Field plot extraction method and system based on boundary extraction and breakpoint connection post-processing

Abstract

The application relates to the technical field of intelligent agriculture, and discloses a field plot extraction method and system based on boundary extraction and breakpoint connection post-processing, which comprises the following steps: collecting farmland remote sensing image data and constructing a labeled data set. A boundary extraction model comprising a connectivity attention module is constructed. Image data in the labeled data set is input into the boundary extraction model to output a boundary probability graph and a directional connectivity probability graph. Double-threshold boundary post-processing is performed on the boundary probability graph to obtain a boundary region binary graph, and center line extraction is performed on the boundary region binary graph to obtain a first boundary extraction result. Breakpoint detection is performed on the first boundary extraction result, and elongation processing is performed on the breakpoints until no breakpoint is detected on the boundary line, so that a second boundary extraction result is obtained. Vector conversion is performed on the second boundary extraction result to obtain a farmland field plot extraction result. The application can obtain a complete farmland field plot extraction result, and improves the precision of farmland field plot extraction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent agriculture, and more particularly to a field plot extraction method and system based on boundary extraction and breakpoint connection post-processing. BACKGROUND

[0002] In modern agricultural technology, fine extraction of farmland plots can provide basic data for grain production and monitoring and evaluation of farmland conditions, and can achieve effective, high-quality and sustainable allocation of farmland resources.

[0003] In current farmland plot extraction methods, a boundary extraction algorithm is often introduced to extract the edges of the plots, plot the boundaries of each plot, and convert the boundary extraction result into a farmland plot result. However, the current farmland plot extraction method based on the boundary extraction algorithm mainly detects the gray level change of the image, plots the boundaries of each plot, and introduces a large amount of non-plot boundary redundancy, which cannot obtain complete farmland plots, resulting in low precision of the farmland plot extraction result. SUMMARY

[0004] To overcome the defects of the prior art that cannot obtain complete farmland plots and the low precision of the farmland plot extraction result, the present application provides a farmland plot extraction method and system based on boundary extraction and breakpoint connection post-processing.

[0005] To solve the above technical problems, the technical solutions of the present application are as follows:

[0006] In a first aspect, the present application provides a farmland plot extraction method based on boundary extraction and breakpoint connection post-processing, comprising:

[0007] Collecting farmland remote sensing image data and labeling the boundaries of the plots of the farmland remote sensing image data to obtain a labeled data set.

[0008] A boundary extraction model including a connectivity attention module is constructed. The connectivity attention module is used to learn the directional connectivity in the plot boundary.

[0009] The image data in the labeled data set is input into the trained boundary extraction model for boundary extraction, and the boundary extraction model outputs a boundary probability map and a directional connectivity probability map.

[0010] The boundary probability map is subjected to double-threshold boundary post-processing to obtain a boundary region binary map, and the center line of the boundary region binary map is extracted to obtain a first boundary extraction result.

[0011] The first boundary extraction result is subjected to breakpoint detection, and the breakpoints of the first boundary extraction result are subjected to extension processing according to the boundary probability graph and the direction connectivity probability graph until the boundary lines in the first boundary extraction result no longer detect breakpoints, and a second boundary extraction result is obtained.

[0012] The second boundary extraction result is subjected to vector conversion, and a farmland plot extraction result is obtained.

[0013] In a second aspect, the present application further provides a plot extraction system based on boundary extraction and breakpoint connection post-processing, comprising:

[0014] A data collection module is configured to collect farmland remote sensing image data, and mark the boundaries of the plots in the farmland remote sensing image data to obtain a labeled data set.

[0015] A boundary extraction model is configured to input the labeled data set into the boundary extraction model, and output a boundary probability graph and a direction connectivity probability graph. The boundary extraction model comprises a connectivity attention module configured to learn the direction connectivity in the boundaries of the plots.

