Agricultural remote sensing image cultivated land extraction method based on boundary guidance
Through the deep learning method of boundary guidance, the features of high-resolution remote sensing images are extracted using texture enhancement and boundary perception modules, and feature fusion is combined with the boundary guidance module, which solves the problem of high-resolution image farmland extraction in the existing technology and achieves more accurate and efficient identification of cultivated land.
Patent Information
- Application Number
- CN202411322915.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-23
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-09-23
AI Technical Summary
When processing high-resolution images, the existing methods of farmland extraction have incomplete characterization due to cropping and downsampling, and the feature resolution is reduced, making it difficult to distinguish visually similar terrain. The mainstream models do not fully consider agricultural remote sensing images. The characteristics of high and low dimensional feature maps in the middle lead to difficulty in identification.
Deep learning method based on boundary guidance is adopted to extract cultivated land through convolutional neural networks, multi-level features are extracted using texture enhancement modules and boundary perception modules, feature fusion is combined with boundary guidance modules, and network parameters are optimized using smooth label cross entropy loss to improve segmentation result accuracy.
It significantly improves the accuracy and generalization of arable land extraction, reduces the difficulty of identifying land texture details, enhances the perception of different land texture information, optimizes the foreground and background feature representation, and improves the accuracy of segmentation results.
Smart Images

Figure CN119027822A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of deep learning and agricultural application technology, and in particular to a method for extracting cultivated land from agricultural remote sensing images based on boundary guidance. Background Art
[0002] Optical remote sensing images are images generated by obtaining information about the earth's surface through remote sensing technology. Optical remote sensing images can be generated through data collected by remote optical sensors (such as satellites, aircraft, drone cameras, etc.). Optical remote sensing images have great advantages in resolution, spectral information, and temporal resolution. They can provide rich surface information and provide sufficient and powerful data support for remote sensing image farmland extraction.
[0003] Remote sensing image segmentation is the pixel-level classification of remote sensing images. It divides the pixel categories of different areas in the image and divides the areas with the same semantics into one category. Cultivated land image extraction is to accurately divide the cultivated land areas in remote sensing images. In order to promote agricultural and rural modernization and achieve the goal of high-yield, high-efficiency and ecologically coordinated agricultural development with the help of scientific and technological innovation, cultivated land image extraction can provide important help for accurately and quickly grasping the utilization of cultivated land resources. In addition, cultivated land image extraction is also of great significance in ensuring national food security and maintaining environmental sustainable development.
[0004] Early methods for extracting cultivated land from remote sensing images mainly include threshold segmentation, region merging, mean shift, watershed algorithm and other methods. They mainly rely on the surface visual features of the image. The features extracted by this type of segmentation method lack semantic information, and the classification process is mainly based on the similarity between pixels rather than deep semantic associations.
[0005] Although the existing methods for extracting cultivated land from remote sensing images have been well developed, due to the characteristics of remote sensing images themselves, cultivated land extraction still has the following shortcomings: (1) High-resolution imaging technology brings rich detailed information, but the diversity of object categories and the complexity of topography significantly increase the difficulty of extracting cultivated land in agricultural remote sensing images. (2) The inevitable cropping and downsampling operations when processing high-resolution remote sensing images will lead to incomplete representation of objects and loss of contextual information. The reduction in feature map resolution leads to the loss of details and boundary information, making it difficult to identify pixels at the junction of different object categories. (3) The current mainstream encoding and decoding network model has not fully considered the characteristics of high- and low-dimensional feature maps of objects in agricultural remote sensing images, and lacks an accurate and efficient optimization strategy to fuse feature representations at different levels, which will make visually similar objects difficult to distinguish. Summary of the invention
[0006] In order to make up for the above shortcomings, the present invention provides a method for extracting cultivated land from agricultural remote sensing images based on boundary guidance, aiming to gradually replace traditional segmentation methods with deep learning methods and show good segmentation effects. The remote sensing image cultivated land extraction model based on deep learning is trained through a large number of data sets, and convolutional neural networks are used to predict segmentation results, thereby improving the accuracy and generalization of cultivated land extraction.
