CN116012709B - High-resolution remote sensing image building extraction method and system - Google Patents

Abstract

The invention provides a high-resolution remote sensing image building extraction method and a system, wherein the method comprises the following steps: acquiring aerial image data in a research area; performing image preprocessing and data enhancement on the aviation image data, and dividing the processed data into a training set, a test set and a verification set; respectively constructing a spatial information encoder, a semantic encoder, a dual-attention extractor and an attention characteristic definer; constructing a depth space attention network according to the space information encoder, the semantic encoder, the dual attention extractor and the attention feature definer, and training the depth space attention network according to the training set; respectively testing and verifying the depth space attention network by using the test set and the verification set to obtain an optimal remote sensing image building extraction model; and inputting the data to be detected into a remote sensing image building extraction model to obtain a detection result. The invention can improve the extraction accuracy of the remote sensing image building.

Description

A method and system for extracting buildings from high-resolution remote sensing images

technical field

The invention relates to the technical field of image processing, in particular to a method and system for extracting buildings from high-resolution remote sensing images.

Background technique

With the rapid development of economic activities, the number of buildings has increased significantly, and the area where buildings are built has become more widely distributed. At the same time, the external contours of buildings are also showing a trend of becoming larger and more complex. Therefore, in order to effectively and reasonably grasp the regional economic development and population distribution, better arrange public service points such as infrastructure and service stations, and efficiently and accurately extract buildings from high-resolution remote sensing images is an urgent need. solved problem.

In the field of computer image processing, deep learning can effectively extract objects after good training, and exhibit excellent computing performance. However, building segmentation is a binary classification task that includes building and non-building elements, and there are still several problems in extracting buildings from high-resolution remote sensing images only through the basic convolutional neural network in the basic model of deep learning . For example, in the face of complex ground object scenes, the accuracy of buildings within a certain range is not high, and in the face of vegetation occlusion, there are incomplete or blurred boundaries. In the face of the "convolutional layer" in the current model, a large number of accumulations are used to extract the buildings. Features are too scattered and so on. All of these have caused great difficulties in efficiently and accurately extracting buildings from high-resolution remote sensing images.

Therefore, the prior art needs a method for improving the accuracy of extracting buildings from remote sensing images, so as to solve the above-mentioned problems.

Contents of the invention

In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a remote sensing image with noise and appropriate size as input data in complex scenes of high-resolution aerial images, based on the attention mechanism in deep learning. A method and system for effectively extracting buildings from high-resolution remote sensing images based on a deep spatial attention network.

To achieve the above object, the present invention provides the following scheme:

A method for extracting buildings from high-resolution remote sensing images, comprising:

Obtain aerial image data in the study area;

Carrying out image preprocessing and data enhancement to the aerial image data, and dividing the processed data into a training set, a test set and a verification set;

Construct spatial information encoder, semantic encoder, dual attention extractor and attention feature definer respectively;

Construct a deep spatial attention network according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and perform the deep spatial attention network according to the training set conduct training;

Using the test set and the verification set to test and verify the deep spatial attention network respectively, to obtain the optimal remote sensing image building extraction model;

The data to be detected is input into the remote sensing image building extraction model to obtain a detection result.

Preferably, said acquisition of aerial image data in the research area includes:

Utilizing the unmanned aerial vehicle to fly in the research area according to the set flight route, so as to collect the remote sensing data sets in the research area;

Generating a remote sensing image of the research area according to the remote sensing data set with geographic coordinates taken by the drone;

The spatial resolution of the remote sensing image is adjusted, and the aerial image data is produced according to the remote sensing image whose spatial resolution reaches a preset resolution threshold.

Preferably, image preprocessing and data enhancement are performed on the aerial image data, and the processed data are divided into training set, test set and verification set, including:

identifying building samples in said aerial imagery data;

determining a building label based on said building sample;

Performing impact clipping on the aerial image data and the corresponding building tags to obtain a clipped image;

performing data enhancement on the cropped image to obtain an expanded image;

The expanded image is randomly divided according to a preset ratio to obtain the training set, the test set and the verification set.

