Farmland plot fine extraction method and system based on GF-7 image

Abstract

The invention belongs to the technical field of agricultural remote sensing, and relates to a farmland plot fine extraction method and system based on a GF-7 image. The method comprises the following steps: collecting multispectral remote sensing data and carrying out correction and image fusion processing; taking ridges as field block distribution boundaries to obtain label raster data; superposing, cutting and slicing the multispectral remote sensing data and the label raster data to obtain sample data; constructing a deep learning model, and training to obtain a field boundary recognition model; and collecting current multispectral remote sensing data and inputting the data into the trained field boundary recognition model to obtain a field boundary recognition result. According to the method, the ridges in the natural boundary of the farmland serve as the field distribution boundary, the multispectral remote sensing data and the label raster data are overlaid and then cut and sliced to obtain the sample data, efficient extraction of the field boundary is achieved in combination with the deep learning model, and the field-level segmentation precision and efficiency are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of agricultural remote sensing, and in particular, relates to a method and system for fine extraction of farmland plots based on GF-7 images. Background Art

[0002] Fine identification of farmland plots has important application value in the fields of farmland precision management and agricultural insurance. It is the foundation of digital agriculture and precision agriculture. The adjustment of planting structure helps to increase agricultural product output and promote the development of precision agriculture. At present, the acquisition of planting plot data in my country is mainly based on the second and third national land survey data, mainly relying on remote sensing interpretation combined with manual reporting methods, facing the problem of difficult to ensure the accuracy of data reporting and difficulty in monitoring and updating crop planting information.

[0003] The idea of ​​plot-oriented classification was first proposed to address the problems of spectral variation and spectral mixing in traditional pixel-oriented classification. Many research results show that plot-oriented crop extraction methods can achieve higher extraction accuracy. At present, sub-meter Google images or drone images or medium- and low-spatial resolution remote sensing images are generally used for plot-scale extraction. Traditional methods have the disadvantages of much manual intervention and poor timeliness. The method and system for plot extraction from high-resolution data, on the one hand, develops the application of high-resolution series data in precision agriculture and smart agriculture in China, and on the other hand, improves the accuracy of plot extraction while saving the cost of plot extraction. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method and system for fine extraction of farmland plots based on GF-7 images.

[0005] In a first aspect, the present invention provides a method for finely extracting farmland plots based on GF-7 images, comprising: Collect multispectral remote sensing data and perform panchromatic and multispectral orthorectification and image fusion processing; The ridges in the natural boundaries of farmland are used as the boundaries of field distribution, and field sample vector data is established through manual processing. The field sample vector data is converted into raster data with the same size as the farmland area in the multispectral remote sensing data, and the farmland area and non-farmland data in the raster data are defined to obtain the label raster data. The multispectral remote sensing data and the label grid data are superimposed and the clips are cut to obtain the sample data; Construct a deep learning model, input sample data into the deep learning model to train the deep learning model, calculate the loss function of the deep learning model, use the Adam optimization method to optimize the deep learning model, and obtain the trained field boundary recognition model; Collect current multispectral remote sensing data and input it into the trained field boundary recognition model to obtain the field boundary recognition result; The field boundary is expanded by dilation method, and the gaps between field boundaries are connected by closing operation method. The field boundary recognition result is obtained by increasing the continuity of field boundaries and removing surface objects.

[0006] In a second aspect, the present invention provides a farmland plot fine extraction system based on GF-7 images, comprising a first collection unit, a conversion unit, a sample data construction unit, a model construction and training unit, a second collection unit, and a data processing and output unit; The first collection unit is used to collect multispectral remote sensing data and perform panchromatic and multispectral orthorectification and image fusion processing; The conversion unit is used to use the ridge in the natural boundary of the farmland as the field distribution boundary, establish the field sample vector data through manual processing, convert the field sample vector data into raster data with the same size as the farmland area in the multispectral remote sensing data, and define the farmland area and non-farmland data in the raster data to obtain the label raster data; A sample data construction unit is used to superimpose the multispectral remote sensing data and the label grid data and then cut out the clip to obtain the sample data; The model building and training unit is used to build a deep learning model, input sample data into the deep learning model to train the deep learning model, calculate the loss function of the deep learning model, optimize the deep learning model using the Adam optimization method, and obtain a trained field boundary recognition model; The second collection unit is used to collect the current multispectral remote sensing data and input it into the trained field boundary recognition model to obtain the field boundary recognition result; The data processing and output unit is used to expand the boundary of the field by the dilation method, connect the gaps between the field boundaries by the closing operation method, and obtain the field boundary recognition result by increasing the continuity of the field boundary and removing the surface objects.