[0016] A first boundary result extraction module is configured to perform double-threshold boundary post-processing on the boundary probability graph to obtain a boundary region binary graph, and perform center line extraction on the boundary region binary graph to obtain a first boundary extraction result.

[0017] A second boundary result extraction module is configured to perform breakpoint detection on the first boundary extraction result, and perform extension processing on the breakpoints of the first boundary extraction result according to the boundary probability graph and the direction connectivity probability graph until the boundary lines in the first boundary extraction result no longer detect breakpoints, and obtain a second boundary extraction result.

[0018] A conversion module is configured to perform vector conversion on the second boundary extraction result to obtain a farmland plot extraction result.

[0019] Compared with the prior art, the technical scheme of the present application has the beneficial effects that: the boundary extraction network based on the deep learning idea and the introduction of the connectivity attention module learns the boundary information and the direction connectivity information of the farmland remote sensing image data, can learn more detailed boundary information, and performs breakpoint connection post-processing on the boundary extraction result by using the direction connectivity probability until the breakpoints of the boundary are completely connected, so that more boundaries can be completely closed into a plane, thereby obtaining a high-precision and complete farmland plot extraction result. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the plot extraction method based on boundary extraction and breakpoint connection post-processing in the embodiments of the present application.

[0021] Figure 2 The network structure diagram of the boundary extraction model of the embodiment of the present application.

[0022] Figure 3 The flowchart of the second boundary extraction result in the embodiment of the present application.

[0023] Figure 4 The architecture diagram of the field block extraction system based on boundary extraction and breakpoint connection post-processing in the embodiment of the present application. DETAILED DESCRIPTION

[0024] The accompanying drawings are only for illustrative purposes and cannot be understood as a limitation of the patent;

[0025] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0026] Embodiment one

[0027] Please refer to Figure 1 The embodiment proposes a field block extraction method based on boundary extraction and breakpoint connection post-processing, which comprises:

[0028] S1: Collect farmland remote sensing image data, and perform field block boundary labeling on the farmland remote sensing image data to obtain a labeled data set.

[0029] S2: A boundary extraction model including a connectivity attention module is constructed. The connectivity attention module is used to learn the directional connectivity in the field block boundary.

[0030] S3: The image data in the labeled data set is input into the trained boundary extraction model for boundary extraction, and the boundary extraction model outputs a boundary probability map and a directional connectivity probability map.

[0031] S4: Double-threshold boundary post-processing is performed on the boundary probability map to obtain a boundary region binary map, and center line extraction is performed on the boundary region binary map to obtain a first boundary extraction result.

[0032] S5: Breakpoint detection is performed on the first boundary extraction result, and according to the boundary probability map and the directional connectivity probability map, the breakpoints of the first boundary extraction result are elongated until the boundary lines in the first boundary extraction result no longer detect breakpoints, to obtain a second boundary extraction result.

[0033] S6: Vector conversion is performed on the second boundary extraction result to obtain a farmland field block extraction result.

[0034] The field block extraction method based on boundary extraction and breakpoint connection post-processing provided in the embodiment can learn more detailed boundary information by using a boundary extraction network based on the deep learning idea and introducing a connectivity attention module to learn boundary information and directional connectivity information of farmland remote sensing image data, and can use directional connectivity probability to perform breakpoint connection post-processing on the boundary extraction result until the breakpoints of the boundary are completely connected, so that more boundaries can be completely closed into a plane, thereby obtaining a high-precision and complete farmland field block extraction result, which is more suitable for high-resolution remote sensing images and greatly reduces the workload of manual correction.

[0035] Embodiment two

[0036] Reference Figure 1 The method provided in the embodiment is improved on the basis of the field block extraction method based on boundary extraction and breakpoint connection post-processing provided in embodiment one, and the method comprises the following steps.

[0037] S1: Collect farmland remote sensing image data, and perform field block boundary labeling on the farmland remote sensing image data to obtain a labeled data set.

