Boundary-guided farmland extraction method for agricultural remote sensing images

Abstract

The application relates to the technical field of deep learning and agricultural application, and discloses a farmland extraction method for agricultural remote sensing images based on boundary guidance, which comprises the following steps: extracting multi-level features from an input image through an encoder network, that is, using a texture enhancement module TEM to enhance the texture representation of low-level features, then using a boundary awareness module BAM to mine boundary representations in objects from low-level features containing texture details and high-level features containing abstract semantics, and then using multiple boundary guidance modules BGM to combine boundary details obtained from the BAM with multi-level backbone features of each layer (Fk), output a decoder feature map, and perform farmland image segmentation prediction. Through reverse gradient propagation of a boundary loss function and a segmentation loss function, network model parameters are trained and updated, farmland segmentation accuracy is improved, and the training accuracy of the network model in remote sensing farmland image segmentation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and agricultural application, and particularly relates to an agricultural remote sensing image cultivated land extraction method based on boundary guidance. BACKGROUND

[0002] Optical remote sensing image is an image generated by the information of the earth's surface obtained by remote sensing technology. The optical remote sensing image can be generated by the data collected by a remote optical sensor (such as a satellite, an airplane, a camera of a drone, etc.). The optical remote sensing image has great advantages in resolution, spectral information, time resolution, etc., and can provide rich ground information, providing sufficient and powerful data support for remote sensing image cultivated land extraction.

[0003] Remote sensing image segmentation is a pixel-level classification of a remote sensing image, and different regions in the image are classified by pixels, and regions with the same semantics are classified into one class. Cultivated land image extraction is to accurately divide the cultivated land region in the remote sensing image. In order to promote the modernization of agriculture and rural areas, and realize the goal of high yield, high efficiency and ecological coordination of development of agriculture by means of scientific and technological innovation, cultivated land image extraction can accurately and quickly grasp the utilization of cultivated land resources and provide important help. In addition, cultivated land image extraction is also of great significance in ensuring national food security and maintaining environmental sustainable development.

[0004] Early remote sensing image cultivated land extraction methods mainly include threshold segmentation method, region merging method, mean shift method, watershed algorithm and other methods, which mainly rely on the surface visual features of the image. The features extracted by this kind of segmentation method lack semantic level information, and the classification process is mainly based on the similarity between pixels, rather than deep semantic association.

[0005] Although the existing remote sensing image cultivated land extraction method has good development, due to the characteristics of the remote sensing image itself, the cultivated land extraction still has the following deficiencies: (1) High-resolution imaging technology brings rich detailed information, and the diversity of ground object categories and the complexity of topography significantly increase the difficulty of cultivated land extraction 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 ground objects and loss of context information. The reduction of feature map resolution leads to the loss of detail and boundary information, making it difficult to identify pixel points at the junction of different ground object categories. (3) The current mainstream coding and decoding network model has not fully considered the characteristics of high-dimensional and low-dimensional feature maps of ground objects in agricultural remote sensing images, and lacks an accurate and efficient optimization strategy to fuse feature representations at different levels, which will make it difficult to distinguish visually similar ground objects. SUMMARY

[0006] In order to make up for the above shortcomings, the application provides a boundary guidance-based agricultural remote sensing image cultivated land extraction method, aiming to gradually replace the traditional segmentation method with the deep learning method, and showing good segmentation effect, the remote sensing image cultivated land extraction model based on deep learning carries out model training through a large amount of data sets, carries out segmentation result prediction by using a convolutional neural network, and the accuracy and generalization of cultivated land extraction are improved.

[0007] In order to achieve the above object, the application provides the following technical scheme: a boundary guidance-based agricultural remote sensing image cultivated land extraction method, comprising the following steps:

[0008] A) data set preprocessing: the high-resolution optical remote sensing cultivated land image is cropped and preprocessed to form a preprocessed data set, and the data set is divided into a training set, a validation set and a test set;

[0009] B) data enhancement: the data set is subjected to general data enhancement or standardization processing to obtain a data-enhanced remote sensing cultivated land image data set;

[0010] C) feature extraction: the data set image is input into a feature extraction module of a deep convolutional neural network in batches to extract multi-level features;

