. 1 is method flowchart of the present invention;

Fig. 2 is the structural diagram of CAD-UNet network model among the present invention;

Fig. 3 is the structural diagram of coordinate attention CA module in the present invention;

Fig. 4 is the structural diagram of data-dependent upsampling DUp module among the present invention;

Fig. 5 is the training data example among the present invention;

Fig. 6 is a graph of prediction results in the present invention.

Detailed ways

As shown in Figure 1, a remote sensing image building extraction method based on coordinate attention and data correlation upsampling, the method includes the following sequential steps:

(1) Obtain remote sensing data: download WHU building data set and Massachusetts building data set; The WHU building data set is Wuhan University building data set, and the Massachusetts building data set is Massachusetts building data set;

(2) Data preprocessing and data enhancement: the preprocessing refers to cropping the large images in the data set, and performing data enhancement on the cropped remote sensing images and label images; The ratio of 8:1:1 is divided into training set, verification set and test set;

(3) Constructing the CAD-UNet network model: Based on the UNet network, improve and construct the building extraction network model including the encoder, the coordinate attention CA module, and the data-dependent upsampling DUp module, namely the CAD-UNet network model;

(4) Model training and evaluation: Based on the joint loss function of the training set data BCE Loss binary classification cross entropy loss and Focal Loss focal loss, the CAD-UNet network model is trained, and the test set is used to evaluate CAD-UNet after training Building extraction accuracy and effect of network model;

(5) Automatic extraction of buildings: after data preprocessing of the new remote sensing images to be extracted, the trained CAD-UNet network model is input, and the CAD-UNet network model outputs predicted images to obtain building extraction results.

Described step (2) specifically comprises the following steps:

(2a) By means of sliding window, crop the remote sensing large image and label image in the Massachusetts building dataset, cut it into a 512×512 image, and put the WHU building dataset and the Massachusetts building dataset in the label image into the building The pixel value of the background is marked as 1, and the pixel value of the background is marked as 0;

(2b) Carry out data enhancement to the remote sensing images and label images in the WHU building dataset and the cropped Massachusetts building dataset remote sensing images and label images to expand the amount of data. The data enhancement includes:

Horizontal flip: Use the image processing library OpenCV to flip the remote sensing image and the label image horizontally;

Vertical flip: Use the image processing library OpenCV to flip the remote sensing image and the label image vertically;

Horizontal and vertical flip: Use the image processing library OpenCV to flip the remote sensing image and the label image first horizontally and then vertically;

Shift, zoom, random crop and add noise: perform operations such as shift, zoom, random crop and add noise on remote sensing images and label images respectively;

(2c) Divide the remote sensing images and label images after data enhancement into training set, verification set and test set according to the ratio of 8:1:1. The training set is used to directly participate in the training of CAD-UNet network model. Feature extraction; the verification set is used to adjust the hyperparameters of the CAD-UNet network model; the test set is used to test the accuracy and extraction effect of the CAD-UNet network model after the training is completed.

Described step (3) specifically comprises the following steps:

(3a) Replace the UNet network encoder: replace the UNet network encoder with the VGG16 network module. The VGG16 network module is composed of the VGG16 network without the last pooling layer and the fully connected layer. The VGG16 network module passes multiple convolutions and four times Maximum pooling is used for down-sampling, extracting building features in remote sensing images, and outputting feature maps of four different scales;

(3b) Construct the coordinate attention CA module: embed the coordinate attention CA module into the skip connection of the UNet network obtained through step (3a);

The coordinate attention CA module captures long-range dependence and retains positional information along two spatial directions, and encodes the feature maps separately to form two feature maps, which are direction-aware and position-sensitive, respectively, and any intermediate tensor X=[x ₁ , x ₂ , x ₃ ..., x _C ]∈R ^C×H×W as input, and output a tensor of the same length Y=[y ₁ , y ₂ , y ₃ ..., y _C ] , specifically for the input X, use a pooling kernel of size (H, 1) and (1, W) to encode each channel along the horizontal and vertical directions, and the expression of the cth channel with a height of H as follows:

In the formula, H represents the height of the image, W represents the width of the image, C represents the total number of channels of the image, c represents the c-th channel, X _c represents the image of the c-th channel, i represents the abscissa of the image, and R represents the middle sheet volume collection;

Similarly, the output of the cth channel of width W is expressed as follows:

Formulas (1), (2) are two transformations of feature aggregation, which are aggregated along two spatial directions respectively, and return two-direction perceptual attention maps; the coordinate attention CA module generates two feature layers before cascading, Then share a 1×1 convolution operation to transform, as shown in formula (3):

f=δ(F ₁ ([Z ^H , Z ^W ])) (3)

