Methods, devices and electronic equipment for detecting building changes in high-resolution remote sensing images

Abstract

This invention discloses a method, apparatus, and electronic device for detecting building changes in high-resolution remote sensing imagery. The method includes acquiring a dual-temporal image dataset of buildings in the same region, preprocessing it, dividing it into training, validation, and test sets, and performing data augmentation. A convolutional neural network FERA-Net with an encoder-decoder structure is constructed. The encoder includes a high-frequency feature enhancement module guided by residual attention, and the decoder includes a skip connection module for feature enhancement. Data from the training and validation sets are input into the network. Multi-level features are extracted in the encoder, and full-scale information is aggregated in the decoder to better recover the feature map. A hybrid loss function is designed to optimize network training. The training results are validated using the validation set. After training, data from the test set is input into the trained network model to predict building change detection results. This invention effectively solves the problems of multi-scale building change detection and building edge blurring, significantly improving building change detection performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image information processing in computer vision, and more particularly to a high-resolution remote sensing image building change detection method, device and electronic equipment. BACKGROUND

[0002] Building is an important part of urban development, and building change detection is usually used for land use management, illegal building management and urban development. Change detection aims to analyze multi-temporal images in the same area and detect the differences of objects or phenomena in the area. Remote sensing images can usually be used to identify and evaluate land surface changes, and the use of multi-temporal remote sensing images and their accompanying data obtained for the same land surface area is one of the hottest research fields in the field of remote sensing.

[0003] With the continuous development of earth observation technology, a large amount of remote sensing data with high spectral, spatial and temporal resolution can now be obtained, thereby providing a large amount of information about land cover. The rapid development of earth observation, computing and artificial intelligence technology has also led to the development of various remote sensing image change detection methods. However, the existing building change detection model cannot effectively extract multi-scale building features, nor can it fully utilize the local and global information of the feature map, such as building edges. These defects affect the detection accuracy and may limit the further application of the model.

[0004] Due to the breakthroughs in the field of computer vision, deep learning networks, especially convolutional networks, have been applied to remote sensing image analysis. Deep learning networks are applied to change detection tasks to automatically extract relevant semantic features and obtain abstract feature representations of image information. Compared with traditional methods based on methods, deep learning-based methods can automatically learn features, which significantly reduces the need for professional knowledge. Due to the excellent performance of deep learning models such as convolutional neural networks in capturing and representing image features, the performance of the change detection task has greatly improved the performance and pushed the change detection technology into a new era. Although deep learning methods have performed outstandingly in the change detection task, there are still some problems. On the one hand, the existing detection methods cannot effectively extract multi-scale building features from high-resolution remote sensing images. Due to the continuous downsampling operation of the encoder part and the continuous upsampling operation of the decoder part, details such as building boundary information and multi-scale building information are gradually diluted. On the other hand, when passing context information through skip connections, existing change detection networks often only connect feature maps of the same level. This causes information differences when connecting upper and lower layers, and cannot fully utilize the local and global information of the feature map, thereby affecting the efficiency of feature extraction and limiting the performance of the model.

[0005] Therefore, the multi-scale building change detection problem and the building change detection edge blur problem are technical problems that need to be solved. SUMMARY

[0006] The present application aims to solve the problem of multi-scale building change detection and the problem of building change detection edge blur, in order to solve the technical problem, the present application provides a high-resolution remote sensing image building change detection method, device and electronic equipment, based on residual attention mechanism and high-frequency feature, building change detection is carried out on double-time image data.

[0007] According to the first aspect of the present application, the present application provides a high-resolution remote sensing image building change detection method, which specifically comprises the following steps:

[0008] S1: obtaining double-time building data set of the same area, after preprocessing, dividing into training set, validation set and test set, and carrying out data enhancement processing;

[0009] S2: constructing a convolutional neural network FERA-Net with encoder-decoder structure, the encoder of the FERA-Net includes a residual attention guided high-frequency feature enhancement module, and the decoder of the FERA-Net includes a feature enhanced skip connection module;

[0010] S3: inputting the data in the training set and the validation set into the FERA-Net, extracting multi-level features in the encoder, aggregating full-scale information in the decoder to better restore the feature map, designing a hybrid loss function, optimizing network training, and verifying the training result through the validation set, after training, obtaining a trained network model;

[0011] S4: inputting the data in the test set into the trained network model to predict the building change detection result, testing the model performance, and obtaining the model precision evaluation result.