[0038] In the embodiment, the collected farmland remote sensing image data needs to be geometrically and radiometrically corrected to ensure the consistency of the geometric and radiometric conditions of the images used for training. The farmland remote sensing images need to be preliminarily screened to retain most of the images containing farmland and appropriately introduce a small amount of images containing only non-farmland features as negative samples. In the labeling process, relevant professionals visually interpret, judge the farmland range and boundary, and then draw and label. The labeling content is the farmland range and the farmland boundary. The labeling result generates a farmland range binary image and a boundary binary image, and is converted into vector data with geographic coordinates. In the embodiment, a total of 2874 image sample pairs of 512x512 size are labeled.

[0039] In the embodiment, after obtaining the labeled data set, the image data in the labeled data set is subjected to data enhancement processing to improve the data volume and enhance the generalization ability of the network, and the image data subjected to the data enhancement processing is subjected to standardization processing.

[0040] The data enhancement processing of the image data in the labeled data set specifically includes horizontal flipping, vertical flipping, deformation, cropping, brightness adjustment, and adding Gaussian noise to increase the training samples, so that the number of image samples meets the feature learning requirements of the network model, and can solve the problem of unbalanced farmland samples to a certain extent.

[0041] The standardization processing of the image data subjected to the data enhancement processing has the following expression:

[0042]

[0043] wherein x is the pixel value of the image, and μ and σ are the mean and standard deviation of the image data corresponding to the waveband, respectively.

[0044] S2: a boundary extraction model including a connectivity attention module is constructed. The connectivity attention module is used to learn the directional connectivity in the field boundary.

[0045] In this embodiment, the labeled data set is divided into a training set and a test set in a ratio of 80% and 20%, and the training set is used to train the boundary extraction model until the loss function Loss(p, y) of the boundary extraction model converges, thereby obtaining the trained boundary extraction model.

[0046] The loss function Loss(p, y) of the boundary extraction model includes cross-entropy loss Loss CE (p, y) and Dice loss Loss Dice (p, y), and the expressions are as follows:

[0047] Loss(p, y) = aLoss CE (p, y) + bLoss Dice (p, y),

[0048]

[0049]

[0050] wherein a and b are adjustable hyperparameters, y i represents the true value of the label i, i = 0 represents the background class label, and i = 1 represents the real field boundary label, p i is the predicted value of the label i, and N is the number of labels.

[0051] As Figure 2 shown, Figure 2 is the network structure diagram of the boundary extraction model of the embodiment of the present application. In this embodiment, the boundary extraction model includes an encoder, a decoder, and a connectivity attention module connected in sequence. The boundary extraction model adopts an encoder-decoder structure based on a convolutional neural network. The encoder extracts image features through multiple convolutional layers, and the decoder mines specific boundary information based on the extracted features.

[0052] In this embodiment, the encoder is a multi-scale high-resolution network including four parallel residual convolution modules, and the image data in the labeled data set is transmitted to the four parallel residual convolution modules respectively. After the four parallel residual convolution modules output feature maps of different scales respectively, the first boundary feature map is obtained after splicing processing. The encoder structure includes multiple parallel multi-scale resolution feature branches, which enables the network to learn stronger semantic information and accurate position information at the same time. Each branch is composed of multiple residual modules. Among them, the residual module F is composed of a convolution layer, a batch normalization layer and an activation function. The input and output of each block are connected by a jump connection. The output of each layer of the network can be represented as:

[0053] x l+1 =x l +F(x l )

[0054] where x l represents the feature of the lth layer. The decoder as a whole can be divided into four stages. Each stage will perform feature down-sampling to obtain features with lower spatial resolution and deeper representation. In order to preserve the detailed information at high resolution scale, the network structure is parallelized to retain features at different scales, so that the features of the next stage of the network can obtain additional high-resolution features of the previous stage. Information exchange is performed between each stage through a conversion fusion layer. For s input features {X1,X2,…,X s}, the output of the corresponding resolution s features {Y1,Y2,…,Y s} after the exchange unit, each output is aggregated from s input features:

[0055]

[0056] where the function a(X i ,k) includes up-sampling or down-sampling of the input feature X i from resolution i to k, wherein the down-sampling is realized by different number of 3×3 convolution with step 2; the up-sampling uses nearest neighbor sampling, and a 1×1 convolution is used after it to align the number of feature channels. If i=k, let a(X i ,k)=X i .