Preferably, the spatial information encoder includes three stages, each stage contains two layers of shallow structure; each layer of the shallow structure is composed of convolution operation, batch normalization and parameter correction linear unit; each The first and second layers of the stage have the same number of filters and output feature map size; the semantic encoder includes channel segmentation and re-partitioning modules and 3 downsampling modules; the dual-attention extractor is based on kernel The channel attention mechanism of attention mechanism and dot product is constructed; The attention feature definer includes a Projection layer and an upsampling module; Each layer of the attention feature definer consists of a convolutional layer, a BN layer, and a PReLU composition.

Preferably, a deep spatial attention network is constructed according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and the depth spatial attention is analyzed according to the training set. The attention network is trained, including:

Constructing a spatial information encoding network branch network according to the spatial information encoder;

Constructing a high-dimensional feature extraction branch network according to the semantic coder, the double-attention extractor and the attention feature definer;

Constructing the deep spatial attention network according to the spatial information encoding network branch network, the high-dimensional feature extraction branch network and the preset feature fusion device;

The training set is input into the deep spatial attention network, and hyperparameters are adjusted for training to obtain a trained deep spatial attention network.

Preferably, the depth spatial attention network is tested and verified by using the test set and the verification set respectively to obtain an optimal remote sensing image building extraction model, including:

Initialize the training parameters of the deep spatial attention network;

Test the deep spatial attention network according to the test set, and evaluate the tested deep spatial attention network according to preset evaluation indicators, so as to obtain the tested deep spatial attention network;

The tested deep spatial attention network is verified according to the verification set, and the optimal remote sensing image building extraction model is obtained.

A system for extracting buildings from high-resolution remote sensing images, including:

Image acquisition module, used to acquire aerial image data in the research area;

A data processing module, for performing image preprocessing and data enhancement on the aerial image data, and dividing the processed data into a training set, a test set and a verification set;

Module building blocks for building spatial information encoder, semantic encoder, dual attention extractor and attentional feature definer respectively;

A network training module for constructing a deep spatial attention network according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and according to the training set for all The deep spatial attention network described above is trained;

The model optimization module is used to test and verify the deep spatial attention network by using the test set and the verification set respectively, so as to obtain the optimal remote sensing image building extraction model;

The detection module is used to input the data to be detected into the remote sensing image building extraction model to obtain detection results.

According to the specific embodiments provided by the invention, the invention discloses the following technical effects:

The invention provides a method and system for extracting buildings from high-resolution remote sensing images. The method includes: acquiring aerial image data in the research area; performing image preprocessing and data enhancement on the aerial image data, and converting the processed data to Divided into training set, test set and verification set; Build spatial information encoder, semantic encoder, double attention extractor and attention feature definer respectively; According to described spatial information encoder, described semantic encoder, described Two attention extractors and the attention feature definer construct depth space attention network, and according to described training set, described depth space attention network is trained; The deep spatial attention network is tested and verified to obtain the optimal remote sensing image building extraction model; the data to be detected is input into the remote sensing image building extraction model to obtain the detection result. The invention can improve the accuracy rate of remote sensing image building extraction.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the embodiments. Obviously, the accompanying drawings in the following description are only some of the present invention. Embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without paying creative labor.

Fig. 1 is the flow chart of the method provided by the embodiment of the present invention;

FIG. 2 is a schematic diagram of a building extraction process provided by an embodiment of the present invention;

Fig. 3 is a data set display diagram provided by the embodiment of the present invention;

4 is a framework diagram of a deep attention model provided by an embodiment of the present invention;

Fig. 5 is a graph of experimental results provided by the embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Reference herein to an "embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The occurrences of this phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is understood explicitly and implicitly by those skilled in the art that the embodiments described herein can be combined with other embodiments.

The terms "first", "second", "third" and "fourth" in the specification and claims of the present application and the drawings are used to distinguish different objects, rather than to describe a specific order . Furthermore, the terms "include" and "have", as well as any variations thereof, are intended to cover a non-exclusive inclusion. For example, including a series of steps, processes, methods, etc. are not limited to the listed steps, but optionally also include steps that are not listed, or optionally also include inherent aspects of these processes, methods, products or equipment. other steps.

The purpose of the present invention is to provide a high-resolution remote sensing image building extraction method and system, which can improve the accuracy of remote sensing image building extraction.