[0007] To achieve the above object, the present invention provides the following technical solution: a method for extracting cultivated land from agricultural remote sensing images based on boundary guidance, comprising the following steps: A) Dataset preprocessing: crop and preprocess the high-resolution optical remote sensing farmland images to form a preprocessed dataset, and divide the dataset into training set, validation set, and test set; B) Data enhancement: Perform general data enhancement or standardization on the dataset to obtain a remote sensing cultivated land image dataset after data enhancement; C) Feature extraction: The training set images of the dataset are input into the feature extraction module of the deep convolutional neural network in batches to extract multi-level features; D) Texture feature enhancement: The texture enhancement module TEM is used to enhance the texture representation of low-level features extracted by the deep convolutional neural network; E) Boundary feature enhancement: The boundary perception module BAM is used to mine the boundary representation of the object from the low-level features containing texture details and the high-level features containing abstract semantics, and a boundary loss function based on geometric similarity is used to supervise the boundary extraction process; F) Boundary-guided feature fusion: Multiple boundary-guided modules (BGMs) are used to combine the boundary details obtained from BAM with the multi-level backbone features of each layer to guide the decoder feature learning; G) Segmentation prediction: Finally, the decoder output feature map is used for segmentation prediction to obtain the final farmland image segmentation result; H) Network model training: The segmentation results predicted by the farmland segmentation network are compared with the actual farmland segmentation labels through smoothed label cross entropy loss calculation, and the network parameters are updated through back propagation to improve the accuracy of the farmland segmentation network prediction results.
[0008] Preferably, in the step D), the texture enhancement module TEM independently adjusts the weight of each group of features after grouping the features, and assigns different attention weights to different spatial positions, thereby improving the ability of feature expression and reducing noise interference.
[0009] Preferably, the step D) comprises the following steps: D1) for a height of H , width is W , the number of channels is CThe convolution feature map is divided into K For each spatial position in each feature map, it is represented as a one-dimensional vector, denoted as ; Specifically, by applying the global spatial average function To approximate the semantic vector of the global features of this group learning representation : (1) Using global semantic features TEM generates a corresponding importance coefficient for each local feature and evaluates it with the dot product operation The similarity between , then, through spatial normalization, the differences between different samples and the influence of noise are suppressed: (2) (3) in, and Represents the similarity coefficient The mean and variance of represents a constant for numerical stability, to ensure that the normalization inserted into the network can represent the identity transformation, for each coefficient A pair of parameters are introduced and , these parameters scale and shift the normalized values: (4) Finally, to obtain the enhanced texture feature vector , the original Scaling is performed and a residual structure is introduced to ensure stable back propagation of the gradient: (5) All enhanced features processed by the texture enhancement module form the final feature set By calculating the texture similarity of different spatial positions in the feature map and assigning higher weights to spatial positions with similar texture characteristics, the texture enhancement module significantly improves the network's ability to perceive the texture information of different objects in remote sensing images.
[0010] Preferably, in step E), the boundary perception module BAM combines the low-level features generated by the encoder through the texture enhancement module TEM Texture details and advanced features The semantic information in the object is used to model the boundary information related to the object, thus forming a richer and more accurate boundary representation.
[0011] Preferably, the step E) comprises the following steps: E1) First in the horizontal direction and vertical direction The Sobel operator is applied in the direction to obtain the gradient map. Specifically, first initialize two The parameters are fixed and a convolution operation with a step size of 1 is applied. The parameters of the two convolution kernels are initialized as and : (6) E2) After that, these two convolutions are applied to the input feature maps respectively To get the gradient map , , : (7) (8) (9) in, Indicates the use of a specific convolution kernel Features Perform convolution operation, gradient map and The element sizes represent the features In the horizontal direction and vertical direction The boundary strength, Represents the total boundary strength regardless of direction, which represents the horizontal gradient and vertical gradient The vector sum, then the gradient map Depend on Function normalization and fusion with the input feature map to obtain a feature map with boundary enhancement : (10) in, represents element-wise multiplication, express function; E3) After that, and The boundary enhancement feature map is fused with a simple stacked convolutional layer. Specifically, the feature map is first Apply a bilinear upsampling operation and Convolution to obtain Feature maps of the same size , then apply Convolution operation is used to adjust the channel size of the two features. Finally, the two feature maps are connected along the channel dimension, and multiple layers of stacked convolution layers are applied to obtain the final boundary enhancement feature and pass Function Get : (11) (12) (13) in, represents the upsampling operation, express convolution, represents the feature concatenation operation, express function, Indicates that they are passed in order. Convolution, batch normalization, ReLU activation function and convolution.