[0007] Based on the above technical solution, the present invention can also be improved as follows.

[0008] Furthermore, the multispectral remote sensing data and the label grid data are superimposed and then the slices are cut, including: setting the slice image size, retaining the set ratio of overlapping pixels for rectangular sliding window slicing, setting is the number of image slice rows, is the number of image slice columns, is the image length, is the image width, is the pixel length of the image slice, is the pixel length of the image slice, is the horizontal length of overlapping pixels of adjacent image slices, is the longitudinal length of overlapping pixels of adjacent image slices, and the image slice formula is: ; .

[0009] Furthermore, by adding reverse sliding window slices to preserve data integrity, is the number of rows in the final slice, is the number of columns of the final slice, is the floor function, is the number of image slice rows, is the number of image slice columns, then the calculation formula for expanded slices is: ; .

[0010] Furthermore, the deep learning model is trained to obtain a trained field boundary recognition model, including: the deep learning model performs feature extraction and feature fusion on the cropped input sample data through a single-layer convolution module and a double-layer convolution module to obtain a multi-scale feature map, uses a single-layer convolution module and a densely connected block of a dense layer to reuse inter-layer features, and restores the spatial resolution of the multi-scale feature map through jump connections and upsampling blocks, and finally fuses the multi-scale features through a custom fusion layer to output a field boundary recognition result to obtain a field boundary recognition model.

[0011] Furthermore, the multispectral remote sensing data includes spectral information of infrared band, red band, green band and blue band; the GF-7 image data is subjected to panchromatic and multispectral orthorectification respectively, and the true color image texture information is enhanced through image fusion.

[0012] Furthermore, the deep learning model includes a base layer, a dense connection block, an upsampling block, a fusion layer and an output layer; the base layer includes a first double-layer convolution module and a single-layer convolution module, which are used to extract basic semantic information and compress input image information; the dense connection block includes several repeatedly superimposed convolution layers; the upsampling block uses a transposed convolution method to gradually increase the size of the feature map; the fusion layer uses a convolution method to extract attention and then performs comprehensive processing with the feature map to complete the fusion of different features before outputting the image; the first output end of the base layer is connected to the input end of the dense connection block; the second output end of the base layer and the output end of the dense connection block are connected to the input end of the upsampling block; the output end of the upsampling block is connected to the output layer through the fusion layer.

[0013] Furthermore, calculating the loss function of the deep learning model includes: calculating the normal loss, texture loss and boundary loss, assigning weighted values ​​to the normal loss, texture loss and boundary loss and summing them to obtain the loss value of the loss function; using weighted binary cross entropy loss to calculate the difference between the prediction and the label to calculate the normal loss; smoothing the image through the convolution kernel to calculate the texture loss between the prediction and the true label; calculating the difference between the boundary of the predicted multi-scale feature image and the image of the label raster data to obtain the boundary loss.

[0014] Furthermore, the dilation algorithm in image morphological operations is used to expand the edges, and then the closing operation is used to connect the gaps between the field boundaries, and the erosion operation is used to restore the edge position information.

[0015] The beneficial effects of the present invention are as follows: the present invention aims at the problem that the field boundary recognition in the prior art is not accurate enough and inefficient, and provides a method for field boundary recognition based on GF-7 images, wherein the GF-7 images have high resolution and rich spectral information. The ridges in the natural boundaries of the farmland are used as the field distribution boundaries, and the multispectral remote sensing data and the label grid data are superimposed and then the clips are cut to obtain sample data, and the field boundaries are efficiently extracted in combination with the deep learning model. Specifically, the present invention uses convolutional neural networks and self-attention mechanisms to extract features from remote sensing images to improve the recognition accuracy of field boundaries, and can automatically recognize field boundaries, which is suitable for large-scale agricultural monitoring and precision agricultural management, and solves the shortcomings of traditional methods such as much manual intervention and poor timeliness, and effectively improves the segmentation accuracy and efficiency at the field level. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The schematic diagram of the method for fine extraction of farmland plots based on GF-7 images provided in Example 1 of the present invention; Figure 2 This is the principle block diagram of the deep learning model; Figure 3 This is a schematic diagram of a farmland plot fine extraction system based on GF-7 images provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0018] Example 1 As an example, Figure 1As shown, in order to solve the above technical problems, this embodiment provides a method for fine extraction of farmland plots based on GF-7 images, including: Collect multispectral remote sensing data and perform panchromatic and multispectral orthorectification and image fusion processing; The ridges in the natural boundaries of farmland are used as the boundaries of field distribution, and field sample vector data is established through manual processing. The field sample vector data is converted into raster data with the same size as the farmland area in the multispectral remote sensing data, and the farmland area and non-farmland data in the raster data are defined to obtain the label raster data. The multispectral remote sensing data and the label grid data are superimposed and the clips are cut to obtain the sample data; Construct a deep learning model, input sample data into the deep learning model to train the deep learning model, calculate the loss function of the deep learning model, and use the Adam (Adaptive Moment Estimation) optimization method to optimize the deep learning model to obtain a trained field boundary recognition model; Collect current multispectral remote sensing data and input it into the trained field boundary recognition model to obtain the field boundary recognition result; The field boundary is expanded by dilation method, and the gaps between field boundaries are connected by closing operation method. The field boundary recognition result is obtained by increasing the continuity of field boundaries and removing surface objects.

