Remote sensing image shelterbelt semantic segmentation model construction method, segmentation method and device

Abstract

The application relates to the technical field of remote sensing image recognition, and is a windbreak forest semantic segmentation model construction method, a segmentation method and a device, which comprise the following steps: obtaining a remote sensing image to be processed, and performing complete cropping on the remote sensing image to be processed; inputting all the cropped images into a windbreak forest semantic segmentation model to obtain corresponding segmentation prediction results; and splicing all the segmentation prediction results according to the segmentation sequence of the remote sensing image to be processed to obtain a complete prediction result image. The improved main network in the improved DeepLabV3+ network model is used to extract deep feature information of the remote sensing image, the improved ASPP module is used to extract multi-scale context information from the extracted deep feature information, and the attention module is used to learn the relevance between different channels, so that important feature information of the image can be better captured, the discontinuity problem of row-shaped windbreak forest segmentation is relieved, and the windbreak forest segmentation effect is improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image recognition technology, specifically a method, segmentation method, and apparatus for constructing a semantic segmentation model of windbreak forests in remote sensing images. Background Technology

[0002] In recent years, with the rapid development of deep learning technology, significant breakthroughs have been achieved in remote sensing image research. Semantic segmentation, as an important task in the field of computer vision, has been widely applied in various fields such as land cover classification, change detection, urban planning, and environmental monitoring.

[0003] The core objective of semantic segmentation is to accurately classify each pixel in an image into its corresponding semantic category, thereby achieving precise segmentation of different objects and regions within the image. To achieve semantic segmentation, deep learning models are frequently used for this task. Common deep learning models include FCN, U-Net, PSPNet, and DeepLabV3+, which learn image features and contextual information to achieve pixel-level classification of images.

[0004] The rich and complex information contained in remote sensing images presents multiple challenges for semantic segmentation of windbreaks. DeepLabV3+ is commonly used for this task. DeepLabV3+ is a network model with good performance in semantic segmentation, primarily using a backbone network for feature extraction, an encoder module to acquire multi-scale information, and a decoder to fuse low-level and high-level features for segmentation prediction. However, DeepLabV3+, based on the ResNet18 backbone, suffers from a limited receptive field during feature extraction, hindering the acquisition of broader contextual information. The ASPP module uses depthwise separable convolutions with different dilation rates to acquire multi-scale information; however, larger dilation rates can lead to poor performance, potentially causing discontinuous segmentation of row-shaped windbreaks and negatively impacting the overall performance of the network model. Summary of the Invention

[0005] This invention provides a method for constructing a semantic segmentation model of windbreak forests in remote sensing images, a segmentation method, and an apparatus, which overcomes the shortcomings of the prior art. It can effectively solve the problem of discontinuous semantic segmentation of windbreak forests in remote sensing images when using the DeepLabV3+ network model for semantic segmentation, which leads to low segmentation accuracy.

[0006] One of the technical solutions of this invention is achieved through the following measures: a method for constructing a semantic segmentation model of windbreak forests in remote sensing images, comprising: A small batch of historical image samples were obtained and divided into training sample set and test sample set. Each historical image sample included a historical remote sensing image of the windbreak forest area and a label indicating the corresponding semantic segmentation result of the windbreak forest. The improved DeepLabV3+ network model was trained using the training sample set to obtain the windbreak semantic segmentation model. The improved DeepLabV3+ network model is a DeepLabV3+ network model that improves the backbone network and ASPP model. The semantic segmentation model of windbreak forest is trained by iterating through the number of iterations and the loss function to obtain the final semantic segmentation model of windbreak forest. The semantic segmentation model of windbreak forest was tested using a test sample set, and the model was optimized based on the test results.

[0007] The following are further optimizations and / or improvements to the above-mentioned technical solution: The above-mentioned training of the improved DeepLabV3+ network model using the training sample set yields a semantic segmentation model for windbreaks, including: Historical image samples from the training sample set are input into the improved backbone network and the improved ASPP module, respectively, to obtain low-level features and high-level features. The low-level and high-level features are fused using a decoder. The fused features are then mapped to the number of categories required for the windbreak semantic segmentation task. Finally, an image with the same size as the input historical image sample is obtained by upsampling, thus completing the windbreak semantic segmentation. The segmentation results of the windbreak semantic segmentation model are evaluated by combining the loss function, and the model network parameters are updated by backpropagation and optimization algorithm based on the evaluation results.