[0011] D) texture feature enhancement: a texture enhancement module TEM is used to enhance the texture representation of the low-level features extracted by the deep convolutional neural network;

[0012] E) boundary feature enhancement: a boundary awareness module BAM is used to mine the boundary representation in 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 adopted to supervise the extraction process of the boundary;

[0013] F) boundary guidance feature fusion: a plurality of boundary guidance modules BGM are used to combine the boundary details obtained from the BAM with the multi-level backbone features of each layer to guide the feature learning of the decoder;

[0014] G) segmentation prediction: finally, the decoder output feature map is subjected to segmentation prediction to obtain the final cultivated land image segmentation result;

[0015] H) network model training: the segmentation result predicted by the cultivated land segmentation network is subjected to smooth label cross-entropy loss calculation with the real cultivated land segmentation label, the network parameters are updated through back propagation, and the accuracy of the cultivated land segmentation network prediction result is improved.

[0016] Preferably, in the step D), the texture enhancement module TEM adjusts the weight of each group of features independently after grouping the features, and gives different spatial positions different attention weights, so as to improve the feature expression ability and reduce noise interference.

[0017] Preferably, the step D) comprises the following steps:

[0018] D1) for a convolutional feature map with height H , width W , and channel number C , divide it into K groups along the channel dimension, and for each spatial position in each group of feature maps, represent it as a one-dimensional vector, denoted as ;

[0019] Specifically, by applying a global spatial average function to approximate the semantic vector of global feature of this group learning representation :

[0020] (1)

[0021] Using the semantic vector of global feature , TEM generates a corresponding importance coefficient for each local feature, and evaluates the similarity between and by dot product operation, and then suppresses the difference between different samples and the influence of noise by spatial normalization:

[0022] (2)

[0023] (3)

[0024] where and represent the mean and variance of the similarity coefficient , respectively, denotes a constant for numerical stability, and in order to ensure that the normalization inserted into the network can represent the identity transformation, a pair of parameters and are introduced for each coefficient , which scale and move the normalization value to obtain the offset value :

[0025] (4)

[0026] Finally, in order to obtain the enhanced texture feature vector , the original is scaled by the sigmoid function in the whole vector group space, and the residual structure is introduced to ensure the stable back propagation of the gradient:

[0027] (5)

[0028] All the enhanced features processed by the texture enhancement module form the final feature set. By calculating the texture similarity of different spatial locations in the feature map and assigning higher weights to spatial locations with similar texture characteristics, the texture enhancement module significantly improves the network's ability to perceive the texture information of different land features in remote sensing images.

[0029] Preferably, in step E), 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 model is used to model the boundary information related to the object, thereby forming a richer and more accurate boundary representation.

[0030] Preferably, step E) includes the following steps:

[0031] E1) First in the horizontal direction and vertical direction The Sobel operator is applied in the direction to obtain the gradient map. Specifically, two are first initialized. With parameters fixed and a convolution operation with a stride of 1 applied, the parameters of the two convolution kernels are initialized as follows: and :

[0032] (6)

[0033] After E2), these two convolutions are applied to the input feature maps respectively. Obtain gradient map :

[0034] (7)

[0035] (8)

[0036] (9)

[0037] in, and 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 Boundary strength, gradient map Depend on Function normalization and fusion with the input feature map to obtain a boundary-enhanced feature map :

[0038] (10)

[0039] in, This represents element-wise multiplication. express function;

[0040] After E3, and The boundary enhancement feature map is fused with a simple stacked convolutional layer. Specifically, the feature map is first... Applying bilinear upsampling operation and Convolution to obtain Feature maps of the same size Then, apply respectively Convolutional operations are used to adjust the channel sizes of these two features. Finally, the two feature maps are concatenated along the channel dimension, and multiple stacked convolutional layers are applied to obtain the final boundary enhancement features. The function obtains :

[0041] (11)

[0042] (12)

[0043] (13)

[0044] in, Indicates an upsampling operation. express convolution, This indicates a feature concatenation operation. express function, It means to pass through in order. Convolution, batch normalization, ReLU activation function and convolution.