[0057] The convolution layer of the decoder uses a 3×3 convolution function, a SiLU activation function and a batch normalization layer. Among them, the SiLU activation function is represented as:

[0058] In this embodiment, the decoder uses the real farmland boundary label to supervise the learning of the first boundary feature map to obtain a boundary probability map, and the specific steps include: using the real farmland boundary label to perform semantic segmentation on the first boundary feature map to obtain a semantic segmentation result of the first boundary feature map; performing object-based context information learning on the first boundary feature map to obtain a second boundary feature map; the second boundary feature map is subjected to feature enhancement learning through a self-attention module to obtain a third boundary feature map; and the semantic segmentation result and the third boundary feature map are subjected to splicing processing to obtain the boundary probability map.

[0059] The decoder part increases the feature representation of the model to the context information enhancement, generates multi-scale features representing object information and background information at the same time through convolution results, and strengthens the extraction of field block features. The input of the decoder is the multi-scale first boundary feature map extracted by the decoder in the previous step. First, the real result is introduced for supervised training to obtain a coarse boundary extraction result, then based on the coarse boundary extraction result and the deep network features, the semantic class feature representation corresponding to each vector in the feature is obtained through the convolution layer, and finally the class self-attention mechanism is used to obtain the context feature through the relationship matrix between the deep feature and the object region feature. Ultimately, the object context feature representation and the deep feature of the original main network are spliced to obtain the enhanced feature for outputting the segmentation result. When the object context feature representation and the feature representation input into the deepest layer of the network are spliced as the context information enhanced feature representation, the semantic class of each pixel can be predicted based on the enhanced feature representation. The object-based context feature representation can calculate a group of target region feature representations, so as to propagate these object region feature representations to each pixel according to the similarity between the region feature representation at the object level and the pixel level. The loss function of the decoder is represented as: L aux =Loss(y soft ,y label ), y soft is the preliminary semantic segmentation result, and y label is the real farmland boundary label.

[0060] In this embodiment, the connectivity attention module learns the direction information in the field block boundary based on the boundary probability map to output a direction connectivity probability map, and the specific steps include: after the boundary probability map is subjected to receptive field expansion through a plurality of convolution layers with different expansion rates, the feature compression in the channel dimension is performed through global average pooling to obtain a feature vector z c , representing global information. The expression of the global average pooling is as follows:

[0061]

[0062] wherein, x cwhere c is the number of input channels, H and W are the size of the feature map, and i and j represent the corresponding spatial position in the feature map.

[0063] The feature vector is sequentially passed through a fully connected layer and a sigmoid activation function to obtain the global attention weight, which is expressed as follows:

[0064] The feature vector representing global information is passed through a fully connected layer and a sigmoid function to calculate the attention weight corresponding to each channel, which can be expressed as:

[0065] a = σ(W2δ(W1z))

[0066] where z is the input feature vector, W1 and W1 are the weights of the fully connected layer, δ is the ReLU activation function, and σ is the sigmoid function. By further learning the feature vector and using the sigmoid function to convert the feature to an attention weight ranging from 0 to 1.

[0067] The boundary probability map is multiplied by the global attention weight, and the multiplication result is passed through a residual module to obtain the directional connectivity probability map.