[0008] The above-mentioned annotation of windbreak forest areas in historical remote sensing images includes using ENVI software to annotate historical remote sensing images, generating a corresponding label for each pixel, storing the generated label file in the form of an XML file, exporting and converting it into a masked image, and obtaining historical remote sensing images with annotated windbreak forest areas.

[0009] The second technical solution of the present invention is achieved through the following measures: a semantic segmentation method for windbreak forests in remote sensing images, comprising: Obtain the remote sensing image to be processed and perform a complete cropping of the remote sensing image to be processed, wherein the cropping size is consistent with the size of the historical image samples in the training sample set; All cropped images are input into the windbreak semantic segmentation model to obtain the corresponding segmentation prediction results, wherein the windbreak semantic segmentation model is obtained by the method described in any one of claims 1 to 4; All segmentation prediction results are stitched together according to the segmentation order of the remote sensing image to be processed to obtain a complete prediction result image.

[0010] The third technical solution of the present invention is achieved through the following measures: a device for constructing a semantic segmentation model of windbreak forests in remote sensing images, comprising: The sample acquisition unit acquires a small batch of historical image samples and divides them into a training sample set and a test sample set. Each historical image sample includes a historical remote sensing image of the windbreak forest area and a label that identifies the corresponding semantic segmentation result of the windbreak forest. The training unit uses the training sample set to train the improved DeepLabV3+ network model to obtain the windbreak semantic segmentation model. The improved DeepLabV3+ network model is a DeepLabV3+ network model that improves the backbone network and ASPP model. The iterative unit trains the windbreak semantic segmentation model based on the number of iterations and the loss function to obtain the final windbreak semantic segmentation model; The test unit uses a test sample set to test the windbreak semantic segmentation model, and optimizes the windbreak semantic segmentation model based on the test results.

[0011] The third technical solution of the present invention is achieved through the following measures: a semantic segmentation device for windbreak forests in remote sensing images, comprising: The image acquisition unit acquires the remote sensing image to be processed and performs a complete cropping of the remote sensing image to be processed, wherein the cropping size is consistent with the size of the historical image samples in the training sample set; The prediction unit inputs all cropped images into the windbreak forest semantic segmentation model to obtain the corresponding segmentation prediction results, wherein the windbreak forest semantic segmentation model is obtained by the method described in any one of claims 1 to 4. The stitching unit stitches together all the segmentation prediction results according to the segmentation order of the remote sensing image to be processed, and obtains a complete prediction result image.

[0012] This invention utilizes the improved backbone network in the improved DeepLabV3+ network model to extract deep feature information from remote sensing images, providing strong support for subsequent segmentation tasks. The improved ASPP module extracts multi-scale contextual information from the extracted deep feature information, and the attention module learns the correlation between different channels to enhance the representation ability of important features. Therefore, it can better capture important feature information of the image, alleviate the problem of discontinuity in the segmentation of row-shaped windbreaks, and thus improve the segmentation effect of windbreaks. Attached Figure Description

[0013] Appendix Figure 1 This is a flowchart of the model construction method of the present invention.

[0014] Appendix Figure 2 This is a schematic diagram of the backbone network structure of the improved DeepLabV3+ network model in this invention.

[0015] Appendix Figure 3 This is a schematic diagram of the overall structure of the improved DeepLabV3+ network model of the present invention.

[0016] Appendix Figure 4 This is a flowchart of the semantic segmentation method for windbreak forests according to the present invention.

[0017] Appendix Figure 5 The images shown are a portion of the original image to be tested and a portion of the windbreak forest segmentation results used in the simulation of this invention in Example 3.

[0018] Appendix Figure 6 This is a partial effect diagram comparing the present invention with other methods.

[0019] Appendix Figure 7 This is a schematic diagram of the structure of the remote sensing image windbreak forest semantic segmentation model construction device of the present invention.

[0020] Appendix Figure 8 This is a schematic diagram of the semantic segmentation device for windbreak forests in remote sensing images according to the present invention. Detailed Implementation

[0021] The present invention is not limited to the following embodiments, and the specific implementation can be determined according to the technical solution of the present invention and the actual situation.

[0022] The present invention will be further described below with reference to embodiments and accompanying drawings: Example 1: As shown in the attached document Figure 1 As shown in the figure, this invention discloses a method for constructing a semantic segmentation model of windbreak forests in remote sensing images, including: Step S110: Obtain a small batch of historical image samples and divide them into training sample set and test sample set. Each historical image sample includes a historical remote sensing image of the windbreak forest area and a label that identifies the corresponding semantic segmentation result of the windbreak forest.