In order to make the above objects, features and advantages of the present invention more comprehensible, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Fig. 1 is a flow chart of the method provided by the embodiment of the present invention. As shown in Fig. 1, the present invention provides a method for extracting buildings from high-resolution remote sensing images, including:

Step 100: Obtain aerial image data in the research area;

Step 200: Perform image preprocessing and data enhancement on the aerial image data, and divide the processed data into training set, test set and verification set;

Step 300: Constructing a spatial information encoder, a semantic encoder, a double-attention extractor and an attention feature definer respectively;

Step 400: Construct a deep spatial attention network according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and classify the depth spatial attention according to the training set attention network for training;

Step 500: Using the test set and the verification set to test and verify the deep spatial attention network respectively, to obtain the optimal remote sensing image building extraction model;

Step 600: Input the data to be detected into the remote sensing image building extraction model to obtain a detection result.

Preferably, said acquisition of aerial image data in the research area includes:

Utilizing the unmanned aerial vehicle to fly in the research area according to the set flight route, so as to collect the remote sensing data sets in the research area;

Generating a remote sensing image of the research area according to the remote sensing data set with geographic coordinates taken by the drone;

The spatial resolution of the remote sensing image is adjusted, and the aerial image data is produced according to the remote sensing image whose spatial resolution reaches a preset resolution threshold.

As shown in Figure 2, the building extraction process in this embodiment includes the following steps:

Step 1, making Weinan aerial image data through the high-resolution aerial image collected by UAV, including the following sub-steps:

Step 1.1. On the basis of field investigation, select Helan Village, Fuping County, Weinan City, Shaanxi Province as the research area, and use drones to fly over the research area to collect high-resolution remote sensing images in the research area to create a dataset.

In step 1.2, the experimental data used in this study comes from a small quadrotor unmanned aerial vehicle (UAV), which can be used to autonomously plan flight paths within the study area. Among them, the flight route is from the southeast corner to the northwest corner of the research area, and the photos are taken along the "S route".

Step 1.3, after importing the remote sensing images with geographical coordinates taken by the UAV into the relevant software, the remote sensing images of the entire research area can be generated, and then the spatial resolution of the images is adjusted, using a camera with a spatial resolution of 2.3 cm UAV data is made into Weinan aerial image data.

Preferably, image preprocessing and data enhancement are performed on the aerial image data, and the processed data are divided into training set, test set and verification set, including:

identifying building samples in said aerial imagery data;

determining a building label based on said building sample;

Performing impact clipping on the aerial image data and the corresponding building tags to obtain a clipped image;

performing data enhancement on the cropped image to obtain an expanded image;

The expanded image is randomly divided according to a preset ratio to obtain the training set, the test set and the verification set.

Further, step 2 in this embodiment is: based on the Weinan aerial image data produced in step 1, first perform image preprocessing and data enhancement and then divide it into training set, test set and verification set in proportion, including the following step:

Step 2.1, in the ArcGIS 10.2 software, manually select architectural samples, which contain various types of buildings in the study area, and convert the marked vector files into ".GIF" format through the "Vector to Raster" tool The image label label, and then complete the building label of image annotation, as shown in Figure 3.

In step 2.2, in order to facilitate deep learning calculations, image cropping is performed on the original image and the corresponding label, and the entire remote sensing image of the research area and the corresponding label are uniformly cropped in the ArcGIS 10.2 software, and the cropped size is 512× Small image size of 512 pixels.

Step 2.3: Carry out data enhancement on small-sized images to expand the amount of data, rotate each image and the corresponding label label by 90°, 180°, and 270°, and then perform random horizontal and vertical flip operations with a probability of 0.5. A total of 830 images of 512 × 512 pixels were generated.

In step 2.4, 830 images of 512 × 512 pixels are randomly divided into training set, test set and verification set according to the ratio of 6:2:2.

Preferably, the spatial information encoder includes three stages, each stage contains two layers of shallow structure; each layer of the shallow structure is composed of convolution operation, batch normalization and parameter correction linear unit; each The first and second layers of the stage have the same number of filters and output feature map size; the semantic encoder includes channel segmentation and re-partitioning modules and 3 downsampling modules; the dual-attention extractor is based on kernel The channel attention mechanism of attention mechanism and dot product is constructed; The attention feature definer includes a Projection layer and an upsampling module; Each layer of the attention feature definer consists of a convolutional layer, a BN layer, and a PReLU composition.