[0068] The connectivity attention module is introduced to supervise the training of the boundary probability map, and the directional information of the farmland boundary is learned. The connectivity attention module first inputs the image features output by the decoder into 3x3 convolution modules with dilation rates of 1 and 3, respectively, to expand the receptive field and extract more local information. Then, the features are globally averaged and compressed in the channel level, and two fully connected layers are used to learn the feature attention to obtain a vector ranging from 0 to 1. Multiplying this vector with the input feature can obtain the connectivity attention, which is finally supervised by the boundary probability map generated by the real label to obtain the directional connectivity probability map, and its loss function is expressed as:

[0069]

[0070] where C O is the number of channels of the directional connectivity probability map, N = H x W is the total number of pixels, is the connectivity label of the pixel at position i and its adjacent pixels in a certain direction of the directional connectivity probability map, is the predicted value of the directional connectivity probability map.

[0071] S3: input image data in the labeled dataset into the trained boundary extraction model for boundary extraction, the boundary extraction model outputs a boundary probability map and a direction connectivity probability map. S4: double-threshold boundary post-processing is performed on the boundary probability map to obtain a boundary region binary graph, and center line extraction is performed on the boundary region binary graph to obtain a first boundary extraction result, the specific steps comprising:

[0072] Step 1: set a high threshold and a low threshold; in this embodiment, the high threshold is set to 0.5 and the low threshold is set to 0.1.

[0073] Step 2: set a strong boundary region for the region in the boundary probability map where the boundary probability prediction value is higher than the high threshold;

[0074] Step 3: set a non-boundary region for the region in the boundary probability map where the boundary probability prediction value is lower than the low threshold;

[0075] Step 4: set a weak boundary region for the region in the boundary probability map where the boundary probability prediction value is lower than the high threshold and higher than the low threshold;

[0076] Step 5: perform 8-neighborhood direction summation on a certain pixel in the weak boundary region to obtain the 8-neighborhood probability sum of the pixel;

[0077] Step 6: if the 8-neighborhood probability sum of the pixel is higher than the high threshold, set the pixel as a strong boundary region;

[0078] Step 7: repeat steps 5-6 until all pixels in the weak boundary region are judged as strong boundary regions, and generate a boundary region binary graph using the set strong boundary regions.

[0079] S5: perform breakpoint detection on the first boundary extraction result, and perform elongation processing on the breakpoints of the first boundary extraction result according to the boundary probability map and the direction connectivity probability map, until no breakpoint is detected in the boundary line of the first boundary extraction result, to obtain a second boundary extraction result, as shown in Figure 3 Figure 3 is a flowchart of the second boundary extraction result in the embodiments of the present application, and the specific steps comprise:

[0080] Step A: stack all breakpoints in the first boundary extraction result to obtain a stack including breakpoint elements;

[0081] Step B: judge whether the stack is empty, if the stack is not empty, execute step C, if the stack is empty, output the second boundary extraction result and end;

[0082] ​Step C: pop the breakpoint from the stack, and read the elongation direction of the breakpoint in the first boundary extraction result based on the position of the breakpoint, and set the boundary probability and the connected direction probability of the pixel position in the opposite direction of the elongation direction to 0;

[0083] Step D: load the boundary probability and the connected direction probability of the 8-neighborhood directions at the breakpoint popped from the stack, and perform weighted summation on the boundary probability and the connected direction probability of the 8-neighborhood directions respectively; in this embodiment, the weight of the boundary probability is set to 0.4, and the weight of the connected direction probability is set to 0.6.

[0084] Step E: compare the summation values of the boundary probability and the connected direction probability of the 8-neighborhood directions, and set the pixel position of the neighborhood direction with the larger summation value of the probability as a new boundary value;

[0085] Step F: determine whether the position of the new boundary value is a breakpoint, if yes, push the position of the breakpoint into the stack and jump to Step B, if not, directly jump to Step B.