[0012] Preferably, in step F), the boundary guidance module BGM receives two inputs: one is the boundary enhancement feature and the corresponding characteristics of the encoder network The mixed features obtained by multiplication; the second is the features from the previous decoder layer ,These two inputs are fed into the foreground path and background path of BGM, aiming to promote the feature representation of foreground and background, respectively.
[0013] Preferably, the step F) comprises the following steps: F1) In the foreground path, by and Directly connect along the channel dimension, and then pass through Conv-BN-ReLU in sequence to obtain foreground features : (14) in, represents element-wise multiplication, It means that Conv-BN-ReLU is passed through in sequence. F2) For the background path, the present invention designs a background attention component to allow the model to selectively focus on background information. The background features obtained by the background attention component It is expressed as: (15) in, Represents a three-layer stacked Conv-BN-ReLU, represents the Sigmoid function, represents element-wise multiplication, represents the background attention map, which is generated by applying the sigmoid function to the feature map of the previous decoder layer and subtracting the resulting foreground attention map from 1; F3) Finally, the foreground features , Background characteristics and the previous decoder feature Concatenate them in the channel dimension to form the final decoder output : (16).
[0014] A system for extracting cultivated land from agricultural remote sensing images based on boundary guidance, using the method for extracting cultivated land from agricultural remote sensing images based on boundary guidance, comprising a preprocessing and data enhancement module, a texture enhancement module TEM, a boundary perception module BAM and a boundary guidance module BGM; The preprocessing and data enhancement module is used to perform image cropping and data enhancement on the input high-resolution optical remote sensing cultivated land segmentation image to form a preprocessed image data set; The texture enhancement module TEM is used to independently adjust the weight of each group of features after grouping the features, and assign different attention weights to different spatial positions, thereby improving the ability of feature expression and reducing noise interference; The boundary perception module BAM combines the low-level features generated by the encoder through the texture enhancement module TEM Texture details and advanced features The semantic information in the object is used to model the boundary information related to the object, thus forming a richer and more accurate boundary representation; The boundary guidance module BGM is used to extract boundary priors, improve the feature representation capability generated by the decoder, adopt a dual-path boundary fusion strategy, optimize the feature representation of the foreground and background, obtain the final decoder output feature map, perform segmentation prediction on it, and output the predicted remote sensing cultivated land segmentation result to the user.
[0015] Preferably, the boundary-guided deep convolutional neural network adopted by the system is an encoder-decoder architecture, the encoder includes a feature extraction module, a texture enhancement module, and a boundary perception module, and the decoder includes a boundary guidance module and a segmentation prediction module.
[0016] The present invention has the following beneficial effects: 1. In the present invention, a texture enhancement module TEM is proposed. After grouping the features, the weight of each group of features is independently adjusted, and different attention weights are assigned to different spatial positions, thereby improving the ability of feature expression and reducing noise interference, significantly reducing the difficulty of farmland extraction in agricultural remote sensing images. This module is particularly critical for enhancing the representation of texture features, especially in agricultural remote sensing image scenes that pay great attention to the texture details of the ground objects.
[0017] 2. In the present invention, the boundary perception ability of the texture feature and semantic feature enhancement model is further utilized, and a boundary perception module BAM is proposed. This module combines the texture details of the low-level features generated by the encoder through the texture enhancement module TEM and the semantic information in the high-level features to model the boundary information related to the object, thereby forming a richer and more accurate boundary representation, reducing the difficulty of identifying pixel points at the junction of different land object categories.
[0018] 3. In the present invention, the boundary prior extracted by the boundary perception module is used to improve the feature representation capability generated by the decoder.