[0012] Further, step S1 specifically comprises:

[0013] The original data set is cropped to HxW pixel size;

[0014] The cropped data set is divided into training set, validation set and test set;

[0015] The data in the training set and the validation set is enhanced by translation, rotation and vertical mirroring, and the label data is re-encoded and converted into a binary graph with 0 and 1 encoding.

[0016] Further, the network structure of the FERA-Net adopts a twin structure with shared weights to extract double-time image features.

[0017] Further, in step S2, the high-frequency feature enhancement module guided by the multi-layer residual attention in the encoder part enhances the high-frequency features of the remote sensing image and excavates multi-level features to obtain low-level spatial information and deep-level semantic information.

[0018] The high-frequency feature enhancement module guided by the residual attention includes a residual spatial attention feature extraction stage and a channel attention high-frequency feature enhancement stage.

[0019] In the residual spatial attention feature extraction stage, the spatial relationship in the feature map is captured through the spatial attention mechanism, and the building part in the remote sensing image is focused on.

[0020] In the channel attention high-frequency feature enhancement stage, the high-frequency information in the remote sensing image is extracted through the omnidirectional Sobel operator, and the high-frequency features in the feature map are enhanced through the channel attention mechanism.

[0021] In the decoder part, the multi-layer feature maps obtained by the encoder part are enhanced through the feature enhancement skip connection module to enhance the local information and global information in the feature map, and the full-scale information is aggregated through the full-scale skip connection for the restoration of the feature map.

[0022] Further, the full-scale information aggregated through the full-scale skip connection is the first i layer feature map The specific calculation formula is as follows:

[0023]

[0024] wherein, represents the number of down-sampling layers, N represents the total number of feature maps, represents the extracted first layer double-time image feature map, k =1,..., i -1, represents a feature enhancement module, represents a down-sampling operator, represents an activation function and a normalization function, represents an up-sampling operator.

[0025] Further, in step S3, the hybrid loss function includes a main loss function and an edge loss function; the main loss function is used to supervise the main part of the building, and the Focal Loss function is used; the edge loss function is used to supervise the edge part of the building, and the MSE Loss function is used.

[0026] Further, the expression of the hybrid loss function is as follows:

[0027]

[0028] wherein, represents a main loss function, represents an edge loss function, , is a proportional coefficient.

[0029] According to a second aspect of the present application, the present application provides a high-resolution remote sensing image building change detection device for implementing the method, comprising the following modules:

[0030] An image acquisition and preprocessing module is configured to acquire a dual-time building data set of the same area, divide the data set into a training set, a validation set and a test set after preprocessing, and perform data enhancement processing.

[0031] A network construction module is configured to construct a convolutional neural network FERA-Net with an encoder-decoder structure, wherein the encoder of the FERA-Net comprises a residual attention guided high-frequency feature enhancement module, and the decoder of the FERA-Net comprises a feature enhanced skip connection module.

[0032] A network training module is configured to input data in the training set and the validation set into the FERA-Net, extract multi-level features in the encoder, aggregate full-scale information in the decoder to better restore feature maps, design a hybrid loss function, optimize network training, verify the training result through the validation set, and obtain a trained network model after the training is completed.

[0033] A network testing module is configured to input data in the test set into the trained network model to predict building change detection results, test model performance, and obtain model precision evaluation results.

[0034] According to a third aspect of the present application, the present application provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the high-resolution remote sensing image building change detection method when executing the program.

[0035] In addition, according to another aspect of the present application, the present application also provides a storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the high-resolution remote sensing image building change detection method.