[0045] Preferably, in step F), the boundary guidance module BGM accepts two inputs: one is boundary enhancement features. and corresponding features of the encoder network The first is the mixed feature obtained by multiplication; the second is the feature from the previous decoder layer. These two inputs are fed into the foreground and background paths of the BGM, respectively, to facilitate feature representations of the foreground and background.

[0046] Preferably, step F) includes the following steps:

[0047] F1) In the foreground path, by and The foreground features are obtained by directly connecting along the channel dimension and then sequentially passing them through Conv-BN-ReLU. :

[0048] (14)

[0049] in, This represents element-wise multiplication. This indicates that the sequence passes through Conv-BN-ReLU in order.

[0050] F2) For background paths, a background attention component is used to allow the model to selectively focus on background information. The background attention component obtains background features. Represented as:

[0051] (15)

[0052] in, This represents a three-layer stacked Conv-BN-ReLU. This represents the Sigmoid function. This represents element-wise multiplication. The background attention map is represented by applying the sigmoid function to the feature map of the previous decoder layer and subtracting the resulting foreground attention map from 1. To generate;

[0053] F3) Finally, foreground features Background features and the features of the previous decoder These are concatenated along the channel dimension to form the final decoder output. :

[0054] (16).

[0055] A boundary-guided agricultural remote sensing image farmland extraction system, using the aforementioned boundary-guided agricultural remote sensing image farmland extraction method, includes a preprocessing and data augmentation module, a texture enhancement module (TEM), a boundary awareness module (BAM), and a boundary guidance module (BGM).

[0056] The preprocessing and data augmentation module is used to perform image cropping and data augmentation on the input high-resolution optical remote sensing farmland segmentation image to form a preprocessed image dataset;

[0057] The texture enhancement module TEM is used to independently adjust the weight of each group of features after grouping them, and to assign different attention weights to different spatial locations, thereby improving the ability to express features and reducing noise interference.

[0058] The boundary awareness module BAM combines low-level features generated by the encoder through the texture enhancement module TEM. Texture details and advanced features The semantic information in the model is used to model the boundary information related to the object, thereby forming a richer and more accurate boundary representation;

[0059] 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, thereby obtaining the final output feature map of the decoder. This map is then used for segmentation prediction, and the predicted remote sensing farmland segmentation results are output to the user.

[0060] Preferably, the system employs a boundary-guided deep convolutional neural network with an encoder-decoder architecture. The encoder includes a feature extraction module, a texture enhancement module, and a boundary awareness module, while the decoder includes a boundary guidance module and a segmentation prediction module.

[0061] The present invention has the following beneficial effects:

[0062] 1. In this invention, a texture enhancement module TEM is proposed. By grouping features and independently adjusting the weight of each group of features, and assigning different attention weights to different spatial locations, the feature representation ability is improved and noise interference is reduced. This significantly reduces the difficulty of extracting farmland in agricultural remote sensing images. This module is particularly important for enhancing the representation of texture features, especially in agricultural remote sensing image scenarios where attention is paid to the details of ground texture.

[0063] 2. In this invention, texture features and semantic features are further used to enhance the model’s ability to perceive boundaries. A boundary awareness 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) with 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 and reducing the difficulty of pixel identification at the junction of different land cover categories.

[0064] 3. In this invention, the boundary prior extracted by the boundary perception module is used to improve the feature representation capability generated by the decoder.

[0065] 4. This invention proposes a boundary guidance module called BGM. BGM adopts a dual-path boundary fusion strategy to optimize the feature representation of foreground and background. Attached Figure Description

[0066] Figure 1 This is a network architecture diagram of the boundary-guided method for extracting cultivated land from agricultural remote sensing images in this invention.

[0067] Figure 2 This is a schematic diagram of the texture enhancement module TEM in this invention;

[0068] Figure 3 This is a schematic diagram of the Boundary Awareness Module (BAM) in this invention;

[0069] Figure 4 This is a schematic diagram of the boundary guidance module (BGM) in this invention;

[0070] Figure 5 The image shows the segmentation effect and segmentation label of the remote sensing image farmland extraction network of the present invention on the Huiwen Town dataset.

[0071] Figure 6 The image shows the segmentation effect and segmentation label of the remote sensing image farmland extraction network of the present invention on the Puqian Town dataset.

