Remote sensing image building segmentation method based on deep learning technology

Abstract

The invention relates to a remote sensing image building segmentation method based on a deep learning technology, and the method comprises the following steps: (1) collecting and preprocessing remote sensing image data, and constructing a training set, a verification set and a test set containing building labels based on the remote sensing image data; (2) constructing a building segmentation model based on the improved HR-Net; (3) introducing a pseudo convolution attention module PCAM into the encoder; (4) introducing a dynamic hybrid cavity convolution (DMAC) module into the decoder; (5) training the building segmentation model by using a training set; and (6) performing performance verification on the trained building segmentation model by using the test set, and outputting a building segmentation result. The method has the advantages that in sudden disaster rescue, the building segmentation model can play a key role in quickly positioning a disaster-affected building, evaluating the influence of a disaster, guiding a rescue path and distributing rescue resources, the rescue efficiency and effect are improved, and casualties and property losses are reduced.

Description

Technical Field

[0001] This invention relates to a method for building segmentation in remote sensing images based on deep learning technology, belonging to the field of remote sensing image processing. Background Technology

[0002] Semantic segmentation of buildings in remote sensing images has become a widely studied research area in recent years. Images of the Earth's surface acquired through remote sensing equipment such as aircraft and satellites typically possess multispectral, high-resolution, and wide-area coverage characteristics. Accurate extraction of buildings from remote sensing images has significant practical application value, and accurate building segmentation is increasingly being used in geographic information systems, urban planning, natural disaster management, and logistics.

[0003] With the continuous development of remote sensing satellite and aerospace technologies, remote sensing imaging technology has become increasingly sophisticated, significantly improving the quality of remote sensing images. This has laid a solid foundation for the widespread application of remote sensing images. In recent years, China has launched a series of high-resolution remote sensing satellites, such as Gaofen-1 to Gaofen-6, providing sub-meter spatial resolution images and greatly enhancing the ability to observe details of the Earth's surface. Furthermore, advancements in multispectral and hyperspectral imaging technologies, along with the development of SAR technology, have enabled images to capture more spectral information, facilitating more refined classification of ground features and material identification. This allows remote sensing imaging to overcome the limitations of clouds and nighttime conditions, achieving all-weather, high-resolution surface observation.

[0004] Traditional semantic segmentation methods are generally based on machine learning and image processing algorithms. Pixel-based classification methods assign each pixel in an image to a different category. These methods typically use machine learning algorithms to classify pixel-level features, such as Support Vector Machines, k-Nearest Neighbors, and Random Forests. Image feature-based methods segment remote sensing images into different regions and then classify these regions. Traditional segmentation algorithms such as region growing, thresholding, edge detection, and HOG are commonly used during segmentation. Traditional algorithms often rely on hand-designed features and rules, and their effectiveness is limited when dealing with complex terrain features or highly variable remote sensing images. Deep learning methods have stronger representation learning capabilities and can handle more complex semantic segmentation tasks, thus their use in remote sensing image semantic segmentation is becoming increasingly widespread.

[0005] Semantic segmentation of remote sensing images based on deep learning has developed rapidly in recent years and has achieved significant breakthroughs. However, remote sensing images are high-resolution images, which means that networks trained on ordinary images need to handle more pixels. In addition, the class imbalance problem in remote sensing images can easily cause the model to be more biased towards common classes, resulting in a decline in overall performance. Finally, remote sensing images are easily affected by changes in satellite attitude and atmospheric interference during acquisition, resulting in unstable targets in the images. At the same time, the cost of labeling remote sensing images is expensive, and it is very difficult to obtain large-scale remote sensing image labeling data.