[0036] The technical solution provided by the present application has the following beneficial effects:

[0037] This invention provides a method, apparatus, and electronic device for detecting building changes in high-resolution remote sensing imagery. It utilizes a residual attention mechanism and high-frequency features to detect building changes in dual-temporal imagery. Specifically, it includes preprocessing the dual-temporal high-resolution images, performing data augmentation through methods such as random gamma transformation, noise point addition, and rotation to improve the diversity of data samples. An end-to-end encoder-decoder architecture network, FERA-Net, based on the residual attention mechanism and high-frequency features, is constructed. In the encoder, a stacked multi-layer residual attention-guided high-frequency feature enhancement module enhances building edge features, and an extraction module mines multi-level features from the remote sensing imagery. In the decoder, to fully utilize feature map information at different levels, this invention designs a feature-enhanced skip connection module to enhance local and global information in the feature maps and aggregates multi-scale and multi-level feature map information for feature map recovery in the decoder. To better focus on building change information and prior building boundary information, this invention designs a hybrid loss function. Preprocessed dual-temporal high-resolution remote sensing images from the training and validation sets are used as input to the network, and training parameters are set to obtain the corresponding network model. The data from the test set is input into the trained model to detect building changes in high-resolution remote sensing images, yielding the building change detection results. The FERA-Net network proposed in this invention can effectively solve the problems of multi-scale building change detection and building edge blurring, thus significantly improving the performance of building change detection. Attached Figure Description

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0039] Figure 1 This is a technical flowchart of a high-resolution remote sensing image building change detection method based on residual attention mechanism and high-frequency features according to the present invention.

[0040] Figure 2 This is a network structure diagram of a high-resolution remote sensing image building change detection method based on residual attention mechanism and high-frequency features according to the present invention.

[0041] Figure 3 This is a diagram of the residual attention-guided high-frequency feature enhancement module in a high-resolution remote sensing image building change detection method based on residual attention mechanism and high-frequency features, according to the present invention.

[0042] Figure 4 This is a diagram illustrating the residual spatial attention feature extraction stage of a high-resolution remote sensing image building change detection method based on residual attention mechanism and high-frequency features, according to the present invention.

[0043] Figure 5This is a stage diagram of channel attention high-frequency feature enhancement in a high-resolution remote sensing image building change detection method based on residual attention mechanism and high-frequency features according to the present invention.

[0044] Figure 6 This is a feature enhancement skip connection module diagram of a high-resolution remote sensing image building change detection method based on residual attention mechanism and high-frequency features according to the present invention.

[0045] Figure 7 This is an example of the results of building boundary extraction in the high-resolution remote sensing image building change detection method based on residual attention mechanism and high-frequency features of the present invention.

[0046] Figure 8 This is an example of the results of a high-resolution remote sensing image building change detection method based on residual attention mechanism and high-frequency features in different ground environments according to the present invention.

[0047] Figure 9 This is a schematic diagram of the structure of a high-resolution remote sensing image building change detection device based on residual attention mechanism and high-frequency features according to the present invention.

[0048] Figure 10 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0050] Please refer to Figure 1 This invention provides a method for detecting building changes in high-resolution remote sensing images, more specifically, a method for detecting building changes in high-resolution remote sensing images based on residual attention mechanism and high-frequency features, the main steps of which are as follows:

[0051] S1: Obtain a dual-time image dataset of buildings in the same region, preprocess it, divide it into training, validation and test sets, and perform data augmentation.

[0052] Specifically, a dual-timeframe building dataset for the same region was acquired. The original dataset underwent data preprocessing, including cropping, and then data augmentation, such as rotation, was applied to the preprocessed dataset to improve sample data diversity. The open-source building dataset, WHU-CD, was used. WHU-CD records images from 2012 and 2016, containing complex scenes such as villages, farms, residential areas, commercial areas, and industrial areas. The WHU-CD dataset contains five images: two original images, building labels for each time period, and a change label. The image size is 32207×15354 pixels, with a spatial resolution of 0.3m. The WHU-CD dataset was randomly cropped to 256×256 pixel images in a 7:1:2 ratio, resulting in 5460 training data, 780 validation data, and 1560 test data. Data augmentation was performed on the image data in both the training and validation sets using methods such as translation, rotation, scaling, vertical mirroring, illumination adjustment, and the addition of salt-and-pepper noise to obtain augmented image data. The labeled data in the training and validation sets are re-encoded and converted into binary images with 0 and 1 encoding.