[0072] Figure 7 The images show the segmentation results and segmentation labels of the remote sensing image farmland extraction network of this invention on the IFLYTEK dataset. Detailed Implementation

[0073] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] Reference Figures 1-7 The present invention provides an embodiment of a method for extracting cultivated land from agricultural remote sensing images based on boundary guidance, comprising the following steps:

[0075] A) Dataset Preprocessing: High-resolution optical remote sensing farmland images are cropped and preprocessed to form a preprocessed dataset, which is then divided into training, validation, and test sets.

[0076] B) Data augmentation: Perform general data augmentation or standardization on the dataset to obtain a data-augmented remote sensing farmland image dataset;

[0077] C) Feature extraction: Input the dataset images into the feature extraction module of a deep convolutional neural network in batches to extract multi-level features;

[0078] D) Texture Feature Enhancement: The texture enhancement module TEM is used to enhance the texture representation of low-level features extracted by deep convolutional neural networks;

[0079] E) Boundary Feature Enhancement: The boundary-aware module BAM is used to mine boundary representations in objects from low-level features containing texture details and high-level features containing abstract semantics. A boundary loss function based on geometric similarity is used to supervise the boundary extraction process.

[0080] F) Boundary-guided feature fusion: Multiple boundary guidance modules (BGM) are used to combine the boundary details obtained from BAM with the multi-level backbone features of each layer to guide the decoder's feature learning.

[0081] G) Segmentation prediction: Finally, the feature map output by the decoder is used for segmentation prediction to obtain the final farmland image segmentation result;

[0082] H) Network model training: The segmentation results predicted by the farmland segmentation network are compared with the actual farmland segmentation labels to calculate the smooth label cross-entropy loss. The network parameters are updated through backpropagation to improve the accuracy of the farmland segmentation network prediction results.

[0083] Among them, reference Figure 2 In step D), the texture enhancement module TEM independently adjusts the weight of each group of features after grouping them, and assigns different attention weights to different spatial locations, thereby improving the ability to express features and reducing noise interference.

[0084] Step D) includes the following steps:

[0085] D1) For height is H Width is W The number of channels is C The convolutional feature map is divided along the channel dimension into K For each feature map in a group, each spatial location is represented as a one-dimensional vector, denoted as . ;

[0086] Specifically, by applying the global spatial averaging function To approximate the semantic vector of the global features learned by this group :

[0087] (1)

[0088] semantic vectors utilizing global features TEM generates an importance coefficient for each local feature and evaluates it using a dot product operation. Similarity between Then, spatial normalization is used to suppress differences and noise effects between different samples:

[0089] (2)

[0090] (3)

[0091] in, and They represent the similarity coefficients respectively. The mean and variance, The constants used for numerical stability, to ensure normalization in the inserted network, can represent identity transformations, with each coefficient... A pair of parameters were introduced. and These parameters scale and shift the normalized values:

[0092] (4)

[0093] Finally, in order to obtain enhanced texture feature vectors The sigmoid function is applied to the original vectors throughout the entire vector space. Scaling is performed, and a residual structure is introduced to ensure stable backpropagation of the gradient:

[0094] (5)

[0095] All the enhanced features processed by the texture enhancement module form the final feature set. By calculating the texture similarity of different spatial locations in the feature map and assigning higher weights to spatial locations with similar texture characteristics, the texture enhancement module significantly improves the network's ability to perceive the texture information of different land features in remote sensing images.

[0096] refer to Figure 3 In step E), 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 model is used to model the boundary information related to the object, thereby forming a richer and more accurate boundary representation.

[0097] Step E) includes the following steps:

[0098] E1) First in the horizontal direction and vertical direction The Sobel operator is applied in the direction to obtain the gradient map. Specifically, two are first initialized. With parameters fixed and a convolution operation with a stride of 1 applied, the parameters of the two convolution kernels are initialized as follows: and :

[0099] (6)

[0100] E2) Apply these two convolutions to the input feature map respectively. Obtain gradient map :

[0101] (7)

[0102] (8)

[0103] (9)

[0104] in, and 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 Boundary strength, This represents the total boundary strength regardless of direction; it indicates the horizontal gradient. and vertical gradient The vector sum, followed by the gradient map. Depend on Function normalization and fusion with the input feature map to obtain a boundary-enhanced feature map :