[0019] 4. The present invention proposes a boundary guidance module named BGM, which adopts a dual-path boundary fusion strategy to optimize the feature representation of the foreground and background. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 A network architecture diagram of the method for extracting cultivated land from agricultural remote sensing images based on boundary guidance in the present invention; Figure 2 is a schematic diagram of a texture enhancement module TEM in the present invention; Figure 3 Schematic diagram of the boundary perception module BAM in the present invention; Figure 4 Schematic diagram of the boundary guidance module BGM in the present invention; Figure 5 It is a segmentation effect diagram and segmentation label diagram of the remote sensing image cultivated land extraction network on the Huiwen Town data set of the present invention; Figure 6 It is a segmentation effect diagram and a segmentation label diagram of the remote sensing image cultivated land extraction network of the present invention on the Puqian Town data set; Figure 7 It is the segmentation effect diagram and segmentation label diagram of the remote sensing image farmland extraction network on the IFLYTEK dataset of the present invention. DETAILED DESCRIPTION
[0021] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0022] Reference Figure 1-Figure 7 The present invention provides an embodiment: a method for extracting cultivated land from agricultural remote sensing images based on boundary guidance, comprising the following steps: A) Dataset preprocessing: crop and preprocess the high-resolution optical remote sensing farmland images to form a preprocessed dataset, and divide the dataset into training set, validation set, and test set; B) Data enhancement: Perform general data enhancement or standardization on the dataset to obtain a remote sensing cultivated land image dataset after data enhancement; C) Feature extraction: The training set images of the dataset are input into the feature extraction module of the deep convolutional neural network in batches to extract multi-level features; D) Texture feature enhancement: The texture enhancement module TEM is used to enhance the texture representation of low-level features extracted by the deep convolutional neural network; E) Boundary feature enhancement: The boundary perception module BAM is used to mine the boundary representation of the object from the low-level features containing texture details and the high-level features containing abstract semantics, and a boundary loss function based on geometric similarity is used to supervise the boundary extraction process; F) Boundary-guided feature fusion: Multiple boundary-guided modules (BGMs) are used to combine the boundary details obtained from BAM with the multi-level backbone features of each layer to guide the decoder feature learning; G) Segmentation prediction: Finally, the decoder output feature map is used for segmentation prediction to obtain the final farmland image segmentation result; H) Network model training: The segmentation results predicted by the farmland segmentation network are compared with the actual farmland segmentation labels through smoothed label cross entropy loss calculation, and the network parameters are updated through back propagation to improve the accuracy of the farmland segmentation network prediction results.
[0023] Among them, reference Figure 2 ,In step D), the texture enhancement module TEM independently adjusts the weight of each group of features after grouping the features, and assigns different attention weights to different spatial positions, thereby improving the ability of feature expression and reducing noise interference.
[0024] Step D) comprises the following steps: D1) for a height of H , width is W , the number of channels is CThe convolution feature map is divided into K For each spatial position in each feature map, it is represented as a one-dimensional vector, denoted as ; Specifically, by applying the global spatial average function To approximate the semantic vector of the global features of this group learning representation : (1) Using global semantic features TEM generates a corresponding importance coefficient for each local feature and evaluates it with the dot product operation The similarity between , then, through spatial normalization, the differences between different samples and the influence of noise are suppressed: (2) (3) in, and Represents the similarity coefficient The mean and variance of represents a constant for numerical stability, to ensure that the normalization inserted into the network can represent the identity transformation, for each coefficient A pair of parameters are introduced and , these parameters scale and shift the normalized values: (4) Finally, to obtain the enhanced texture feature vector , the original Scaling is performed and a residual structure is introduced to ensure stable back propagation of the gradient: (5) All enhanced features processed by the texture enhancement module form the final feature set By calculating the texture similarity of different spatial positions in the feature map and assigning higher weights to spatial positions with similar texture characteristics, the texture enhancement module significantly improves the network's ability to perceive the texture information of different objects in remote sensing images.
[0025] refer to Figure 3 In step E), the boundary perception module BAM combines the low-level features generated by the encoder through the texture enhancement module TEM Texture details and advanced features The semantic information in the object is used to model the boundary information related to the object, thus forming a richer and more accurate boundary representation.
[0026] Step E) comprises the following steps: E1) First in the horizontal direction and vertical direction The Sobel operator is applied in the direction to obtain the gradient map. Specifically, first initialize two The parameters are fixed and a convolution operation with a step size of 1 is applied. The parameters of the two convolution kernels are initialized as and : (6) E2) Apply these two convolutions to the input feature maps separately To get the gradient map , , : (7) (8) (9) in, Indicates the use of a specific convolution kernel Features Perform convolution operation, gradient map and The element sizes represent the features In the horizontal direction and vertical direction The boundary strength, Represents the total boundary strength regardless of direction, which represents the horizontal gradient and vertical gradient The vector sum, then the gradient map Depend on Function normalization and fusion with the input feature map to obtain a feature map with boundary enhancement : (10) in, represents element-wise multiplication, express function; E3) After that, and The boundary enhancement feature map is fused with a simple stacked convolutional layer. Specifically, the feature map is first Apply a bilinear upsampling operation and Convolution to obtain Feature maps of the same size , then apply Convolution operation is used to adjust the channel size of the two features. Finally, the two feature maps are connected along the channel dimension, and multiple layers of stacked convolution layers are applied to obtain the final boundary enhancement feature and pass Function Get : (11) (12) (13) in, represents the upsampling operation, express convolution, represents the feature concatenation operation, express function, Indicates that they are passed in order. Convolution, batch normalization, ReLU activation function and convolution.