Concretely, step 3 in the present embodiment is the four major components required for constructing the deep spatial attention network, including the following sub-steps:

In step 3.1, the spatial information encoder SIEN is constructed, which effectively combines high-level semantics and low-level details to improve the performance of semantic segmentation. A shallow SIEN with three stages and six layers is designed, where each layer consists of convolution operation (Conv), batch normalization (BN) and parameter rectified linear unit (PReLU). The first and second layers of each stage have the same number of filters (with a stride of 2) and output feature map size. Therefore, SIEN extracts 1/8 of the original input, improves the channel capacity and enriches the spatial details.

Step 3.2, constructing the semantic encoder SEN, by applying 5 downsampling to set the ratio of input image resolution, the size of the final output is only 1/32 of the input image size, which may lead to significant loss of spatial details. SEN is based on the channel segmentation and resegmentation module FDSS_nbt module and applies 3 downsampling (the output resolution is only 1/8 of the original image resolution with 32, 64 and 128 channels). In stages 3 and 4, special convolutions are introduced to increase the receptive field.

In step 3.3, the dual-attention extractor DAB is constructed, and a kernel-based attention mechanism (KAM) is designed in terms of the spatial dimension of the image. In the channel dimension, using the dot product-based channel attention mechanism (CAM), the number of input channels C is usually much smaller than the number of pixels contained in the feature map, and the complexity of the Softmax function for multiple channels is not high. The Dual Attention Block (DAM) combines the KAM that simulates the position dependence and the CAM that simulates the channel dependence, which can enhance the ability of the extracted features of each layer feature map to recognize the target.

Step 3.4, build the attention feature definer AFRB, which specifically includes the Projection layer and upsampling, each layer consists of a convolutional layer, a BN layer, and a PReLU. Set the number of input channels equal to the number of output channels of the double-attention extractor DAM in step 3.3, and set the number of output channels to the following formula:

AFRB = conv3(conv3(F _att ));

Among them, F _att represents the number of output channels of DAM, conv3 represents the fusion convolution of the third layer, and AFRB represents the number of buildings.

Preferably, a deep spatial attention network is constructed according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and the depth spatial attention is analyzed according to the training set. The attention network is trained, including:

Constructing a spatial information encoding network branch network according to the spatial information encoder;

Constructing a high-dimensional feature extraction branch network according to the semantic coder, the double-attention extractor and the attention feature definer;

Constructing the deep spatial attention network according to the spatial information encoding network branch network, the high-dimensional feature extraction branch network and the preset feature fusion device;

The training set is input into the deep spatial attention network, and hyperparameters are adjusted for training to obtain a trained deep spatial attention network.

Further, step 4 in this embodiment is: combine the four major components constructed in step 3 into a remote sensing image building extraction algorithm based on a deep spatial attention network, and then train the training set for building extraction, including the following Substeps:

Step 4.1, an overview of the deep spatial attention network architecture shown in Figure 4, where the two branch networks are the spatial information encoding network branch and the high-dimensional feature extraction branch, and the spatial information encoding network branch is composed of the spatial information encoder SIEN (Fig. shown in a in Figure 4), the high-dimensional feature extraction branch is composed of the semantic encoder SEN (shown in b in Figure 4) as the main building block for constructing the advanced feature extraction branch, and SEN is composed of the feature fuser FFM (shown in c in Figure 4 ), a dual attention extractor DAB (shown in d in Figure 4) and an attention feature definer AFRB (shown in e in Figure 4).

Step 4.2, input the input image to the SEN module to generate 4 feature maps with reduced spatial resolution (F _h,1 , F _h,2 , F _h,3 , F _h,4 ). The feature maps F _h,3 and F _h,4 have the same number of channels with different expansion rates to expand the receptive field convolution filter, and adopt a 1/8 downsampling strategy to convert the high-resolution feature map F _h,3 Combined with the spatial features of F _h,4 , the final model is obtained.

In step 4.3, the images and corresponding labels in the images and labels folders in the training set are input into the constructed deep spatial attention network model, and training is performed after adjusting the hyperparameters to obtain an effective training parameter model.

Preferably, the depth spatial attention network is tested and verified by using the test set and the verification set respectively to obtain an optimal remote sensing image building extraction model, including:

Initialize the training parameters of the deep spatial attention network;

Test the deep spatial attention network according to the test set, and evaluate the tested deep spatial attention network according to preset evaluation indicators, so as to obtain the tested deep spatial attention network;

The tested deep spatial attention network is verified according to the verification set, and the optimal remote sensing image building extraction model is obtained.