[0105] (10)

[0106] in, This represents element-wise multiplication. express function;

[0107] After E3, and The boundary enhancement feature map is fused with a simple stacked convolutional layer. Specifically, the feature map is first... Applying bilinear upsampling operation and Convolution to obtain Feature maps of the same size Then, apply respectively Convolutional operations are used to adjust the channel sizes of these two features. Finally, the two feature maps are concatenated along the channel dimension, and multiple stacked convolutional layers are applied to obtain the final boundary enhancement features. The function obtains :

[0108] (11)

[0109] (12)

[0110] (13)

[0111] in, Indicates an upsampling operation. express convolution, This indicates a feature concatenation operation. express function, It means to pass through in order. Convolution, batch normalization, ReLU activation function and convolution.

[0112] refer to Figure 4 In step F), the boundary guidance module (BGM) accepts two inputs: one is the boundary enhancement feature. and corresponding features of the encoder network The first is the mixed feature obtained by multiplication; the second is the feature from the previous decoder layer. These two inputs are fed into the foreground and background paths of the BGM, respectively, to facilitate feature representations of the foreground and background.

[0113] Preferably, step F) includes the following steps:

[0114] F1) In the foreground path, by and The foreground features are obtained by directly connecting along the channel dimension and then sequentially passing them through Conv-BN-ReLU. :

[0115] (14)

[0116] in, This represents element-wise multiplication. This indicates that the sequence passes through Conv-BN-ReLU in order.

[0117] F2) For background paths, a background attention component is used to allow the model to selectively focus on background information. The background attention component obtains background features. Represented as:

[0118] (15)

[0119] in, This represents a three-layer stacked Conv-BN-ReLU. This represents the Sigmoid function. This represents element-wise multiplication. The background attention map is represented by applying the sigmoid function to the feature map of the previous decoder layer and subtracting the resulting foreground attention map from 1. To generate;

[0120] F3) Finally, foreground features Background features and the features of the previous decoder These are concatenated along the channel dimension to form the final decoder output. :

[0121] (16).

[0122] A boundary-guided agricultural remote sensing image farmland extraction system, using the aforementioned boundary-guided agricultural remote sensing image farmland extraction method, includes a preprocessing and data augmentation module, a texture enhancement module (TEM), a boundary awareness module (BAM), and a boundary guidance module (BGM).

[0123] The preprocessing and data augmentation module is used to perform image cropping and data augmentation on the input high-resolution optical remote sensing farmland segmentation images to form a preprocessed image dataset;

[0124] The Texture Enhancement Module (TEM) is used to improve the ability to represent features and reduce noise interference by independently adjusting the weight of each group of features after grouping them and assigning different attention weights to different spatial locations.

[0125] The Boundary Awareness Module (BAM) combines low-level features generated by the encoder through the Texture Enhancement Module (TEM). Texture details and advanced features The semantic information in the model is used to model the boundary information related to the object, thereby forming a richer and more accurate boundary representation;

[0126] The Boundary Guidance Module (BGM) is used to extract boundary priors, improving the feature representation capabilities of the decoder. It employs a dual-path boundary fusion strategy to optimize the feature representations of the foreground and background, obtaining the final output feature map of the decoder. This map is then used for segmentation prediction, and the predicted remote sensing farmland segmentation results are output to the user.

[0127] Boundary-guided deep convolutional neural networks employ an encoder-decoder architecture. The encoder includes a feature extraction module, a texture enhancement module, and a boundary awareness module, while the decoder includes a boundary guidance module and a segmentation prediction module.

[0128] In this embodiment, the deep convolutional neural network is a ResNet50 structure.

[0129] refer to Figures 5-7 This embodiment was compared with several state-of-the-art methods in terms of segmentation accuracy, including U-Net, PSPNet, DANet, DeepLabV3+, ABCNet, HRCNet, and MANet. The remote sensing farmland segmentation datasets used in this experiment were the Huiwen Town dataset, Puqian Town dataset, and IFLYTEK dataset. The comparative experiments used the model methods U-Net, PSPNet, DANet, DeepLabV3+, ABCNet, HRCNet, and MANet, which were performed as described in the original paper.