In the formula, δ is a nonlinear activation function, and f is an intermediate feature map, which is the result of feature encoding of spatial information in both horizontal and vertical directions; then f is decomposed into two separate Tensor, f ^H ∈ R ^C/r×H and f ^W ∈ R ^C/r×W , r is used to control the reduction ratio of the number of channels, and then use two other 1×1 convolutions to transform F _H and F _W transforms f ^H and f ^W into two tensors g ^H and g ^W with the same number of feature layers, respectively:

g ^H ＝σ(F _H (f ^H )) (4)

g ^W =σ(F _H (f ^W ))(11)

In the formula, F _H and F _W are two 1×1 convolution transformations, f ^H and f ^W are two separate tensors obtained by decomposing f, g ^H and g ^W are convolution transformations and activation functions The obtained tensor, σ is the sigmoid activation function. During the transformation process, use the reduction ratio r to reduce the number of channels of f, and then expand the output g ^H and g ^W , respectively as attention weights; coordinate attention CA The final output of the module is shown in formula (6):

(3c) Construct a data-dependent upsampling DUp module: combine the convolutional layer and the data-dependent upsampling module to construct a data-dependent upsampling DUp module, and use it to extract the boundary information of high-resolution buildings. For the four different scale features of the input Figure, first pass through a 3×3 convolutional layer to reduce the number of channels of the feature map; then perform data-related upsampling to directly restore the feature map to a size of 512×512, and perform point-by-point processing on the four feature maps obtained after upsampling After addition and fusion, output from the data correlation upsampling DUp module;

(4) Obtain the CAD-UNet network model.

Described step (4) specifically comprises the following steps:

(4a) Construct a joint loss function: Construct a joint loss function that combines BCE Loss binary cross-entropy loss and Focal Loss focal loss. The formulas of BCE Loss binary cross-entropy loss and Focal Loss focal loss are as follows:

BL(p _t , target)=-ω*(target*ln(p _t )+(1-target)*ln(1-p _t )) (7)

In the formula, p _t is the predicted value of CAD-UNet network model, target is the label value, and ω is the weight value;

FL(p _t )＝-α(1-p _t ) ^γ log(p _t ) (8)

In the formula, p _t is the prediction value of the CAD-UNet network model, α is a balance parameter, used to balance the proportion of positive and negative samples, and the value range is (0, 1]; γ is a focusing parameter, used to reduce the number of easy-to-classify samples Loss, value range [0, +∞);

The joint loss function is shown in Equation (9):

Loss＝BL+FL (9)

(4b) Parameter setting: set ω=1, α=0.5, γ=2;

(4c) Training strategy: use the pre-training weight of the VGG16 network during training, adopt the frozen training method, freeze the parameters of the backbone network for the first 100 epochs for training, and train normally for the next 100 epochs, and train a total of 200 epochs for each experiment;

(4d) Evaluation of model accuracy: The accuracy of evaluation indicators Precision and IoU are used to evaluate the accuracy. The calculation formulas of evaluation indicators are shown in formulas (10) and (11):

In the formula, TP means that the true value is positive, and the model judges it as positive; FP means that the true value is negative, and the model judges it as positive; FN means that the true value is positive, and the model judges it as negative.

Described step (5) specifically comprises the following steps:

(5a) After performing data preprocessing on the new remote sensing image to be extracted, the size is adjusted to 512×512;

(5b) Input the adjusted image into the trained CAD-UNet network model, output the predicted image through the CAD-UNet network model, and obtain the building extraction result. The CAD-UNet network model predicts that the pixel value of the building is 255 , the CAD-UNet network model predicts that the pixel value of the background is 0, so the white part in the predicted image is the building area, and the black part is the background area.

In order to verify the effectiveness of the present invention, Unet is selected as a comparative example, and the standard building data set is used to compare the results, and compare the accuracy and intersection ratio of the algorithm.

Table 1: Comparison of the results of the examples and comparative examples on the data set

As shown in Figure 2, the CAD-UNet network model in the present invention also adopts a codec structure, the left side is an encoder for down-sampling and feature extraction; the middle solid arrow represents the coordinate attention CA module, which is used to pay attention to the building Position information; on the right is the decoder, which is used for feature fusion and upsampling; the dotted arrow in the lower right corner represents the data-related upsampling DUp module, which is used to extract the boundary information of the building; finally, after adjusting the number of channels by 1×1 convolution, the output Building extraction results.

As shown in Figure 3, for the input feature tensor, the coordinate attention CA module in the present invention first encodes the channel from the horizontal direction and the vertical direction respectively; then, after aggregating the features in two different directions, the middle tensor is obtained Quantity; Finally, the intermediate tensor is split along the spatial dimension, and the final output is obtained after going through the convolutional layer and the Sigmoid function respectively.