[0019] In the actual application process, in order to ensure data consistency, the size of the raster data is kept consistent with the multispectral remote sensing data of the GF-7 image, and the pixel farmland area and non-farmland area are defined.

[0020] Optionally, the multispectral remote sensing data and the label grid data are superimposed and then the slices are cut, including: setting the slice image size, retaining the set ratio of overlapping pixels for rectangular sliding window slicing, setting is the number of image slice rows, is the number of image slice columns, is the image length, is the image width, is the pixel length of the image slice, is the pixel length of the image slice, is the horizontal length of overlapping pixels of adjacent image slices, is the longitudinal length of overlapping pixels of adjacent image slices, and the image slice formula is: ; .

[0021] Optionally, data integrity can be preserved by adding reverse sliding window slices. is the number of rows in the final slice, is the number of columns of the final slice, is the floor function, is the number of image slice rows, is the number of image slice columns, then the calculation formula for expanded slices is: ; .

[0022] The spectral image and label grid data are superimposed and then cut into slices, which is conducive to efficient data access through deep learning models. The slice image size is set to 512*512*(4 / 1), the first 8 layers are the spectral information of the infrared band, red band, green band and blue band of the GF-7 optical image, and the 5th layer is the processed sample label data.

[0023] Optionally, the deep learning model is trained to obtain a trained field boundary recognition model, including: the deep learning model performs feature extraction and feature fusion on the cropped input sample data through a single-layer convolution module and a double-layer convolution module to obtain a multi-scale feature map, uses a single-layer convolution module and a densely connected block of a dense layer to reuse inter-layer features, and restores the spatial resolution of the multi-scale feature map through jump connections and upsampling blocks, and finally fuses the multi-scale features through a custom fusion layer to output a field boundary recognition result to obtain a field boundary recognition model.

[0024] Optionally, multispectral remote sensing data includes spectral information of infrared band, red band, green band and blue band; after panchromatic and multispectral orthorectification of GF-7 image data, the texture information of true color image is enhanced through image fusion. The highest spatial resolution is 0.65 meters, and the correction and image fusion of GF-7 image data can ensure data availability.

[0025] Optional, as attached Figure 2 As shown in the figure, the deep learning model includes a base layer, a dense connection block, an upsampling block, a fusion layer and an output layer; the base layer includes a first double-layer convolution module and a single-layer convolution module, which are used to extract basic semantic information and compress input image information; the dense connection block includes several repeatedly superimposed convolution layers; the upsampling block uses a transposed convolution method to gradually increase the size of the feature map; the fusion layer uses a convolution method to extract attention and then performs comprehensive processing with the feature map to complete the fusion of different features before outputting the image; the first output end of the base layer is connected to the input end of the dense connection block; the second output end of the base layer and the output end of the dense connection block are connected to the input end of the upsampling block; the output end of the upsampling block is connected to the output layer through the fusion layer.

[0026] Optionally, calculating the loss function of the deep learning model includes: calculating normal loss, texture loss and boundary loss, assigning weighted values ​​to the normal loss, texture loss and boundary loss and summing them to obtain the loss value of the loss function; using weighted binary cross entropy loss to calculate the difference between the prediction and the label to calculate the normal loss; smoothing the image through a convolution kernel, and calculating the texture loss between the prediction and the true label; calculating the difference between the boundary of the predicted multi-scale feature image and the image of the label raster data to obtain the boundary loss.

[0027] Optionally, the dilation algorithm in the image morphological operation is used to expand the edge, and then a closing operation is used to connect the gaps between the field boundaries, and an erosion operation is used to restore the edge position information.

[0028] By expanding the edge lines and using closing operations to connect the gaps between field boundaries, the output field boundaries are made more accurate. By increasing the continuity of the identified boundaries and removing small surface features such as ridges, the position accuracy is ensured, and then the erosion operation is used to restore the edge line position information, reducing erroneous boundary recognition and improving the accuracy of the results.