[0006] In 2015, Long et al. proposed fully convolutional neural networks, which for the first time introduced fully convolutional layers into deep learning models, achieving end-to-end pixel-level segmentation and breaking through the limitations of image segmentation methods that relied on manually designed features and intermediate steps. In the same year, Ronneberger et al. proposed U-Net to address application problems in fields such as medical image segmentation. U-Net's network structure consists of a symmetrical encoder and decoder, while also introducing skip connections and multi-scale feature maps. These deep learning methods improved the automation, accuracy, and robustness of segmentation, but there is still room for improvement in the segmentation of building targets. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this method improves upon the high-resolution semantic segmentation algorithm of HR-Net, proposing a PD-DBnet remote sensing image building segmentation algorithm. Considering the problems of jagged edges in building target segmentation, incomplete overall segmentation of building targets, inconsistent building colors and shapes, and scattered distribution, and taking into account the technical characteristics of convolutional neural networks in deep learning, a personalized deep learning architecture is designed and trained to obtain a personalized remote sensing image building segmentation model, achieving accurate segmentation of building target contours.

[0008] This invention provides a method for building segmentation in remote sensing images based on deep learning technology. The technical solution of this invention is as follows: A deep learning-based method for building segmentation in remote sensing images includes the following steps: (1) Collect and preprocess remote sensing image data, and construct training, validation and test sets containing building labels based on the remote sensing image data; (2) Construct a building segmentation model based on the improved HR-Net, wherein the building segmentation model includes an encoder and a decoder; (3) Introduce a pseudo-convolutional attention module (PCAM) into the encoder to enhance the ability to extract building edges and details from the feature map; (4) Introduce a Dynamic Hybrid Dilated Convolutional Module (DMAC) in the decoder to improve the continuity and integrity of building edges during upsampling; (5) Train the building segmentation model using the training set and adjust the hyperparameters using the validation set; (6) Use the test set to verify the performance of the trained building segmentation model. If it passes the test, save the building segmentation model and output the building segmentation results; otherwise, return to step (5).

[0009] In step (3), the pseudo-convolutional attention module (PCAM) includes: Multi-layer attention stacking unit is used to stack channels and assign weights to feature maps of different sizes; Multi-scale large-kernel convolutional fusion units are used to extract feature information from different receptive fields.

[0010] The specific stacking formula for the multi-layer attention stacking unit is as follows: Assuming a given input feature map The superposition of multiple attention layers is expressed as Equation (1), and the multiplication of weight matrices is expressed as Equation (2). (1) (2) in, It is represented as a weight matrix after multiple feature maps are superimposed. The total number of feature maps, For the first Channel addition, For the first Each feature map For 1×1 convolution, This is a matrix multiplication operation.

[0011] The specific fusion steps of the multi-scale large-kernel convolutional fusion unit are as follows: By fusing large kernel convolutions of different scales, the dimensionality is compressed through convolution, and feature information is output. This is achieved using large kernel convolution. Multiple convolutional kernels of 7×7, 9×9, and 13×13 are set to capture multi-scale information. Different sizes of convolutional kernels capture features of different ranges. This part is expressed as formula (3). The multiplication of weight matrices can be expressed as formula (4). (3) (4) in, This represents the weight matrix after fusing convolutional kernels of different scales. The total number of feature maps, Represents convolution kernels of different scales. The value is the kernel size, and it takes the value of . , For the first Each feature map For 1×1 convolution, This refers to matrix multiplication. The 7×7 convolution kernel is replaced with a linear combination of 1×7 and 7×1 convolution kernels; the 9×9 convolution kernel is replaced with a linear combination of 1×9 and 9×1 convolution kernels; the 13×13 convolution kernel is replaced with a linear combination of 1×13 and 13×1 convolution kernels; through asymmetric decomposition, the convolution kernels are all split into linear combinations of two one-dimensional convolutions.

[0012] In step (4), the Dynamic Hybrid Dilated Convolution Module (DMAC) is composed of parallel dilated convolutions, and the formula for calculating the receptive field of the dilated convolution is as shown in formula (5). (5) in, The kernel size is [size]. The total number of feature maps, Indicates the number of convolutional layers, with a value range of 1. , This represents the initial expansion rate, with an initial value of 1. Indicates the size of the receptive field; The feature maps are dilated and convolved with different dilation rates and then linearly combined with the original feature maps one by one between channels to dynamically select features of different channels. Finally, feature fusion is performed. The linear combination is expressed by formula (6). (6) in, The combined pixels, The scalar value represents the weight coefficients of the feature map after dilated convolution. This is the offset.