[0023] The above method involves labeling historical remote sensing images using ENVI software, generating a corresponding label for each pixel. The generated label files are stored in XML format, exported, and converted into masked images to obtain historical remote sensing images labeled with windbreak forest areas. Furthermore, because the original historical remote sensing images are too large and have too high resolution to be directly input into the model for training, the historical remote sensing images and their corresponding labels need to be cropped to ensure they correspond spatially. The cropping size and step size are set according to the actual situation. Specifically, the cropped image size can be, but is not limited to, 256×256, and the step size can be, but is not limited to, 128. Parts that cannot be cropped to 256×256 size at the final edge are discarded. Then, the proportion of pixels of each category in the label is calculated. For categories with a proportion lower than a preset threshold, the corresponding images and labels are removed, thus forming a small batch of historical image samples labeled with windbreak forest areas. Finally, the cropped dataset is divided into training and testing sample sets according to the sample partitioning ratio to obtain the final training and testing sample sets. Therefore, this step uses random flipping, cropping, and scaling as data augmentation operations to increase data diversity.

[0024] Step S120: Train the improved DeepLabV3+ network model using the training sample set to obtain the windbreak semantic segmentation model, wherein the improved DeepLabV3+ network model is a DeepLabV3+ network model that improves the backbone network and ASPP model.

[0025] The above-mentioned training of the improved DeepLabV3+ network model using the training sample set yields a semantic segmentation model for windbreaks, including: (1) Input the historical image samples in the training sample set into the improved backbone network and the improved ASPP module respectively to obtain low-level features and high-level features respectively; The aforementioned backbone network is attached. Figure 2As shown, the input image is first preliminarily feature-extracted through initial convolutions, and then downsampled using pooling layers to extract initial low-level features. Subsequently, deep feature extraction is performed in four stages to obtain deep features. Each of the four stages contains two residual blocks, each consisting of a 3×3 convolution + Batch Normalization (BN) + a 5×5 depthwise convolution + BN + a linear layer + GELU + a linear layer + BN. Except for the first stage where the input is directly added to the output via skip connections, the residual connections in the first residual block of stages 2, 3, and 4 consist of 1×1 convolutions + BN, directly adding the input to the output via skip connections. The second residual block also directly adds the input to the output via skip connections, mitigating the vanishing gradient effect. In the second stage, the first convolution uses a stride of 2 for downsampling. This constitutes a deep feature extraction network, the backbone network, capable of capturing a wider range of contextual information, which helps to better understand the semantic relationships and structures in remote sensing images.

[0026] The improved ASPP module is attached. Figure 3 As shown, the algorithm comprises five branches with different structures. The first branch is a 1×1 convolution followed by an attention module. The second to fifth branches consist of 3×3 depthwise separable convolutions. Then, depthwise separable convolutions with dilation rates of 2, 4, 6, and 8 are added progressively for each branch. Finally, the attention module is introduced. The attention module uses average pooling and max pooling to obtain information along the channel dimension. The two pooling operations are concatenated, and then the number of channels C is adjusted using a 1×1 convolution. The kernel size k is adaptively determined, followed by a one-dimensional convolution, and then a sigmoid function is used to learn the channel attention weights. Furthermore, a full-dimensional dynamic convolution is introduced to enhance the adaptability of the features. The result of the full-dimensional dynamic convolution is multiplied by the channel attention weights, and finally added to the original input for output. The outputs of the five branches are concatenated along the channel dimension, and then subjected to a 3×3 convolution for feature fusion to adjust the number of channels. The improved ASPP module uses five branches with different structures to extract more refined multi-scale contextual information from the extracted deep features, thereby obtaining high-level features. The attention module enhances the correlation between different channels, which helps the model better capture the feature information of the windbreak forest and improve the segmentation effect.

[0027] (2) The low-level features and high-level features are fused using the decoder, and the fused features are mapped to the number of categories required for the windbreak semantic segmentation task. The images are then upsampled to obtain an image with the same size as the input historical image samples to complete the windbreak semantic segmentation. The decoder module described above integrates low-level features from the improved backbone network and high-level features from the improved ASPP module, and finally maps the features to the number of categories required for the semantic segmentation task to obtain the final segmentation result.

[0028] (3) The segmentation results of the windbreak semantic segmentation model are evaluated by combining the loss function, and the model network parameters are updated by backpropagation and optimization algorithm based on the evaluation results. The loss function is selected as needed.