[0027] refer to Figure 4 In step F), the boundary guidance module BGM accepts two inputs: one is the boundary enhancement feature and the corresponding characteristics of the encoder network The mixed features obtained by multiplication; the second is the features from the previous decoder layer ,These two inputs are fed into the foreground path and background path of BGM, aiming to promote the feature representation of foreground and background, respectively.
[0028] Preferably, step F) comprises the following steps: F1) In the foreground path, by and Directly connect along the channel dimension, and then pass through Conv-BN-ReLU in sequence to obtain foreground features : (14) in, represents element-wise multiplication, It means that Conv-BN-ReLU is passed through in sequence. F2) For the background path, the present invention designs a background attention component to allow the model to selectively focus on background information. The background features obtained by the background attention component It is expressed as: (15) in, Represents a three-layer stacked Conv-BN-ReLU, represents the Sigmoid function, represents element-wise multiplication, represents the background attention map, which is generated by applying the sigmoid function to the feature map of the previous decoder layer and subtracting the resulting foreground attention map from 1; F3) Finally, the foreground features , Background characteristics and the previous decoder feature Concatenate them in the channel dimension to form the final decoder output : (16).
[0029] A system for extracting cultivated land from agricultural remote sensing images based on boundary guidance, using the above-mentioned method for extracting cultivated land from agricultural remote sensing images based on boundary guidance, comprising a preprocessing and data enhancement module, a texture enhancement module TEM, a boundary perception module BAM and a boundary guidance module BGM; The preprocessing and data enhancement module is used to perform image cropping and data enhancement on the input high-resolution optical remote sensing cultivated land segmentation image to form a preprocessed image dataset; The texture enhancement module TEM is used to group features and independently adjust the weight of each group of features, and assign different attention weights to different spatial positions, thereby improving the ability of feature expression and reducing noise interference; Boundary Awareness Module BAM The boundary awareness module BAM combines the low-level features generated by the encoder through the texture enhancement module TEM Texture details and advanced features The semantic information in the object is used to model the boundary information related to the object, thus forming a richer and more accurate boundary representation; The boundary guidance module (BGM) is used to extract boundary priors, improve the feature representation capability of the decoder, and adopt a dual-path boundary fusion strategy to optimize the feature representation of the foreground and background, obtain the final decoder output feature map, perform segmentation prediction on it, and output the predicted remote sensing cultivated land segmentation results to the user.
[0030] The boundary-guided deep convolutional neural network is an encoder-decoder architecture. The encoder includes a feature extraction module, a texture enhancement module, and a boundary perception module, and the decoder includes a boundary guidance module and a segmentation prediction module.
[0031] In this embodiment, the deep convolutional neural network is a ResNet50 structure.
[0032] refer to Figure 5-Figure 7,The segmentation accuracy of this embodiment is compared with several state-of-the-art methods, including U-Net, PSPNet, DANet, DeepLabV3+, ABCNet, HRCNet, and MANet.,The remote sensing cultivated land segmentation datasets used in this experiment are Huiwen town dataset, Puqian town and IFLYTEK dataset.,The comparative experiment uses the model methods U-Net, PSPNet, DANet, DeepLabV3+, ABCNet, HRCNet, and MANet methods as per the original text.