As shown in Figure 4, the data-dependent up-sampling DUp module in the present invention passes the four input feature maps of different scales through a 3×3 convolution layer to reduce the number of channels; then performs data-dependent up-sampling, and the four The feature map is directly restored to the size of 512×512; finally, the four feature maps are added and fused point by point, and then output from the DUp module.

Figure 5 is an example of the WHU building dataset and the Massachusetts building dataset respectively. The left side is the remote sensing image, and the right side is the corresponding real label.

Fig. 6 is the predicted result of the CAD-UNet network model of the present invention, the first column is the remote sensing image, the second column is the corresponding real label, the third column is the predicted result of the CAD-UNet network model of the present invention, and the fourth column is UNet It can be seen from the figure that the prediction result of the method of the present invention has a smoother and clearer boundary, which is better than that of UNet.

In summary, the network designed in the present invention gradually extracts the deep features of buildings, and then gradually upsamples to the input resolution after feature fusion, which is more friendly to building extraction tasks and significantly improves the accuracy of building extraction; The invention adds and constructs the coordinate attention CA module and the data-related upsampling DUp module, which can effectively capture the location information and boundary information of the building, so that the extracted building can have a smoother boundary and a complete outline; the present invention adopts Coordinate attention is a plug-and-play lightweight attention, and the CAD-UNet network model of the present invention has fewer channels than the original UNet model, reducing network complexity, because the method parameters of the present invention Small amount, easy to train.

Claims (5)

1. A remote sensing image building extraction method based on coordinate attention and data correlation up-sampling, characterized in that: the method comprises the steps of the following order: (1) Obtain remote sensing data: download WHU building dataset and Massachusetts building dataset; (2) Data preprocessing and data enhancement: the preprocessing refers to cropping the large images in the data set, and performing data enhancement on the cropped remote sensing images and label images; The ratio of 8:1:1 is divided into training set, verification set and test set; (3) Constructing the CAD-UNet network model: Based on the UNet network, improve and construct the building extraction network model including the encoder, the coordinate attention CA module, and the data-dependent upsampling DUp module, namely the CAD-UNet network model; (4) Model training and evaluation: Based on the joint loss function of the training set data BCE Loss binary classification cross entropy loss and Focal Loss focal loss, the CAD-UNet network model is trained, and the test set is used to evaluate CAD-UNet after training Building extraction accuracy and effect of network model; (5) Automatic extraction of buildings: after data preprocessing of the new remote sensing images to be extracted, the trained CAD-UNet network model is input, and the CAD-UNet network model outputs predicted images to obtain building extraction results. 2. the remote sensing image building extraction method based on coordinate attention and data correlation upsampling according to claim 1, is characterized in that: described step (2) specifically comprises the following steps: (2a) By means of sliding window, crop the remote sensing large image and label image in the Massachusetts building dataset, cut it into a 512×512 image, and put the WHU building dataset and the Massachusetts building dataset in the label image into the building The pixel value of the background is marked as 1, and the pixel value of the background is marked as 0; (2b) Carry out data enhancement to the remote sensing images and label images in the WHU building dataset and the cropped Massachusetts building dataset remote sensing images and label images to expand the amount of data. The data enhancement includes: Horizontal flip: Use the image processing library OpenCV to flip the remote sensing image and the label image horizontally; Vertical flip: Use the image processing library OpenCV to flip the remote sensing image and the label image vertically; Horizontal and vertical flip: Use the image processing library OpenCV to flip the remote sensing image and the label image first horizontally and then vertically; Shift, zoom, random crop and add noise: perform operations such as shift, zoom, random crop and add noise on remote sensing images and label images respectively; (2c) divide the remote sensing image and label image after data enhancement into training set, verification set and test set according to the ratio of 8:1:1 respectively, and the training set is used to directly participate in the training of CAD-UNet network model, and carry out Feature extraction; the verification set is used to adjust the hyperparameters of the CAD-UNet network model; the test set is used to test the accuracy and extraction effect of the CAD-UNet network model after the training is completed. 3. the remote sensing image building extraction method based on coordinate attention and data correlation upsampling according to claim 1, is characterized in that: described step (3) specifically comprises the following steps: (3a) Replace the UNet network encoder: replace the UNet network encoder with the VGG16 network module. The VGG16 network module is composed of the VGG16 network without the last pooling layer and the fully connected layer. The VGG16 network module passes multiple convolutions and four times Maximum pooling is used for down-sampling, extracting building features in remote sensing images, and outputting feature maps of four different scales; (3b) Construct the coordinate attention CA module: embed the coordinate attention CA module into the skip connection of the UNet network obtained through step (3a); The coordinate attention CA module captures long-range dependencies and retains location information along two spatial directions, and encodes the feature maps separately to form two feature maps, which are direction-aware and position-sensitive, respectively, and any intermediate tensor X=[x ₁ ,x ₂ ,x ₃ ...,x _C ]∈R ^C×H×W as input, and output a tensor of the same length Y=[y ₁ ,y ₂ ,y ₃ ...,y _C ] , specifically for the input X, use a pooling kernel of size (H, 1) and (1, W) to encode each channel along the horizontal and vertical directions, and the expression of the cth channel with a height of H as follows:

In the formula, H represents the height of the image, W represents the width of the image, C represents the total number of channels of the image, c represents the c-th channel, X _c represents the image of the c-th channel, i represents the abscissa of the image, and R represents the middle sheet volume collection; Similarly, the output of the cth channel of width W is expressed as follows:

Formulas (1), (2) are two transformations of feature aggregation, which are aggregated along two spatial directions respectively, and return two-direction perceptual attention maps; the coordinate attention CA module generates two feature layers before cascading, Then share a 1×1 convolution operation to transform, as shown in formula (3): f=δ(F ₁ ([Z ^H , Z ^W ])) (3) In the formula, δ is a nonlinear activation function, and f is an intermediate feature map, which is the result of feature encoding of spatial information in both horizontal and vertical directions; then f is decomposed into two separate Tensor, f ^H ∈ R ^C/r×H and f ^W ∈ R ^C/r×W , r is used to control the reduction ratio of the number of channels, and then use two other 1×1 convolutions to transform F _H and F _W transforms f ^H and f ^W into two tensors g ^H and g ^W with the same number of feature layers, respectively: g ^H ＝σ(F _H (f ^H )) (4) g ^w =σ(F _H (f ^W ))(3) In the formula, F _H and F _W are two 1×1 convolution transformations, f ^H and f ^W are two separate tensors obtained by decomposing f, g ^H and g ^W are convolution transformations and activation functions The obtained tensor, σ is the sigmoid activation function. During the transformation process, use the reduction ratio r to reduce the number of channels of f, and then expand the output g ^H and g ^W , respectively as attention weights; coordinate attention CA The final output of the module is shown in formula (6):

(3c) Construct a data-dependent upsampling DUp module: combine the convolutional layer and the data-dependent upsampling module to construct a data-dependent upsampling DUp module, and use it to extract the boundary information of high-resolution buildings. For the four different scale features of the input Figure, first pass through a 3×3 convolutional layer to reduce the number of channels of the feature map; then perform data-related upsampling to directly restore the feature map to a size of 512×512, and perform point-by-point processing on the four feature maps obtained after upsampling After addition and fusion, output from the data correlation upsampling DUp module; (4) Obtain the CAD-UNet network model. 4. the remote sensing image building extraction method based on coordinate attention and data correlation upsampling according to claim 1, is characterized in that: described step (4) specifically comprises the following steps: (4a) Construct a joint loss function: Construct a joint loss function that combines BCE Loss binary cross-entropy loss and Focal Loss focal loss. The formulas of BCE Loss binary cross-entropy loss and Focal Loss focal loss are as follows: BL(p _t , target)=-ω*(target*ln(p _t )+(1-target)*ln(1-p _t )) (7) In the formula, p _t is the predicted value of CAD-UNet network model, target is the label value, and ω is the weight value; FL(p _t )＝-α(1-p _t ) ^γ log(p _t ) (8) In the formula, p _t is the prediction value of the CAD-UNet network model, α is a balance parameter, used to balance the proportion of positive and negative samples, and the value range is (0, 1]; γ is a focusing parameter, used to reduce the number of easy-to-classify samples Loss, value range [0, +∞); The joint loss function is shown in Equation (9): Loss＝BL+FL (9) (4b) Parameter setting: set ω=1, α=0.5, γ=2; (4c) Training strategy: use the pre-training weight of the VGG16 network during training, adopt the frozen training method, freeze the parameters of the backbone network for the first 100 epochs for training, and train normally for the next 100 epochs, and train a total of 200 epochs for each experiment; (4d) Evaluation of model accuracy: The accuracy of evaluation indicators Precision and IoU are used to evaluate the accuracy. The calculation formulas of evaluation indicators are shown in formulas (10) and (11):

In the formula, TP means that the true value is positive, and the model judges it as positive; FP means that the true value is negative, and the model judges it as positive; FN means that the true value is positive, and the model judges it as negative. 5. the remote sensing image building extraction method based on coordinate attention and data correlation upsampling according to claim 1, is characterized in that: described step (5) specifically comprises the following steps: (5a) After performing data preprocessing on the new remote sensing image to be extracted, the size is adjusted to 512×512; (5b) Input the adjusted image into the trained CAD-UNet network model, output the predicted image through the CAD-UNet network model, and obtain the building extraction result. The CAD-UNet network model predicts that the pixel value of the building is 255 , the CAD-UNet network model predicts that the pixel value of the background is 0, so the white part in the predicted image is the building area, and the black part is the background area.

Images