[0029] Step S130: Train the windbreak semantic segmentation model based on the number of iterations and the loss function to obtain the final windbreak semantic segmentation model.

[0030] The above-mentioned iteration count is used to iteratively train the windbreak semantic segmentation model. The model can be trained using LinearLR and PolyLR learning strategies. For example, if the number of iterations is 320,000, the LinearLR learning strategy is used for the first 8,000 iterations, and the PolyLR learning strategy is used from the 8,000th iteration until the end, so as to effectively train the model. The above steps are repeated for each iteration.

[0031] Step S140: Test the windbreak semantic segmentation model using a test sample set, and optimize the model based on the test results. Here, the test sample set is input into the pre-trained windbreak semantic segmentation model for testing, obtaining the segmentation results of windbreaks in remote sensing images. The model's performance is validated on the test set to evaluate the model's training effect and generalization ability. If the evaluation result is unsatisfactory, the model network parameters can be adjusted in reverse, and retraining can be performed.

[0032] In summary, this invention discloses a method for constructing a semantic segmentation model for windbreaks in remote sensing images. It utilizes the improved backbone network in the improved DeepLabV3+ network model to extract deep feature information from remote sensing images, providing strong support for subsequent segmentation tasks. The improved ASPP module extracts multi-scale contextual information from the extracted deep feature information, and the attention module learns the correlation between different channels, enhancing the representation ability of important features. Therefore, it can better capture important feature information of the image, alleviate the discontinuity problem in the segmentation of row-shaped windbreaks, and thus improve the segmentation effect of windbreaks.

[0033] Example 2: As shown in the attached document Figure 4 As shown in the figure, this invention discloses a semantic segmentation method for windbreak forests in remote sensing images, including: Step S210: Obtain the remote sensing image to be processed and perform a complete cropping of the remote sensing image to be processed, wherein the cropping size is consistent with the size of the historical image samples in the training sample set; Step S220: Input all cropped images into the windbreak semantic segmentation model to obtain the corresponding segmentation prediction results, wherein the windbreak semantic segmentation model is obtained using the method described in the above embodiments; Step S230: All segmentation prediction results are stitched together according to the segmentation order of the remote sensing image to be processed to obtain a complete prediction result image.

[0034] Example 3: Experiments were conducted using PyTorch, with the model implemented based on the mmsegmentation library for segmentation. The specific experimental platform was an Ubuntu 22.04.2 LTS operating system with an RTX 3090 graphics card with 24GB of VRAM. Initial parameters for training the improved DeepLabV3+ semantic segmentation network model were set; four channels of the training dataset were used as input; the AdamW optimizer was used with an initial learning rate of 0.00001; the number of iterations was 320,000; the cross-entropy loss function was set, and loss weights were configured.

[0035] Both LinearLR and PolyLR learning strategies were employed. LinearLR was used for the first 8000 iterations, and PolyLR was used from the 8000th iteration until the end. During training, images were cropped to 256×256 pixels with a stride of 128. The cropped dataset was divided into training and testing sets in a 6:4 ratio. The model was trained using the cropped dataset divided into training and testing sets in a 6:4 ratio, and validated using the testing set. To better evaluate the proposed model, PSPNet and DeepLabV3+, with ResNet18 as the backbone network without pre-training, were compared and evaluated. Comparison images of some original test images and partial windbreak forest segmentation results obtained using the method disclosed in this invention are attached. Figure 5 As shown in the attached diagram, the local effects of each method are compared. Figure 6 As shown in Table 1, the validation accuracy of each method is as follows. In Table 1, mIoU, mAcc, mFscore, mPrecision, mRecall, and Params are all evaluation metrics. Experiments demonstrate that the model proposed in this invention outperforms PSPNet and DeepLabV3+ in all evaluation metrics, achieving an mIoU of 77.11% with fewer parameters.

[0036] Example 4: As shown in the appendix Figure 7 As shown in the figure, an embodiment of the present invention discloses a device for constructing a semantic segmentation model of windbreak forests in remote sensing images, comprising: The sample acquisition unit acquires a small batch of historical image samples and divides them into a training sample set and a test sample set. Each historical image sample includes a historical remote sensing image of the windbreak forest area and a label that identifies the corresponding semantic segmentation result of the windbreak forest. The training unit uses the training sample set to train the improved DeepLabV3+ network model to obtain the windbreak semantic segmentation model. The improved DeepLabV3+ network model is a DeepLabV3+ network model that improves the backbone network and ASPP model. The iterative unit trains the windbreak semantic segmentation model based on the number of iterations and the loss function to obtain the final windbreak semantic segmentation model; The test unit uses a test sample set to test the windbreak semantic segmentation model, and optimizes the windbreak semantic segmentation model based on the test results.