[0013] In step (5), during training, the cross-entropy loss function used in the training process is defined as follows: (The original text contains some inconsistencies and unclear grammatical structures. A more accurate translation would require the full context.) ; in, The values ​​are the actual label values: 1 for buildings and 0 for non-buildings. The probability value for predicting a building. It represents the difference between the true sample label and the predicted probability.

[0014] In step (6), the specific verification steps are as follows: If we define buildings as positive samples and non-buildings as negative samples, then: : Indicates the number of samples that were actually buildings but were correctly predicted as buildings; : This indicates the number of samples that were actually buildings but were incorrectly predicted as non-buildings; : This indicates the number of samples that were actually not buildings but were incorrectly predicted as buildings; The calculation formulas for various indicators are as follows: Precision P represents the probability that a sample is actually positive out of all samples predicted as positive. ; Recall R represents the probability that a pixel in the sample that is actually a building is predicted to belong to a building. ; The F1 score represents a balance between precision and recall, maximizing both simultaneously. ; The mean IoU and mIoU represent the average of the cross-union ratios for positive and negative samples, respectively. ; Where k is the number of categories, and for semantic segmentation of buildings, there are k+1 categories.

[0015] The advantages of this invention are: (1) By combining urban planning schemes with building segmentation models, basic information on the distribution of buildings in the urban area can be obtained in a timely manner, and more scientific and reasonable planning schemes can be formulated to improve the efficiency and effectiveness of urban management and development.

[0016] (2) By combining intelligent transportation systems with building segmentation models, the spatial relationship and distribution between roads and buildings can be obtained in a timely manner, optimizing traffic flow, improving road planning and traffic management, and enhancing the efficiency and safety of urban transportation.

[0017] (3) In the event of a sudden disaster, building segmentation models can play a key role in quickly locating affected buildings, assessing the impact of the disaster, guiding rescue routes and allocating rescue resources, thereby improving rescue efficiency and effectiveness and reducing casualties and property losses. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0020] See Figure 1 This invention relates to a method for building segmentation in remote sensing images based on deep learning technology, comprising the following steps: (1) Collect and preprocess remote sensing image data, and construct training, validation, and test sets containing building labels based on the remote sensing image data; Collect public remote sensing image datasets according to the application objectives, retaining only building targets. The data format includes urban and non-urban areas; (2) Construct a building segmentation model based on the improved HR-Net, which includes an encoder and a decoder. HR-Net has the following advantages in high-resolution remote sensing image detection: 1. Maintaining high-resolution features: HR-Net maintains high-resolution feature maps in a single branch throughout the network, which can capture more detailed information and improve detection accuracy, especially in the extraction of details such as building edges.

[0021] 2. Multi-resolution fusion: By using parallel multi-resolution branches and frequent multi-scale feature fusion, HR-Net integrates information from different scales, improving its adaptability to complex scenes and giving the model better robustness and generalization ability on high-resolution images.

[0022] 3. Efficient feature representation: Multi-resolution feature exchange and fusion are performed at each stage, which enhances the richness and diversity of feature representation and can better handle various details and complex structures in high-resolution images.

[0023] 4. Precise localization: Maintaining high-resolution feature maps makes HR-Net more accurate in target localization, suitable for fine detection of small targets and edges in remote sensing images, such as buildings, roads, and water bodies.

[0024] This invention improves the feature extraction layer in the HR-Net high-resolution network. Taking into account the device environment and image training difficulty, this invention uses HR-Net48 as the special extraction network for feature extraction. However, during image feature extraction, it is difficult to effectively determine the weights and attention levels in the image, and image edges are difficult to capture.