[0086] For example, all the boundary pixels in the first boundary extraction result are judged based on the 8-neighborhood directions, if the boundary pixel is connected with the pixels in the surrounding 8 directions, i.e. the position is connected in the direction, then the connectedness image in the direction is recorded as 1, otherwise as 0. When judging the adjacent pixels, a certain distance d is set, i.e. whether the pixels in the certain direction within the distance d are all boundaries is judged, if yes, the direction is considered to be connected, and the connectedness image value is recorded as 1, and finally the second boundary extraction result, i.e. the connectedness image O ∈ R H×W×8 , is generated, where H × W represents the size of the image, and the connectedness image contains 8 dimensions representing the connectedness of the 8-neighborhood. A certain pixel O i,j,c in the connectedness image represents the connectedness between the pixel and a certain pixel, where i, j represent the spatial position of the image pixel, and c represents the position of the adjacent pixel. Finally, the connectedness information in the 8-neighborhood is obtained by splicing to obtain the final second boundary extraction result, i.e. the connectedness image.

[0087] In this embodiment, the connectedness distance d is set to 1.

[0088] S6: performing vector conversion on the second boundary extraction result to obtain the farmland plot extraction result.

[0089] Embodiment Three

[0090] In order to explore the effectiveness of the connectivity attention module, the boundary extraction model without introducing the connectivity attention module, the boundary extraction model using the connectivity attention module with a connectivity distance d = 1, the model after breakpoint connection post-processing, and other semantic segmentation networks including DeepLabV3+, LinkNet, and DLinkNet are compared. In order to quantify the performance of the model, the accuracy of the model extraction result and the real label is evaluated after the model training. The indexes used include IoU, Recall and F1.

[0091] Table 1 Precision evaluation comparison of the boundary extraction model proposed in the application and other semantic segmentation networks

[0092]

[0093] As shown in Table 1, the boundary extraction model proposed in the application used in the embodiment of the application is better than the models participating in the comparison in each index, and can make more accurate and complete identification of farmland plots. In addition, using the boundary extraction model proposed in the application and introducing breakpoint connection post-processing can significantly improve the accuracy of farmland plots and achieve better results in each index. The post-processing method can effectively complete the area with low boundary prediction probability at the breakpoint, thereby optimizing the farmland plot extraction effect.

[0094] Embodiment three

[0095] Referring to Figure 4 , the embodiment proposes a plot extraction system based on boundary extraction and breakpoint connection post-processing, which comprises:

[0096] A data collection module is used to collect farmland remote sensing image data and mark the boundaries of the plots of the farmland remote sensing image data to obtain a labeled data set.

[0097] A boundary extraction model is used. The labeled data set is input into the boundary extraction model, and the boundary extraction model outputs a boundary probability map and a directional connectivity probability map. The boundary extraction model comprises a connectivity attention module, and the connectivity attention module is used to learn the directional connectivity in the plot boundary.

[0098] A first boundary result extraction module is used to perform double-threshold boundary post-processing on the boundary probability map to obtain a boundary region binary map, and to perform center line extraction on the boundary region binary map to obtain a first boundary extraction result.

[0099] A second boundary extraction module is configured to perform breakpoint detection on the first boundary extraction result, and perform elongation processing on the breakpoints of the first boundary extraction result according to the boundary probability map and the direction connectivity probability map, until no breakpoint is detected in the boundary line of the first boundary extraction result, to obtain a second boundary extraction result.

[0100] A conversion module is configured to perform vector conversion on the second boundary extraction result to obtain a farmland plot extraction result.

[0101] The farmland plot extraction system based on boundary extraction and breakpoint connection post-processing provided in the embodiment can learn more detailed boundary information by using a boundary extraction network based on the deep learning idea and introducing a connectivity attention module to learn boundary information and direction connectivity information of farmland remote sensing image data, and can use the direction connectivity probability to perform breakpoint connection post-processing on the boundary extraction result until the breakpoints of the boundary are completely connected, so that more boundaries can be completely closed into a plane, thereby obtaining a high-precision and complete farmland plot extraction result, which is more suitable for high-resolution remote sensing images and greatly reduces the workload of manual correction.