[0037] Example 5: As shown in the attached document Figure 8 As shown in the figure, an embodiment of the present invention discloses a semantic segmentation device for windbreak forests in remote sensing images, comprising: The image acquisition unit acquires the remote sensing image to be processed and performs a complete cropping of the remote sensing image to be processed, wherein the cropping size is consistent with the size of the historical image samples in the training sample set; The prediction unit inputs all cropped images into the windbreak forest semantic segmentation model to obtain the corresponding segmentation prediction results, wherein the windbreak forest semantic segmentation model is obtained using the method described in Example 1; The stitching unit stitches together all the segmentation prediction results according to the segmentation order of the remote sensing image to be processed, and obtains a complete prediction result image.

[0038] Example 6: This embodiment of the invention discloses a storage medium storing a computer program that can be read by a computer. The computer program is configured to execute a semantic segmentation method for windbreak forests in remote sensing images during runtime.

[0039] The aforementioned storage media may include, but are not limited to, USB flash drives, read-only memory, portable hard drives, magnetic disks, optical disks, and other media capable of storing computer programs.

[0040] Example 7: This embodiment of the invention discloses an electronic device, including a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement a semantic segmentation method for windbreak forests in remote sensing images.

[0041] The processor described above can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. It can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The memory can include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, portable hard drives, magnetic disks, or optical disks.

[0042] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0043] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0044] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0045] The above technical features constitute the preferred embodiment of the present invention, which has strong adaptability and optimal implementation effect. Unnecessary technical features can be added or removed according to actual needs to meet the requirements of different situations.

Claims

1. A method for constructing a semantic segmentation model of windbreak forests in remote sensing images, characterized in that, include: A small batch of historical image samples were obtained and divided into training sample set and test sample set. Each historical image sample included a historical remote sensing image of the windbreak forest area and a label indicating the corresponding semantic segmentation result of the windbreak forest. The improved DeepLabV3+ network model was trained using the training sample set to obtain the following semantic segmentation model for windbreaks: Historical image samples from the training sample set are input into the improved backbone network and the improved ASPP module, respectively, to obtain low-level features and high-level features, specifically: The improved backbone network performs initial feature extraction on the input image through initial convolutions and downsampling using pooling layers to extract initial low-level features. Subsequently, deep feature extraction is performed in four stages to obtain deep features. Each stage contains two residual blocks, which are composed of 3×3 convolution + BN + 5×5 depthwise convolution + BN + linear layer + GELU + linear layer + BN. The residual block in the first stage directly adds the input to the output through skip connections. The residual connections in the first residual block of the second, third, and fourth stages are composed of 1×1 convolution + BN, which directly add the input to the output through skip connections. The second residual block directly adds the input to the output through skip connections. In the first convolution of the second stage, downsampling is performed with a stride of 2. The improved ASPP module includes five branches with different structures, an attention module, and full-dimensional dynamic convolution. The attention module includes average pooling and max pooling. Through the five branches with different structures, multi-scale contextual information is extracted from the extracted deep features to obtain high-level features. The high-level features are multiplied by the channel attention weights of the attention module after full-dimensional dynamic convolution, and finally added to the original deep features and output. The first branch is a 1×1 convolution followed by the attention module. The second to fifth branches are composed of 3×3 depthwise separable convolutions. Then, depthwise separable convolutions with dilation rates of 2, 4, 6, and 8 are added progressively according to each branch. The low-level and high-level features are fused using a decoder. The fused features are then mapped to the number of categories required for the windbreak semantic segmentation task. Finally, an image with the same size as the input historical image sample is obtained by upsampling, thus completing the windbreak semantic segmentation. The segmentation results of the semantic segmentation model for windbreak forests are evaluated by combining the loss function, and the model network parameters are updated by backpropagation and optimization algorithms based on the evaluation results. The semantic segmentation model of windbreak forest is trained by iterating through the number of iterations and the loss function to obtain the final semantic segmentation model of windbreak forest. The semantic segmentation model of windbreak forest was tested using a test sample set, and the model was optimized based on the test results.

2. The method for constructing a semantic segmentation model of windbreak forests in remote sensing images according to claim 1, characterized in that, When training the semantic segmentation model of windbreak forest based on the number of iterations, the model is trained using the LinearLR learning strategy and the PolyLR learning strategy.