[0025] (3) A pseudo-convolutional attention module (PCAM) is introduced into the encoder to enhance the extraction capability of building edges and details in the feature maps. In the encoder part, for high-resolution image feature extraction, HR-Net processes feature maps of different resolutions simultaneously through multiple parallel branches. The parallel structure ensures that it always maintains high-resolution features, while differentiating low-resolution branches and fusing multi-scale information, making it perform well on high-resolution images. The parallel structure performs downsampling through multiple paths, and the information of the branch structure in the same stage is independent. Based on this, information is extracted from the feature maps in the same stage and shared with all feature maps, which can further improve the accuracy of feature map information and thus improve segmentation accuracy.

[0026] The purpose of pseudo-convolutional attention is to improve the local attention of the network by computing the analogy of attention structures through the stacking of inexpensive feature maps. By stacking multiple layers of feature maps, it helps to quickly expand the channels containing useful information features, while the changes in channels lacking useful information are relatively small. Therefore, this structure can positively help strengthen and highlight feature information, while reducing the introduction and interference of irrelevant information.

[0027] (4) A Dynamic Hybrid Dilated Convolutional (DMAC) module is introduced into the decoder to improve the continuity and integrity of building edges during upsampling. The HR-Net decoder is designed with a dynamic hybrid dilated convolutional structure, which is applied to the HR-Net decoder to replace the overly abrupt upsampling method in HR-Net, improve the continuity of building edge segmentation, and reduce jagged segmentation. Dynamic hybrid dilated convolution is a combination of dilated convolutions with different dilation rates. Dilated convolutions are usually used to increase the receptive field, but they are prone to grid effects when processing high-resolution images. The grid effect is gradually amplified through multiple convolutions, and the discontinuity becomes more obvious. Fusing dilated convolutions of different scales can alleviate this problem. At the same time, an exponentially increasing dilation rate is used. The high dilation rate expands the network's receptive field to the global level, while the low dilation rate and regular convolutions preserve the most basic features, thereby enabling the network to better understand the contextual information. (5) Train the building segmentation model using the training set and adjust the hyperparameters using the validation set; (6) Use the test set to verify the performance of the trained building segmentation model. If it passes the test, save the building segmentation model and output the building segmentation results; otherwise, return to step (5).

[0028] In step (3), the pseudo-convolutional attention module (PCAM) includes: Multi-layer attention stacking unit is used to stack channels and assign weights to feature maps of different sizes; Multi-scale large-kernel convolutional fusion units are used to extract feature information from different receptive fields.

[0029] The specific stacking formula for the multi-layer attention stacking unit is as follows: Assuming a given input feature map The superposition of multiple attention layers is expressed as Equation (1), and the multiplication of weight matrices is expressed as Equation (2). (1) (2) in, It is represented as a weight matrix after multiple feature maps are superimposed. The total number of feature maps, For the first Channel addition, For the first Each feature map For 1×1 convolution, This is a matrix multiplication operation.

[0030] The specific fusion steps of the multi-scale large-kernel convolutional fusion unit are as follows: By fusing large kernel convolutions of different scales, the dimensionality is compressed through convolution, and feature information is output. This is achieved using large kernel convolution. By fusing large kernel convolutions of different scales, the dimensionality is compressed through convolution, and feature information is output. Using large kernel convolutions can increase the network's field of view, extract a wider range of features, and make the network more suitable for high-resolution remote sensing images.

[0031] Multiple convolutional kernels of 7×7, 9×9, and 13×13 are set to capture multi-scale information. Different sizes of convolutional kernels capture features of different ranges. This part is expressed as formula (3). The multiplication of weight matrices can be expressed as formula (4). (3) (4) in, This represents the weight matrix after fusing convolutional kernels of different scales. The total number of feature maps, Represents convolution kernels of different scales. The value is the kernel size, and it takes the value of . , For the first Each feature map For 1×1 convolution, This is a matrix multiplication operation; the superposition of multiple feature maps is equivalent to channel feature fusion. The network performs a feature selection at the channel level, emphasizing positively correlated feature channels and discarding negatively correlated channels.