[0102] The terms used to describe the positional relationship in the drawings are only used for illustrative description, and should not be understood as a limitation on the patent;

[0103] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the embodiments of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and impossible to exhaust all the embodiments. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A method for extracting fields based on boundary extraction and post-processing of breakpoint connection, characterized in that, include: Collect farmland remote sensing image data and annotate the field boundaries of the farmland remote sensing image data to obtain an annotated dataset; A boundary extraction model including a connectivity attention module is constructed; the connectivity attention module is used to learn directional connectivity in the field boundaries. Image data from the labeled dataset is input into a trained boundary extraction model for boundary extraction. The boundary extraction model outputs a boundary probability map and a directional connectivity probability map. The boundary probability map is subjected to double threshold boundary post-processing to obtain a binary map of the boundary region, and the centerline of the binary map of the boundary region is extracted to obtain the first boundary extraction result. Breakpoint detection is performed on the first boundary extraction result, and the breakpoints in the first boundary extraction result are extended according to the boundary probability map and the directional connectivity probability map until no breakpoints are detected on the boundary lines in the first boundary extraction result, thus obtaining the second boundary extraction result, including: Step A: Push all breakpoints in the first boundary extraction result onto the stack to obtain a stack containing breakpoint elements; Step B: Determine if the stack is empty. If the stack is not empty, proceed to step C. If the stack is empty, output the second boundary extraction result and end. Step C: Pop the breakpoint from the stack, and based on the position of the breakpoint, read the extension direction of the breakpoint in the first boundary extraction result, and set the boundary probability and connected direction probability of the pixel position in the opposite direction of the extension direction to 0. Step D: Load the boundary probabilities and connected direction probabilities of the 8 neighborhood directions at the popped breakpoint, and perform a weighted summation of the boundary probabilities and connected direction probabilities of the 8 neighborhood directions respectively. Step E: Compare the sum of the boundary probabilities and connected direction probabilities in the 8-neighborhood directions, and set the pixel position of the neighborhood direction with the larger sum of probabilities as the new boundary value; Step F: Determine whether the location of the newly added boundary value is a breakpoint. If it is a breakpoint, push the location of the breakpoint onto the stack and jump to execute step B. If it is not a breakpoint, jump directly to execute step B. The second boundary extraction result is vectorized to obtain the farmland plot extraction result.

2. The field extraction method based on boundary extraction and breakpoint connection post-processing as described in claim 1, characterized in that, After constructing the boundary extraction model including the connectivity attention module, the method further includes: The labeled dataset is divided into a training set and a test set. The boundary extraction model is trained using the training set until the loss function of the boundary extraction model is optimized. The convergence process yields a well-trained boundary extraction model. The loss function of the boundary extraction model Including cross-entropy loss and Dice loss Its expression is as follows: ， ， in, and These are adjustable hyperparameters. Indicates that the label is The true value, Indicates background category tags, Indicates field boundary labels, For the tag The predicted value, N The number of tags.

3. The field extraction method based on boundary extraction and breakpoint connection post-processing as described in claim 1, characterized in that, The boundary extraction model includes an encoder, a decoder, and a connectivity attention module connected in sequence. The encoder is a multi-scale high-resolution network, including four parallel residual convolutional modules. The image data in the labeled dataset is transmitted to the four parallel residual convolutional modules respectively. The four parallel residual convolutional modules output feature maps of different scales respectively, which are then stitched together to obtain the first boundary feature map. The decoder uses real farmland boundary labels to supervise the learning of the first boundary feature map to obtain a boundary probability map; The connectivity attention module learns directional information in the field boundaries based on the boundary probability map and outputs a directional connectivity probability map.

4. The field extraction method based on boundary extraction and breakpoint connection post-processing according to claim 3, wherein the decoder uses real farmland boundary labels to perform supervised learning on the first boundary feature map to obtain a boundary probability map, specifically including: The first boundary feature map is semantically segmented using the real farmland boundary labels to obtain the semantic segmentation result of the first boundary feature map; The first boundary feature map is subjected to object-based contextual information learning to obtain the second boundary feature map; The second boundary feature map is processed by a self-attention module for feature enhancement learning to obtain the third boundary feature map; The semantic segmentation result and the third boundary feature map are concatenated to obtain a boundary probability map.