[0130] Table 1 shows the experimental results of this invention and other methods for farmland segmentation in remote sensing images on the Huiwen Town dataset. OA represents the overall accuracy, F1-Score represents the F1 score, and mIoU represents the mean intersection-union ratio. The results are shown in the table below:

[0131] Table 1

[0132]

[0133] Table 2 shows the experimental results of this invention and other methods for farmland segmentation in remote sensing images on the Puqian Town dataset. OA represents the overall accuracy, F1-Score represents the F1 score, and mIoU represents the mean intersection-union ratio. The results are shown in the table below:

[0134] Table 2

[0135]

[0136] Table 3 shows the experimental results of this invention and other methods for farmland segmentation in remote sensing images on the IFLYTEK dataset. OA represents the overall accuracy, F1-Score represents the F1 score, and mIoU represents the mean intersection-over-union ratio. The results are shown in the table below:

[0137] Table 3

[0138]

[0139] According to Tables 1, 2 and 3, the farmland extraction method based on boundary guidance in agricultural remote sensing images proposed in this invention has higher accuracy than other methods on the three datasets.

[0140] Finally, it should be noted that the above description 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 foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting cultivated land from an agricultural remote sensing image based on boundary guidance, characterized in that, The method comprises the following steps: A) data set preprocessing: cropping and preprocessing high-resolution optical remote sensing cultivated land images to form a preprocessed data set, and dividing the data set into a training set, a validation set and a test set; B) data augmentation: performing general data augmentation or standardization processing on the data set to obtain a data augmented remote sensing cultivated land image data set; C) feature extraction: inputting the data set images into the feature extraction module of the deep convolutional neural network in batches to extract multi-level features; D) texture feature enhancement: using a texture enhancement module TEM to enhance the texture representation of the low-level features extracted by the deep convolutional neural network; E) boundary feature enhancement: using a boundary awareness module BAM to mine the boundary representation in the objects from the low-level features containing texture details and the high-level features containing abstract semantics, and adopting a boundary loss function based on geometric similarity to supervise the extraction process of the boundary; F) boundary guided feature fusion: using multiple boundary guiding modules BGM to combine the boundary details obtained from the BAM with the multi-level backbone features of each layer to guide the decoder feature learning; In step F), the boundary guidance module BGM takes two inputs: one is the boundary enhanced feature and the corresponding feature of the encoder network which are multiplied to get the mixed feature; the other is the feature from the previous decoder layer , which are fed into the foreground path and the background path of the BGM respectively, aiming to promote the feature representation of foreground and background respectively; The step F) comprises the following steps: F1) In the foreground path, by connecting and directly along the channel dimension, then sequentially through Conv-BN-ReLU to obtain foreground features : ; wherein, represents element-wise multiplication, represents sequentially passing through Conv-BN-ReLU, F2) for the background path, using a background attention component to let the model selectively focus on background information, the background features obtained by the background attention component is represented as: ； wherein, denotes a three-layer stacked Conv-BN-ReLU, denotes a Sigmoid function, denotes an element-wise multiplication, denotes a background attention map, which is generated by applying a Sigmoid function to the feature maps of the previous decoder layer and subtracting the resulting foreground attention map from 1 ; F3) Finally, foreground features , background features and previous decoder features are concatenated in the channel dimension to form the output of the final decoder : ； G) segmentation prediction: finally performing segmentation prediction on the decoder output feature map to obtain the final cultivated land image segmentation result; H) network model training: performing smooth label cross-entropy loss calculation on the segmentation result predicted by the cultivated land segmentation network and the real cultivated land segmentation label, updating the network parameters through back propagation, and improving the prediction result accuracy of the cultivated land segmentation network. 2.The boundary-guided agricultural remote sensing image cropland extraction method according to claim 1, characterized in that, In the step D), the texture enhancement module TEM adjusts the weight of each group of features independently after grouping the features, and gives different attention weights to different spatial positions, thereby improving the feature expression ability and reducing noise interference. 