Specifically, step 5 in this embodiment is: use the deep learning semantic segmentation model to perform random testing on the test set data until the random testing achieves satisfactory results, including the following sub-steps:

Step 5.1, the experiment is carried out using the PyTorch deep learning framework, and the experimental environment is built by the PyCharm software, which is constructed in the Python language.

Step 5.2, according to the common evaluation criteria of deep learning semantic segmentation, five most common evaluation indicators are used for quality evaluation, including "overall accuracy (OA)", "precision (Precision)", "recall rate (Recall)" , "F1 Score (F1)" and "Intersection Union (IoU)", the specific evaluation index formula is as follows:

Among them, P represents the number of positive samples, N represents the number of negative samples, TP represents the number of true positives, TN represents the number of true negatives, FP represents the number of false positives, and FN represents the number of false negatives.

Step 5.3, all experiments are equipped with CPU E5-2630v4 (2.20GHz) and two NVIDIA GeForce GTX 1080Ti (11GB) servers.

Step 5.4, when training the deep spatial attention network model, set the parameters of the input data to epoch to 100, batches to 16, the optimizer to select Adam, the initial learning rate to 0.0001, the weight decay to 0.0001, and the loss function to be the cross-entropy loss function to find the minimum The network parameterization loss function is based on the stochastic gradient descent algorithm.

Further, in the present embodiment, step 6 is to input data parameters according to the deep spatial attention network model training set in step 5, and the images and corresponding labels in the training set are input into the model constructed by Python software for training, The resulting parametric model is named "epoch100_batches16_learning0.0001.pth".

In step 6.2, use this parameter model to directly predict the building extraction for the high-resolution remote sensing images that do not input the label label in the verification set, and then evaluate the accuracy of the model according to the five evaluation indicators.

Step 6.3, if the accuracy of this parameter model does not meet the expected effect (IoU reaches 85% and above), you need to modify the hyperparameter content of the model, re-input the data of the training set, use the data of the test set for testing and save the highest accuracy A good parameter model, then evaluate the accuracy of the images in the verification set, and finally save the parameter model with the best evaluation index is the required deep spatial attention network model, as shown in Figure 5 is the extracted high-resolution High-precision results of remote sensing image buildings.

Corresponding to the above method, this embodiment also provides a high-resolution remote sensing image building extraction system, including:

Image acquisition module, used to acquire aerial image data in the research area;

A data processing module, for performing image preprocessing and data enhancement on the aerial image data, and dividing the processed data into a training set, a test set and a verification set;

Module building blocks for building spatial information encoder, semantic encoder, dual attention extractor and attentional feature definer respectively;

A network training module for constructing a deep spatial attention network according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and according to the training set for all The deep spatial attention network described above is trained;

The model optimization module is used to test and verify the deep spatial attention network by using the test set and the verification set respectively, so as to obtain the optimal remote sensing image building extraction model;

The detection module is used to input the data to be detected into the remote sensing image building extraction model to obtain detection results.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same and similar parts of each embodiment can be referred to each other. As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and for the related information, please refer to the description of the method part.

In this paper, specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; meanwhile, for those of ordinary skill in the art, according to the present invention Thoughts, there will be changes in specific implementation methods and application ranges. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims (5)