5. The field extraction method based on boundary extraction and breakpoint connection post-processing as described in claim 4, characterized in that, The object-based context information includes the query statement, key values, and data items.

6. The field extraction method based on boundary extraction and breakpoint connection post-processing as described in claim 3, wherein the connectivity attention module learns the directional connectivity in the field boundaries based on the boundary probability map and outputs a directional connectivity probability map, the specific steps of which include: The boundary probability map is expanded to expand the receptive field through several convolutional layers with different expansion rates, and then subjected to global average pooling for feature compression in the channel dimension to obtain a feature vector representing global information. The feature vectors are passed sequentially through a fully connected layer and a sigmoid activation function to obtain the global attention weights; The boundary probability map is multiplied with the global attention weights, and the result of the multiplication is then passed through a residual module to obtain the directional connectivity probability map.

7. The field extraction method based on boundary extraction and breakpoint connection post-processing according to claim 1, wherein the boundary probability map is subjected to double-threshold boundary post-processing to obtain a binary map of the boundary region, specifically including: Step 1: Set the high threshold and low threshold; Step 2: Set strong boundary regions for areas in the boundary probability map where the predicted boundary probability value is higher than the high threshold; Step 3: Set the regions in the boundary probability map whose predicted boundary probability values ​​are below a low threshold as non-boundary regions; Step 4: Set the regions in the boundary probability map whose predicted boundary probability values ​​are below the high threshold but above the low threshold as weak boundary regions; Step 5: In the weak boundary region, sum the 8-neighbor directions of a certain pixel to obtain the sum of the 8-neighbor probabilities of that pixel; Step 6: If the sum of the probabilities of the 8 neighbors of a pixel is higher than the high threshold, then the pixel is set as a strong boundary region. Step 7: Repeat steps 5-6 until all pixels in the weak boundary region have completed the strong boundary region judgment, and generate a binary map of the boundary region using the set strong boundary region.

8. The field extraction method based on boundary extraction and breakpoint connection post-processing as described in claim 1, characterized in that, After obtaining the labeled dataset, the method further includes: Data augmentation processing is performed on the image data in the labeled dataset, specifically including: horizontal flipping, vertical flipping, deformation, cropping, brightness adjustment, and adding Gaussian noise to the image data; The normalization process for image data that has undergone data augmentation is expressed as follows: Among them, the pixel values ​​in the original image The pixel values ​​of the image. and These represent the mean and standard deviation of the corresponding bands in the image data, respectively.

9. A field extraction system based on boundary extraction and breakpoint connection post-processing, applied to the field extraction method based on boundary extraction and breakpoint connection post-processing as described in any one of claims 1 to 8, comprising: The data collection module is used to collect farmland remote sensing image data and to annotate the field boundaries of the farmland remote sensing image data to obtain an annotated dataset; A boundary extraction model is provided, wherein the labeled dataset is input into the boundary extraction model, and the boundary extraction model outputs a boundary probability map and a directional connectivity probability map; the boundary extraction model includes a connectivity attention module, which is used to learn the directional connectivity in the field boundaries; The first boundary result extraction module is used to perform double-threshold boundary post-processing on the boundary probability map to obtain a binary map of the boundary region, and to extract the centerline of the binary map of the boundary region to obtain the first boundary extraction result. The second boundary result extraction module is used to perform breakpoint detection on the first boundary extraction result, and extend the breakpoints of the first boundary extraction result according to the boundary probability map and the directional connectivity probability map until the boundary line in the first boundary extraction result no longer detects breakpoints, thereby obtaining the second boundary extraction result; The conversion module is used to perform vector conversion on the second boundary extraction result to obtain the farmland plot extraction result.

Images