3.The boundary-guided agricultural remote sensing image cropland extraction method of claim 2, wherein, The step D) comprises the following steps: D1) for a convolutional feature map with height H , width W , and number of channels C , divide it into K groups along the channel dimension, and for each spatial location in each group of feature maps represent it as a one-dimensional vector, denoted as ; In particular, by applying a global spatial average function to approximate the semantic vector of the global feature of this group learning representation : ； semantic vectors utilizing global features TEM generates an importance coefficient for each local feature and evaluates it using a dot product operation. and similarity between Then, spatial normalization is used to suppress differences between different samples and the influence of noise. ； ; in, and They represent the similarity coefficients respectively. The mean and variance, The constants used for numerical stability, to ensure normalization in the inserted network, can represent identity transformations, with each coefficient... A pair of parameters were introduced. and These parameters scale and shift the normalized values ​​to obtain the offset values. : ; Finally, to obtain enhanced texture feature vectors The original are scaled by sigmoid function in the whole vector group space, and the residual structure is introduced to ensure the stable back propagation of gradient: ； All the enhanced features processed by the texture enhancement module form the final feature group The texture enhancement module significantly improves the network's perception of different ground object texture information in remote sensing images by calculating the texture similarity of different spatial positions in the feature map and giving higher weights to spatial positions with similar texture characteristics. 4.The boundary-guided agricultural remote sensing image cropland extraction method of claim 1, wherein, In step E), the boundary-aware module BAM combines the low-level features of texture details and high-level features of semantic information to model the boundary information associated with the object, resulting in a more rich and accurate boundary representation. 5.The boundary-guided agricultural remote sensing image cropland extraction method of claim 4, wherein, The step E) comprises the following steps: E1) first in the horizontal direction and in the vertical direction apply Sobel operator to obtain gradient map, specifically, first initialize two parameters fixed and apply convolution operation with step size 1, two convolution kernel parameters initialization is and : ; E2) Then, these two convolutions are applied to the input feature map separately to obtain the gradient map : ； ； ； wherein, and denotes performing a convolution operation on the feature with a specific convolution kernel , the element size of the gradient map and respectively represent the boundary strength of the feature in the horizontal direction and the vertical direction , the gradient map is normalized by function and fused with the input feature map to obtain the boundary-enhanced feature map : ； wherein denotes element-wise multiplication, denotes function; E3) Then, the boundary enhanced feature map of and is fused with the simple stacked convolutional layers. Specifically, first, the feature map is applied with bilinear up-sampling operation and convolution to obtain a feature map with the same size as , then, respectively, the two features are applied with convolution operation to adjust the channel size of the two features, finally, the two feature maps are concatenated along the channel dimension, and a multi-layer stacked convolutional layer is applied to obtain the final boundary enhanced feature and obtain by function: ; ； ； wherein, denotes an up-sampling operation, denotes convolution, denotes a feature concatenation operation, denotes function, denotes sequentially passing through in order, i.e. convolution, batch normalization, activation function, and convolution. 6.A boundary-guided cropland extraction system based on agricultural remote sensing images, characterized by, The boundary guided agricultural remote sensing image cultivated land extraction method according to any one of claims 1-5 comprises a preprocessing and data enhancement module, a texture enhancement module TEM, a boundary awareness module BAM and a boundary guiding module BGM; The preprocessing and data enhancement module is used for 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 for adjusting the weight of each group of features independently after grouping the features, and giving different attention weights to different spatial positions, thereby improving the feature expression ability and reducing noise interference; The boundary awareness module BAM combines low-level features generated by the encoder through the texture enhancement module TEM. Texture details and advanced features The semantic information in the model is used to model the boundary information related to the object, thereby forming a richer and more accurate boundary representation; The boundary guiding module BGM is used for extracting boundary priori, improving the feature representation ability of the decoder, adopting a double-path boundary fusion strategy, optimizing the feature representation of the foreground and the background, obtaining the output feature map of the final decoder, performing segmentation prediction on the output feature map, and outputting the predicted remote sensing cultivated land segmentation result to the user.

7. The boundary-guided based agricultural remote sensing image cultivated land extraction system according to claim 6, characterized in that, The boundary guided deep convolutional neural network adopted by the system is an encoder-decoder architecture, wherein the encoder comprises a feature extraction module, a texture enhancement module and a boundary awareness module, and the decoder comprises a boundary guiding module and a segmentation prediction module.

Images