[0033] After experiments, Table 1 shows the comparative experimental results of the present invention and other methods for remote sensing image farmland segmentation on the Huiwen Town dataset, where OA is the overall accuracy, F1-Score is the F1 score, and mIoU is the mean intersection over union ratio indicator. The results are shown in the following table: Table 1
[0034] After experiments, Table 2 shows the comparative experimental results of the present invention and other methods for remote sensing image farmland segmentation on the Puqian Town dataset, where OA is the overall accuracy, F1-Score is the F1 score, and mIoU is the mean intersection over union ratio indicator. The results are shown in the following table: Table 2
[0035] After experiments, Table 3 shows the comparative experimental results of the present invention and other methods for remote sensing image farmland segmentation on the IFLYTEK dataset, where OA is the overall accuracy, F1-Score is the F1 score, and mIoU is the mean intersection over union ratio indicator. The results are shown in the following table: Table 3
[0036] According to Table 1, Table 2 and Table 3, the boundary-guided cultivated land extraction method for agricultural remote sensing images proposed in the present invention has higher accuracy than other methods on the three data sets.
[0037] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for extracting cultivated land from agricultural remote sensing images based on boundary guidance, characterized in that: The following steps are involved: A) Dataset preprocessing: crop and preprocess the high-resolution optical remote sensing farmland images to form a preprocessed dataset, and divide the dataset into training set, validation set, and test set; B) Data enhancement: Perform general data enhancement or standardization on the dataset to obtain a remote sensing cultivated land image dataset after data enhancement; C) Feature extraction: The training set images of the dataset are input into the feature extraction module of the deep convolutional neural network in batches to extract multi-level features; D) Texture feature enhancement: The texture enhancement module TEM is used to enhance the texture representation of low-level features extracted by the deep convolutional neural network; E) Boundary feature enhancement: The boundary perception module BAM is used to mine the boundary representation of the object from the low-level features containing texture details and the high-level features containing abstract semantics, and a boundary loss function based on geometric similarity is used to supervise the boundary extraction process; F) Boundary-guided feature fusion: Multiple boundary-guided modules (BGMs) are used to combine the boundary details obtained from BAM with the multi-level backbone features of each layer to guide the decoder feature learning; G) Segmentation prediction: Finally, the decoder output feature map is used for segmentation prediction to obtain the final farmland image segmentation result; H) Network model training: The segmentation results predicted by the farmland segmentation network are compared with the actual farmland segmentation labels through smoothed label cross entropy loss calculation, and the network parameters are updated through back propagation to improve the accuracy of the farmland segmentation network prediction results.
2. The method for extracting cultivated land from agricultural remote sensing images based on boundary guidance according to claim 1, characterized in that: In the step D), the texture enhancement module TEM groups the features and independently adjusts the weight of each group of features, and assigns different attention weights to different spatial positions, thereby improving the ability of feature expression and reducing noise interference.
3. The method for extracting cultivated land from agricultural remote sensing images based on boundary guidance according to claim 2, characterized in that: The step D) comprises the following steps: D1) for a height of H , width is W , the number of channels is C The convolution feature map is divided into K For each spatial position in each feature map, it is represented as a one-dimensional vector, denoted as ; Specifically, by applying the global spatial average function To approximate the semantic vector of the global features of this group learning representation : (1) Using global semantic features TEM generates a corresponding importance coefficient for each local feature and evaluates it with the dot product operation The similarity between , then, through spatial normalization, the differences between different samples and the influence of noise are suppressed: (2) (3) in, and Represents the similarity coefficient The mean and variance of represents a constant for numerical stability, to ensure that the normalization inserted into the network can represent the identity transformation, for each coefficient A pair of parameters are introduced and , these parameters scale and shift the normalized values: (4) Finally, to obtain the enhanced texture feature vector , the original Scaling is performed and a residual structure is introduced to ensure stable back propagation of the gradient: (5) All enhanced features processed by the texture enhancement module form the final feature set By calculating the texture similarity of different spatial positions in the feature map and assigning higher weights to spatial positions with similar texture characteristics, the texture enhancement module significantly improves the network's ability to perceive the texture information of different objects in remote sensing images.
4. The method for extracting cultivated land from agricultural remote sensing images based on boundary guidance according to claim 1, characterized in that: In step E), the boundary perception module BAM combines the low-level features generated by the encoder through the texture enhancement module TEM Texture details and advanced features The semantic information in the object is used to model the boundary information related to the object, thus forming a richer and more accurate boundary representation.