1. A method for extracting buildings from high-resolution remote sensing images, comprising: Obtain aerial image data in the study area; Carrying out image preprocessing and data enhancement to the aerial image data, and dividing the processed data into a training set, a test set and a verification set; Construct spatial information encoder, semantic encoder, dual attention extractor and attention feature definer respectively; Construct a deep spatial attention network according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and perform the deep spatial attention network according to the training set conduct training; Using the test set and the verification set to test and verify the deep spatial attention network respectively, to obtain the optimal remote sensing image building extraction model; Inputting the data to be detected into the remote sensing image building extraction model to obtain a detection result; The spatial information encoder includes three stages, each stage contains two layers of shallow structure; each layer of the shallow structure is composed of convolution operation, batch normalization and parameter correction linear unit; the first layer of each stage The first layer and the second layer have the same number of filters and output feature map size; the semantic encoder includes channel segmentation and repartitioning modules and 3 downsampling modules; the dual attention extractor is based on the kernel attention mechanism and the channel attention mechanism of dot product are constructed; the attention feature definer includes a Projection layer and an upsampling module; each layer of the attention feature definer is composed of a convolutional layer, a BN layer and a PReLU; Construct a deep spatial attention network according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and perform the deep spatial attention network according to the training set Conduct training, including: Constructing a spatial information encoding network branch network according to the spatial information encoder; Constructing a high-dimensional feature extraction branch network according to the semantic coder, the double-attention extractor and the attention feature definer; Constructing the deep spatial attention network according to the spatial information encoding network branch network, the high-dimensional feature extraction branch network and the preset feature fusion device; The training set is input into the deep spatial attention network, and hyperparameters are adjusted for training to obtain a trained deep spatial attention network. 2. The high-resolution remote sensing image building extraction method according to claim 1, wherein said acquisition of aerial image data in the study area includes: Utilizing the unmanned aerial vehicle to fly in the research area according to the set flight route, so as to collect the remote sensing data sets in the research area; Generating a remote sensing image of the research area according to the remote sensing data set with geographic coordinates taken by the drone; The spatial resolution of the remote sensing image is adjusted, and the aerial image data is produced according to the remote sensing image whose spatial resolution reaches a preset resolution threshold. 3. the high-resolution remote sensing image building extraction method according to claim 1, is characterized in that, image preprocessing and data enhancement are carried out to described aerial image data, and the data after processing is divided into training set, test set and a validation set, including: identifying building samples in said aerial imagery data; determining a building label based on said building sample; Performing impact clipping on the aerial image data and the corresponding building tags to obtain a clipped image; performing data enhancement on the cropped image to obtain an expanded image; The expanded image is randomly divided according to a preset ratio to obtain the training set, the test set and the verification set. 4. the high-resolution remote sensing image building extraction method according to claim 1, is characterized in that, utilizes described test set and described verification set to carry out test and verification to described deep space attention network respectively, obtains optimal The remote sensing image building extraction model, including: Initialize the training parameters of the deep spatial attention network; Test the deep spatial attention network according to the test set, and evaluate the tested deep spatial attention network according to preset evaluation indicators, so as to obtain the tested deep spatial attention network; The tested deep spatial attention network is verified according to the verification set, and the optimal remote sensing image building extraction model is obtained. 5. A high-resolution remote sensing image building extraction system, characterized in that it comprises: Image acquisition module, used to acquire aerial image data in the research area; A data processing module, for performing image preprocessing and data enhancement on the aerial image data, and dividing the processed data into a training set, a test set and a verification set; Module building blocks for building spatial information encoder, semantic encoder, dual attention extractor and attentional feature definer respectively; A network training module for constructing a deep spatial attention network according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and according to the training set for all The deep spatial attention network described above is trained; The model optimization module is used to test and verify the deep spatial attention network by using the test set and the verification set respectively, so as to obtain the optimal remote sensing image building extraction model; A detection module, configured to input the data to be detected into the remote sensing image building extraction model to obtain a detection result; The spatial information encoder includes three stages, each stage contains two layers of shallow structure; each layer of the shallow structure is composed of convolution operation, batch normalization and parameter correction linear unit; the first layer of each stage The first layer and the second layer have the same number of filters and output feature map size; the semantic encoder includes channel segmentation and repartitioning modules and 3 downsampling modules; the dual attention extractor is based on the kernel attention mechanism and the channel attention mechanism of dot product are constructed; the attention feature definer includes a Projection layer and an upsampling module; each layer of the attention feature definer is composed of a convolutional layer, a BN layer and a PReLU; Construct a deep spatial attention network according to the spatial information encoder, the semantic encoder, the double-attention extractor and the attention feature definer, and perform the deep spatial attention network according to the training set Conduct training, including: Constructing a spatial information encoding network branch network according to the spatial information encoder; Constructing a high-dimensional feature extraction branch network according to the semantic coder, the double-attention extractor and the attention feature definer; Constructing the deep spatial attention network according to the spatial information encoding network branch network, the high-dimensional feature extraction branch network and the preset feature fusion device; The training set is input into the deep spatial attention network, and hyperparameters are adjusted for training to obtain a trained deep spatial attention network.