[0032] Furthermore, to reduce model parameters and considering that most building samples are rectangular, 1×7 and 7×1 linear combination convolutional kernels are used instead of 7×7 convolutional kernels; 1×9 and 9×1 linear combination convolutional kernels are used instead of 9×9 convolutional kernels; and 1×13 and 13×1 linear combination convolutional kernels are used instead of 13×13 convolutional kernels. Through asymmetric decomposition, all convolutional kernels are split into linear combinations of two one-dimensional convolutions. This approach retains the wide receptive field of large kernel convolutions while significantly reducing computational costs, and simultaneously enhances the edge features of rectangular buildings. The linear combination convolutional kernels can highlight local features in both horizontal and vertical directions, preserving more building edges while saving computational overhead.

[0033] In step (4), the Dynamic Hybrid Dilated Convolution Module (DMAC) is composed of parallel dilated convolutions, and the formula for calculating the receptive field of the dilated convolution is as shown in formula (5). (5) in, The kernel size is [size]. The total number of feature maps, Indicates the number of convolutional layers, with a value range of 1. , This represents the initial expansion rate, with an initial value of 1. Indicates the size of the receptive field; The feature maps are dilated and convolved with different dilation rates and then linearly combined with the original feature maps one by one between channels to dynamically select features of different channels. Finally, feature fusion is performed. The linear combination is expressed by formula (6). (6) in, The combined pixels, The scalar value represents the weight coefficients of the feature map after dilated convolution. This is the offset.

[0034] The feature map of the DMAC module is upsampled using a layer-by-layer bilinear interpolation method. Starting from the bottom layer feature map, the feature map of the current layer is upsampled by bilinear interpolation to double its resolution. The upsampled feature map is then concatenated with the feature map of the previous layer.

[0035] By improving the HR-Net network and progressively extracting feature maps at each stage, more accurate feature extraction of building targets in remote sensing images can be achieved. The HR-Net network has four stages, each extending downwards by one branch. Simultaneously, the feature maps of each stage are input into the PCAM module for attention stacking operations to obtain corresponding attention weights for each stage. These weights are then shared across all feature maps of that stage, thereby enhancing feature extraction capabilities and improving the model's focus on feature maps at different scales. After network encoding, the feature maps of each branch are further enhanced by the DMAC module, specifically improving the edge integrity of the feature maps. Finally, layer-by-layer fusion upsampling is performed, progressively fusing information from each layer's feature maps at different scales, transitioning from low-level features to high-level features. This better converts edge texture and other information into semantic information while minimizing information loss.

[0036] In step (5), during training, the cross-entropy loss function used in the training process is defined as follows: (The original text contains some inconsistencies and unclear grammatical structures. A more accurate translation would require the full context.) ; in, The values ​​are the actual label values: 1 for buildings and 0 for non-buildings. The probability value for predicting a building. It represents the difference between the true sample label and the predicted probability.

[0037] In step (6), the specific verification steps are as follows: If we define buildings as positive samples and non-buildings as negative samples, then: : Indicates the number of samples that were actually buildings but were correctly predicted as buildings; : This indicates the number of samples that were actually buildings but were incorrectly predicted as non-buildings; : This indicates the number of samples that were actually not buildings but were incorrectly predicted as buildings; The calculation formulas for various indicators are as follows: Precision P represents the probability that a sample is actually positive out of all samples predicted as positive. ; Recall R represents the probability that a pixel in the sample that is actually a building is predicted to belong to a building. ; The F1 score represents a balance between precision and recall, maximizing both simultaneously. ; The mean IoU and mIoU represent the average of the cross-union ratios for positive and negative samples, respectively. ; Where k is the number of categories, and for semantic segmentation of buildings, there are k+1 categories.

[0038] In summary, this invention provides a method for building segmentation in remote sensing images based on deep learning technology. Its core working principle is to achieve accurate, continuous, and complete segmentation of building targets in remote sensing images by using an improved HR-Net network structure, combined with a pseudo-convolutional attention module (PCAM) and a dynamically hybrid dilated convolutional module (DMAC). Its working principle is as follows: 1. Data Preparation and Preprocessing Data collection: Acquire high-resolution remote sensing images (such as Gaofen series satellite images) from public remote sensing datasets, which contain building targets and cover both urban and non-urban areas.