[0053] S2: Construct a convolutional neural network FERA-Net with an encoder-decoder structure. The encoder of FERA-Net includes a residual attention-guided high-frequency feature enhancement module (AGMF), and the decoder of FERA-Net includes a feature-enhanced skip connection module (FESCM).

[0054] S3: Input the data from the training set and the validation set into FERA-Net, extract multi-level features in the encoder, aggregate full-scale information in the decoder to better recover the feature map, design a hybrid loss function, optimize network training, and verify the training results through the validation set. After training is completed, the trained network model is obtained.

[0055] Network structure such as Figure 2 As shown, the network employs a shared-weights twin structure to extract bitemporal features. Bitemporal remote sensing images with RGB three-channel bands and a size of 3×H×W from the training set are input into the convolutional neural network FERA-Net for training. The convolutional neural network mainly includes an encoder and a decoder.

[0056] like Figure 3 As shown, in the encoder section, five feature maps are obtained through five consecutive residual attention-guided high-frequency feature enhancement modules. The residual attention-guided high-frequency feature enhancement module includes a residual spatial attention feature extraction stage (…). Figure 4 ) and channel attention high-frequency feature enhancement stage ( Figure 5The residual spatial attention feature extraction stage mainly consists of a mask branch, a trunk branch, and residual units. The trunk branch uses continuous residual units to extract features, while the mask branch uses a top-down structure to generate a mask with spatial information and uses a residual connection to refine the information about changing buildings in the remote sensing image. In the channel attention high-frequency feature enhancement stage, the omnidirectional Sobel operator is used to extract high-frequency information representing building edges in the remote sensing image, and the channel attention mechanism is used to add the extracted high-frequency information to the network, enhancing the network's ability to extract high-frequency information. In the encoder part of FERA-Net, the feature maps obtained from each layer are used as input feature maps for the next layer, and the dimensions of each feature map are as follows: .

[0057] like Figure 6 As shown, in the decoder section, the local and global information in the dual-temporal feature maps obtained from each layer is enhanced through a feature-enhanced skip connection module, and the full-scale feature maps are aggregated for feature recovery. The first step in aggregating the full-scale information... i Layer feature map The specific calculation formula is as follows:

[0058]

[0059] in, Indicates the number of downsampling layers. N Indicates the total number of feature layers. Indicates the extracted first Layered double-time image feature map, k =1, ..., i -1, This indicates the feature enhancement module. Indicates the downsampling operator. This represents the activation function and the normalization function. Describe the upsampling operator.

[0060] Considering the structural characteristics of the FERA-Net network and the specific requirements of building change detection, where the number of changed pixels in the detection result is far less than the number of background pixels, and to better focus on building change information and prior building boundary information, this invention designs a hybrid loss function. The hybrid loss function can be described as follows:

[0061]

[0062] in, Represents the principal loss function. Represents the edge loss function. , This is a proportionality coefficient. In this embodiment, Set to 0.8, Set it to 0.2.

[0063] For the main loss function in a building, the focal loss function is used, and its formula is as follows:

[0064]

[0065] For the building edge loss function, the mean squared loss function (MSE Loss) is used, and its formula is as follows:

[0066]

[0067] Using the pre-constructed FERA-Net network framework, we learn from building data of various types with different structures, shapes, and sizes. Preprocessed training and validation sets are simultaneously fed into the FERA-Net network. The training set is used to train the network to learn building features, while the validation set is used to verify the accuracy of the learned features. The results from the validation set are fed back into the network, allowing it to continuously update its parameters during learning to obtain the optimal learning parameters. Through this iterative process, the optimal network model can be obtained.

[0068] In this embodiment, to test the performance of the proposed model, the following four evaluation metrics are used: accuracy, recall, F1 score, and intersection-over-union ratio (IoU). Specifically, these evaluation metrics can be expressed by the following formulas:

[0069]

[0070]

[0071]

[0072]

[0073] S4: Input the data from the test set into the trained network model to predict the building change detection results, test the model performance, and obtain the model accuracy evaluation results.

[0074] Specifically, the dual-time image data from the test set is input into the trained FERA-Net network model to detect building changes and obtain building detection results.