5. The method for extracting cultivated land from agricultural remote sensing images based on boundary guidance according to claim 4 is characterized in that: The step E) comprises the following steps: E1) First in the horizontal direction and vertical direction The Sobel operator is applied in the direction to obtain the gradient map. Specifically, first initialize two The parameters are fixed and a convolution operation with a step size of 1 is applied. The parameters of the two convolution kernels are initialized as and : (6) E2) After that, these two convolutions are applied to the input feature maps respectively To get the gradient map , , : (7) (8) (9) in, Indicates using a specific convolution kernel Features Perform convolution operation, gradient map and The element sizes represent the features In the horizontal direction and vertical direction The boundary strength, Represents the total boundary strength regardless of direction, which represents the horizontal gradient and vertical gradient The vector sum, then the gradient map Depend on Function normalization and fusion with the input feature map to obtain a feature map with boundary enhancement : (10) in, represents element-wise multiplication, express function; E3) After that, and The boundary enhancement feature map is fused with a simple stacked convolutional layer. Specifically, the feature map is first Apply a bilinear upsampling operation and Convolution to obtain Feature maps of the same size , then apply Convolution operation is used to adjust the channel size of the two features. Finally, the two feature maps are connected along the channel dimension, and multiple layers of stacked convolution layers are applied to obtain the final boundary enhancement feature and pass Function Get : (11) (12) (13) in, represents the upsampling operation, express convolution, represents the feature concatenation operation, express function, Indicates that they are passed in order. Convolution, batch normalization, ReLU activation function and convolution.
6. The method for extracting cultivated land from agricultural remote sensing images based on boundary guidance according to claim 1, characterized in that: In step F), the boundary guidance module BGM receives two inputs: one is the boundary enhancement feature and the corresponding characteristics of the encoder network The mixed features obtained by multiplication; The second is the features from the previous decoder layer ,These two inputs are fed into the foreground path and background path of BGM, aiming to promote the feature representation of foreground and background, respectively.
7. The method for extracting cultivated land from agricultural remote sensing images based on boundary guidance according to claim 6, characterized in that: The step F) comprises the following steps: F1) In the foreground path, by and Directly connect along the channel dimension, and then pass through Conv-BN-ReLU in sequence to obtain foreground features : (14) in, represents element-wise multiplication, It means that Conv-BN-ReLU is passed through in sequence. F2) For the background path, the present invention designs a background attention component to allow the model to selectively focus on background information. The background features obtained by the background attention component It is expressed as: (15) in, Represents a three-layer stacked Conv-BN-ReLU, represents the Sigmoid function, represents element-wise multiplication, represents the background attention map, which is generated by applying the sigmoid function to the feature map of the previous decoder layer and subtracting the resulting foreground attention map from 1; F3) Finally, the foreground features , Background characteristics and the previous decoder feature Concatenate them in the channel dimension to form the final decoder output : (16)。 8. A boundary-guided agricultural remote sensing image farmland extraction system, characterized in that: The method for extracting cultivated land from agricultural remote sensing images based on boundary guidance according to any one of claims 1 to 7 comprises a preprocessing and data enhancement module, a texture enhancement module TEM, a boundary perception module BAM and a boundary guidance module BGM; The preprocessing and data enhancement module is used to perform image cropping and data enhancement on the input high-resolution optical remote sensing cultivated land segmentation image to form a preprocessed image data set; The texture enhancement module TEM is used to independently adjust the weight of each group of features after grouping the features, and assign different attention weights to different spatial positions, thereby improving the ability of feature expression and reducing noise interference; The boundary perception module BAM combines the low-level features generated by the encoder through the texture enhancement module TEM Texture details and advanced features The semantic information in the object is used to model the boundary information related to the object, thus forming a richer and more accurate boundary representation; The boundary guidance module BGM is used to extract boundary priors, improve the feature representation capability generated by the decoder, adopt a dual-path boundary fusion strategy, optimize the feature representation of the foreground and background, obtain the final decoder output feature map, perform segmentation prediction on it, and output the predicted remote sensing cultivated land segmentation result to the user.
9. The agricultural remote sensing image farmland extraction system based on boundary guidance according to claim 8 is characterized in that: The system adopts a boundary-guided deep convolutional neural network as an encoder-decoder architecture, wherein the encoder includes a feature extraction module, a texture enhancement module, and a boundary perception module, and the decoder includes a boundary guidance module and a segmentation prediction module.
Citation Information
Patent Citations
Sea-land segmentation method based on context attention and boundary perception guidance
CN117635628A
Intelligent background selection and image segmentation
US20160092746A1
Cited By
Deep learning-based enteromorpha remote sensing image detection method and system
CN120510528A