[0039] Data preprocessing: The images are cropped, downsampled (e.g., reduced from 0.075 meters to 0.3 meters resolution), and standardized to construct a dataset with pixel-level labels for buildings, which is then divided into training, validation, and test sets.

[0040] 2. Construct an improved HR-Net building segmentation model Model structure: Based on the HR-Net48 architecture, it retains the advantages of multi-branch parallel processing, high-resolution feature preservation and multi-scale fusion.

[0041] Encoder section: A pseudo-convolutional attention module (PCAM) is introduced to enhance the feature extraction capabilities for building edges and details.

[0042] Decoder section: Introducing a Dynamic Hybrid Dilated Convolutional Module (DMAC) to improve edge continuity and integrity during upsampling and reduce jagged segmentation.

[0043] 3. Pseudo-convolutional attention module (PCAM) Function: Enhances the encoder's ability to perceive building edges and details.

[0044] Composition: Multi-layer attention stacking unit: Stacks and weights multiple feature maps of different sizes in the same stage to enhance useful feature channels and suppress irrelevant information.

[0045] Multi-scale large-kernel convolutional fusion unit: Uses large convolutional kernels such as 7×7, 9×9, and 13×13 to extract multi-scale features and enhance the receptive field. To reduce computation, the 7×7 convolution is split into a linear combination of 1×7 and 7×1 to highlight horizontal and vertical edge features.

[0046] 4. How Dynamic Hybrid Dilated Convolutional Module (DMAC) Works Function: Replaces the traditional upsampling method in the decoder, improving the continuity and integrity of building edges.

[0047] Structure: It consists of multiple dilated convolutions with different dilation rates in parallel, fusing multi-scale receptive field features.

[0048] Operating procedures: Perform dilation convolutions on the feature maps with different dilation rates; The results are linearly combined with the original feature map element by element to dynamically select useful features. Finally, feature fusion is performed, and multi-scale information is fused layer by layer by gradually upsampling through bilinear interpolation.

[0049] 5. Model Training and Optimization Loss function: The binary cross-entropy loss function is used to calculate the error between the predicted value and the true label pixel by pixel.

[0050] Training process: Forward and backward propagation are performed using the training set, and hyperparameters (such as learning rate, batch size, etc.) are adjusted using the validation set to optimize model performance.

[0051] 6. Model Validation and Output Performance evaluation: The model's segmentation performance is evaluated by calculating metrics such as precision, recall, F1 score, and mean intersection-union ratio (mIoU) using the test set.

[0052] Output results: If the model performance meets the requirements, save the model and output the building segmentation result image; if it does not meet the requirements, return to the training phase for further optimization.

[0053] This invention achieves accurate, continuous, and complete segmentation of building targets in high-resolution remote sensing images through data preprocessing, an improved HR-Net network structure, the introduction of PCAM and DMAC modules, and a systematic training and validation process. This method has significant application value in fields such as urban planning, intelligent transportation, and disaster relief.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for building segmentation in remote sensing images based on deep learning, characterized in that, Includes the following steps: (1) Collect and preprocess remote sensing image data, and construct training, validation and test sets containing building labels based on the remote sensing image data; (2) Construct a building segmentation model based on the improved HR-Net, wherein the building segmentation model includes an encoder and a decoder; (3) Introduce a pseudo-convolutional attention module (PCAM) into the encoder to enhance the ability to extract building edges and details from the feature map; (4) Introduce a Dynamic Hybrid Dilated Convolutional Module (DMAC) in the decoder to improve the continuity and integrity of building edges during upsampling; (5) Train the building segmentation model using the training set and adjust the hyperparameters using the validation set; (6) Use the test set to verify the performance of the trained building segmentation model. If it passes the test, save the building segmentation model and output the building segmentation results; otherwise, return to step (5).

2. The deep learning-based remote sensing image building segmentation method according to claim 1, characterized in that, In step (3), the pseudo-convolutional attention module (PCAM) includes: Multi-layer attention stacking unit is used to stack channels and assign weights to feature maps of different sizes; Multi-scale large-kernel convolutional fusion units are used to extract feature information from different receptive fields.