[0075] To better evaluate the detection accuracy of the model in this invention, four evaluation indicators were calculated and compared with mainstream building change detection methods (models). The comparison results are shown in Table 1 and... Figure 7 , 8 As shown:

[0076] Table 1. Comparison of accuracy results between the method of this invention and mainstream building change detection methods (models).

[0077]

[0078] From Table 1, Figure 7 and Figure 8 As can be seen, compared with mainstream building change detection methods (models), the high-resolution remote sensing image building change detection method (FERA-Net) based on residual attention mechanism and high-frequency features is more effective in building change detection tasks.

[0079] The following describes a high-resolution remote sensing image building change detection device provided by the present invention. The high-resolution remote sensing image building change detection device described below can be referred to in correspondence with the high-resolution remote sensing image building change detection method described above.

[0080] like Figure 9 As shown, a high-resolution remote sensing image building change detection device, specifically a high-resolution remote sensing image building change detection device based on residual attention mechanism and high-frequency features, includes the following modules:

[0081] Image acquisition and preprocessing module 01 is used to acquire a dual-time image dataset of buildings in the same region, preprocess it, divide it into training set, validation set and test set, and perform data augmentation processing.

[0082] Network building module 02 is used to build a convolutional neural network FERA-Net with an encoder-decoder structure. The encoder of FERA-Net includes a high-frequency feature enhancement module guided by residual attention, and the decoder of FERA-Net includes a feature enhancement skip connection module.

[0083] The network training module 03 is used to input data from the training set and the validation set into FERA-Net, extract multi-level features in the encoder, aggregate full-scale information in the decoder to better recover the feature map, design a hybrid loss function, optimize network training, and verify the training results through the validation set. After training is completed, the trained network model is obtained.

[0084] The network testing module 04 is used to input the data from the test set into the trained network model to predict the building change detection results, test the model performance, and obtain the model accuracy evaluation results.

[0085] like Figure 10As shown, a schematic diagram of the physical structure of an electronic device is illustrated. The electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call the logical instructions in the memory 630 to execute the steps of the above-described high-resolution remote sensing image building change detection method, specifically including: acquiring a dual-temporal image building dataset of the same area, preprocessing it and dividing it into a training set, a validation set and a test set, and performing data augmentation processing; constructing a convolutional neural network FERA-Net with an encoder-decoder structure, wherein the encoder of FERA-Net includes a residual attention-guided high-frequency feature enhancement module, and the decoder of FERA-Net includes a feature enhancement skip connection module; inputting the data from the training set and the validation set into FERA-Net, extracting multi-level features in the encoder, aggregating full-scale information in the decoder to better recover the feature map, designing a hybrid loss function, optimizing network training, and verifying the training results through the validation set, obtaining a trained network model after training; inputting the data from the test set into the trained network model to predict the building change detection results, testing the model performance, and obtaining the model accuracy evaluation results.

[0086] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] As an optional implementation, this embodiment of the invention also provides a storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the above-described high-resolution remote sensing image building change detection method, specifically including: acquiring a dual-temporal image building dataset of the same region, preprocessing it, dividing it into a training set, a validation set, and a test set, and performing data augmentation; constructing a convolutional neural network FERA-Net with an encoder-decoder structure, wherein the encoder of FERA-Net includes a residual attention-guided high-frequency feature enhancement module, and the decoder of FERA-Net includes a feature-enhanced skip connection module; inputting data from the training set and the validation set into FERA-Net, extracting multi-level features in the encoder, aggregating full-scale information in the decoder to better recover the feature map, designing a hybrid loss function, optimizing network training, and validating the training results through the validation set; after training, obtaining a trained network model; inputting data from the test set into the trained network model to predict building change detection results, testing model performance, and obtaining model accuracy evaluation results.