3. The method for constructing a semantic segmentation model of windbreak forests in remote sensing images according to claim 1 or 2, characterized in that, The annotation of windbreak forest areas in historical remote sensing images includes using ENVI software to annotate the historical remote sensing images, generating a corresponding label for each pixel, storing the generated label file in the form of an XML file, exporting and converting it into a masked image to obtain historical remote sensing images with annotated windbreak forest areas.

4. A semantic segmentation method for windbreak forests in remote sensing images, characterized in that, include: Obtain the remote sensing image to be processed and perform a complete cropping of the remote sensing image to be processed, wherein the cropping size is consistent with the size of the historical image samples in the training sample set; All cropped images are input into the windbreak forest semantic segmentation model to obtain the corresponding segmentation prediction results, wherein the windbreak forest semantic segmentation model is obtained by the method described in any one of claims 1 to 3; All segmentation prediction results are stitched together according to the segmentation order of the remote sensing image to be processed to obtain a complete prediction result image.

5. A device for constructing a semantic segmentation model of windbreak forests in remote sensing images using the method described in any one of claims 1 to 3, characterized in that, include: The sample acquisition unit acquires a small batch of historical image samples and divides them into a training sample set and a test sample set. Each historical image sample includes a historical remote sensing image of the windbreak forest area and a label that identifies the corresponding semantic segmentation result of the windbreak forest. The training unit uses the training sample set to train the improved DeepLabV3+ network model, resulting in a windbreak semantic segmentation model, including: Historical image samples from the training sample set are input into the improved backbone network and the improved ASPP module, respectively, to obtain low-level features and high-level features, specifically: The improved backbone network performs initial feature extraction on the input image through initial convolutions and downsampling using pooling layers to extract initial low-level features. Subsequently, deep feature extraction is performed in four stages to obtain deep features. Each stage contains two residual blocks, which are composed of 3×3 convolution + BN + 5×5 depthwise convolution + BN + linear layer + GELU + linear layer + BN. The residual block in the first stage directly adds the input to the output through skip connections. The residual connections in the first residual block of the second, third, and fourth stages are composed of 1×1 convolution + BN, which directly add the input to the output through skip connections. The second residual block directly adds the input to the output through skip connections. In the first convolution of the second stage, downsampling is performed with a stride of 2. The improved ASPP module includes five branches with different structures, an attention module, and full-dimensional dynamic convolution. The attention module includes average pooling and max pooling. Through the five branches with different structures, multi-scale contextual information is extracted from the extracted deep features to obtain high-level features. The high-level features are multiplied by the channel attention weights of the attention module after full-dimensional dynamic convolution, and finally added to the original deep features and output. The first branch is a 1×1 convolution followed by the attention module. The second to fifth branches are composed of 3×3 depthwise separable convolutions. Then, depthwise separable convolutions with dilation rates of 2, 4, 6, and 8 are added progressively according to each branch. The low-level and high-level features are fused using a decoder. The fused features are then mapped to the number of categories required for the windbreak semantic segmentation task. Finally, an image with the same size as the input historical image sample is obtained by upsampling, thus completing the windbreak semantic segmentation. The segmentation results of the semantic segmentation model for windbreak forests are evaluated by combining the loss function, and the model network parameters are updated by backpropagation and optimization algorithms based on the evaluation results. The iterative unit trains the windbreak semantic segmentation model based on the number of iterations and the loss function to obtain the final windbreak semantic segmentation model; The test unit uses a test sample set to test the windbreak semantic segmentation model, and optimizes the windbreak semantic segmentation model based on the test results.

6. A semantic segmentation device for windbreak forests in remote sensing images using the method described in claim 4, characterized in that, include: The image acquisition unit acquires the remote sensing image to be processed and performs a complete cropping of the remote sensing image to be processed, wherein the cropping size is consistent with the size of the historical image samples in the training sample set; The prediction unit inputs all cropped images into the windbreak forest semantic segmentation model to obtain the corresponding segmentation prediction results, wherein the windbreak forest semantic segmentation model is obtained by applying the method described in any one of claims 1 to 3; The stitching unit stitches together all the segmentation prediction results according to the segmentation order of the remote sensing image to be processed, and obtains a complete prediction result image.

7. A storage medium, characterized in that, The storage medium stores a computer program that can be read by a computer, and the computer program is configured to execute the semantic segmentation method for windbreak forests in remote sensing images as described in claim 4 when it is run.

8. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the semantic segmentation method for windbreak forests in remote sensing images as described in claim 4.

Images