3. The deep learning-based remote sensing image building segmentation method according to claim 2, characterized in that, The specific stacking formula for the multi-layer attention stacking unit is as follows: Assuming a given input feature map The superposition of multiple attention layers is expressed as Equation (1), and the multiplication of weight matrices is expressed as Equation (2). （1）； （2）； in, It is represented as a weight matrix after multiple feature maps are superimposed. The total number of feature maps, For the first Channel addition, For the first Each feature map For 1×1 convolution, This is a matrix multiplication operation.

4. The deep learning-based remote sensing image building segmentation method according to claim 3, characterized in that, The specific fusion steps of the multi-scale large-kernel convolutional fusion unit are as follows: By fusing large kernel convolutions of different scales, the dimensionality is compressed through convolution, and feature information is output. This is achieved using large kernel convolution. Multiple convolutional kernels of 7×7, 9×9, and 13×13 are set to capture multi-scale information. Different sizes of convolutional kernels capture features of different ranges. This part is expressed as formula (3). The multiplication of weight matrices can be expressed as formula (4). （3）； （4）； in, This represents the weight matrix after fusing convolutional kernels of different scales. The total number of feature maps, Represents convolution kernels of different scales. The value is the kernel size, and it takes the value of . , For the first Each feature map For 1×1 convolution, This refers to matrix multiplication. The 7×7 convolution kernel is replaced with a linear combination of 1×7 and 7×1 convolution kernels; the 9×9 convolution kernel is replaced with a linear combination of 1×9 and 9×1 convolution kernels; the 13×13 convolution kernel is replaced with a linear combination of 1×13 and 13×1 convolution kernels; through asymmetric decomposition, the convolution kernels are all split into linear combinations of two one-dimensional convolutions.

5. The deep learning-based remote sensing image building segmentation method according to claim 3, characterized in that, In step (4), the Dynamic Hybrid Dilated Convolution Module (DMAC) is composed of parallel dilated convolutions, and the formula for calculating the receptive field of the dilated convolution is as shown in formula (5). （5）； in, The kernel size is [size]. The total number of feature maps, Indicates the number of convolutional layers, with a value range of 1. , This represents the initial expansion rate, with an initial value of 1. Indicates the size of the receptive field; The feature maps are dilated and convolved with different dilation rates and then linearly combined with the original feature maps one by one between channels to dynamically select features of different channels. Finally, feature fusion is performed. The linear combination is expressed by formula (6). （6）； in, The combined pixels, The scalar value represents the weight coefficients of the feature map after dilated convolution. This is the offset.

6. The method for building segmentation in remote sensing images based on deep learning according to claim 1, characterized in that, In step (5), during training, the cross-entropy loss function used in the training process is defined as follows: (The original text contains some inconsistencies and unclear grammatical structures. A more accurate translation would require the full context.) ； in, The values ​​are the actual label values: 1 for buildings and 0 for non-buildings. The probability value for predicting a building. It represents the difference between the true sample label and the predicted probability.

7. The method for building segmentation in remote sensing images based on deep learning according to claim 1, characterized in that, In step (6), the specific verification steps are as follows: If we define buildings as positive samples and non-buildings as negative samples, then: : Indicates the number of samples that were actually buildings but were correctly predicted as buildings; : This indicates the number of samples that were actually buildings but were incorrectly predicted as non-buildings; : This indicates the number of samples that were actually not buildings but were incorrectly predicted as buildings; The calculation formulas for various indicators are as follows: Precision P represents the probability that a sample is actually positive out of all samples predicted as positive. ； Recall R represents the probability that a pixel in the sample that is actually a building is predicted to belong to a building. ; The F1 score represents a balance between precision and recall, maximizing both simultaneously. ； The mean IoU and mIoU represent the average of the cross-union ratios for positive and negative samples, respectively. ； Where k is the number of categories, and for semantic segmentation of buildings, there are k+1 categories.

Images