[0088] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0089] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0090] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A high-resolution remote sensing image building change detection method, characterized in that, The method comprises the following steps: S1: Obtain a dual-time building data set of the same region, divide it into a training set, a validation set and a test set after preprocessing, and perform data enhancement processing; S2: Construct a convolutional neural network FERA-Net with an encoder-decoder structure, wherein the encoder of the FERA-Net comprises a residual attention-guided high-frequency feature enhancement module, and the decoder of the FERA-Net comprises a feature-enhanced skip connection module; S3: input the data in the training set and the validation set into the FERA-Net, extract multi-level features in the encoder, aggregate full-scale information in the decoder to better restore the feature map, design a hybrid loss function, optimize network training, and verify the training result through the validation set; after training is completed, a trained network model is obtained; S4: input the data in the test set into the trained network model to predict the building change detection result, test the model performance, and obtain the model precision evaluation result; In step S2, the residual attention-guided high-frequency feature enhancement module in the encoder part enhances the high-frequency features of the remote sensing image and excavates multi-level features to obtain low-level spatial information and deep semantic information; The residual attention-guided high-frequency feature enhancement module comprises a residual spatial attention feature extraction stage and a channel attention high-frequency feature enhancement stage; The residual spatial attention feature extraction stage is composed of a mask branch, a trunk branch and a residual unit, the spatial attention mask is generated in the mask branch through a max pooling layer, two residual units, up-sampling, a BR block and a Sigmoid function, the features extracted by the trunk branch are multiplied with the spatial attention mask generated by the mask branch, and then the residual addition operation is performed on the features extracted by the trunk branch to enhance the information of the key region in the feature map; In the channel attention high-frequency feature enhancement stage, the high-frequency information in the remote sensing image is extracted through an omnidirectional Sobel operator, and the high-frequency features in the feature map are enhanced through the channel attention mechanism; In the decoder part, the multi-level feature maps obtained in the encoder part are enhanced through the feature-enhanced skip connection module to enhance the local information and global information in the feature map, and the full-scale information is aggregated through the full-scale skip connection for feature map restoration; The aggregated full-scale information is aggregated by full-scale jump connection i Layer feature map The specific calculation formula is as follows: wherein, denotes the number of down-sampling layers, N denotes the total number of feature map layers, denotes the extracted first layered dual-time image feature map, k = 1,..., i - 1, denotes a feature enhancement module, denotes a down-sampling operator, denotes an activation function and a normalization function, denotes an up-sampling operator.

2. The high resolution remote sensing image building change detection method according to claim 1, wherein, Step S1 specifically comprises: The original data set is cropped to HxW pixels; The cropped data set is divided into a training set, a validation set and a test set; The data in the training set and the validation set are enhanced by translation, rotation and vertical mirroring; the label data is re-encoded and converted into a binary image with 0 and 1 encoding.

3. The high resolution remote sensing image building change detection method according to claim 1, wherein, The network structure of the FERA-Net adopts a shared weight twin structure to extract dual-time features. 4.The high-resolution remote sensing image building change detection method of claim 1, wherein, In step S3, the mixed loss function comprises a main body loss function and an edge loss function; the main body loss function is used for supervising the main body part of the building and uses a Focal Loss function; and the edge loss function is used for supervising the edge part of the building and uses an MSE Loss function.

5. The high resolution remote sensing image building change detection method of claim 1, wherein, The expression of the mixed loss function is as follows: wherein, represents the main loss function, represents the edge loss function, , is a proportionality coefficient.

6. A device for detecting building changes in high resolution remote sensing images implementing the method according to any one of claims 1 to 5, characterized in that, The method comprises the following steps: An image acquisition and preprocessing module is configured to acquire a dual-time building dataset of the same region, divide the dataset into a training set, a validation set and a test set after preprocessing, and perform data enhancement processing; A network construction module is configured to construct a convolutional neural network FERA-Net with an encoder-decoder structure, wherein the encoder of the FERA-Net comprises a residual attention-guided high-frequency feature enhancement module, and the decoder of the FERA-Net comprises a feature-enhanced skip connection module; A network training module is configured to input data in the training set and the validation set into the FERA-Net, extract multi-level features in the encoder, aggregate full-scale information in the decoder to better restore the feature map, design a mixed loss function, optimize network training, verify the training result through the validation set, and obtain a trained network model after the training is completed; A network testing module is configured to input data in the test set into the trained network model to predict a building change detection result and test the performance of the model, and obtain a model precision evaluation result.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the building change detection method for high-resolution remote sensing images according to any one of claims 1-5 when executing the program.

8. A storage medium having stored thereon a computer program, characterized in that The computer program implements the steps of the building change detection method for high-resolution remote sensing images according to any one of claims 1-5 when executed by the processor.

Images