[0029] The present invention aims at the problem that the field boundary recognition in the prior art is not accurate enough and inefficient, and provides a method for field boundary recognition based on GF-7 images, which have high resolution and rich spectral information. The ridges in the natural boundaries of farmland are used as the field distribution boundaries, and the multispectral remote sensing data and the label grid data are superimposed and the clips are cut to obtain sample data, and the field boundary is efficiently extracted in combination with the deep learning model. Specifically, the present invention uses convolutional neural networks and self-attention mechanisms to extract features from remote sensing images to improve the recognition accuracy of field boundaries, and can automatically identify field boundaries, which is suitable for large-scale agricultural monitoring and precision agricultural management, and solves the shortcomings of traditional methods such as more manual intervention and poor timeliness, and effectively improves the segmentation accuracy and efficiency of field levels.

[0030] Example 2 Based on the same principle as the method shown in Example 1 of the present invention, as shown in the attached Figure 3 As shown, an embodiment of the present invention also provides a farmland plot fine extraction system based on GF-7 images, including a first collection unit, a conversion unit, a sample data construction unit, a model construction and training unit, a second collection unit, and a data processing and output unit; The first collection unit is used to collect multispectral remote sensing data and perform panchromatic and multispectral orthorectification and image fusion processing; The conversion unit is used to use the ridge in the natural boundary of the farmland as the field distribution boundary, establish the field sample vector data through manual processing, convert the field sample vector data into raster data with the same size as the farmland area in the multispectral remote sensing data, and define the farmland area and non-farmland data in the raster data to obtain the label raster data; A sample data construction unit is used to superimpose the multispectral remote sensing data and the label grid data and then cut out the clip to obtain the sample data; The model building and training unit is used to build a deep learning model, input sample data into the deep learning model to train the deep learning model, calculate the loss function of the deep learning model, optimize the deep learning model using the Adam optimization method, and obtain a trained field boundary recognition model; The second collection unit is used to collect the current multispectral remote sensing data and input it into the trained field boundary recognition model to obtain the field boundary recognition result; The data processing and output unit is used to expand the boundary of the field by the dilation method, connect the gaps between the field boundaries by the closing operation method, and obtain the field boundary recognition result by increasing the continuity of the field boundary and removing the surface objects.

[0031] Optionally, the multispectral remote sensing data and the label grid data are superimposed and then the slices are cut, including: setting the slice image size, retaining the set ratio of overlapping pixels for rectangular sliding window slicing, setting is the number of image slice rows, is the number of image slice columns, is the image length, is the image width, is the pixel length of the image slice, is the pixel length of the image slice, is the horizontal length of overlapping pixels of adjacent image slices, is the longitudinal length of overlapping pixels of adjacent image slices, and the image slice formula is: ; .

[0032] Optionally, data integrity can be preserved by adding reverse sliding window slices. is the number of rows in the final slice, is the number of columns of the final slice, is the floor function, is the number of image slice rows, is the number of image slice columns, then the calculation formula for expanded slices is: ; .

[0033] Optionally, the deep learning model is trained to obtain a trained field boundary recognition model, including: the deep learning model performs feature extraction and feature fusion on the cropped input sample data through a single-layer convolution module and a double-layer convolution module to obtain a multi-scale feature map, uses a single-layer convolution module and a densely connected block of a dense layer to reuse inter-layer features, and restores the spatial resolution of the multi-scale feature map through jump connections and upsampling blocks, and finally fuses the multi-scale features through a custom fusion layer to output a field boundary recognition result to obtain a field boundary recognition model.

[0034] Optionally, the multispectral remote sensing data includes spectral information of infrared bands, red bands, green bands and blue bands; after performing panchromatic and multispectral orthorectification on the GF-7 image data, the true color image texture information is enhanced through image fusion.

[0035] Optionally, the deep learning model includes a base layer, a densely connected block, an upsampling block, a fusion layer and an output layer; the base layer includes a first double-layer convolution module and a single-layer convolution module, which are used to extract basic semantic information and compress input image information; the densely connected block includes several repeatedly superimposed convolution layers; the upsampling block uses a transposed convolution method to gradually increase the size of the feature map; the fusion layer uses a convolution method to extract attention and then performs comprehensive processing with the feature map to complete the fusion of different features before outputting the image; the first output end of the base layer is connected to the input end of the densely connected block; the second output end of the base layer and the output end of the densely connected block are connected to the input end of the upsampling block; the output end of the upsampling block is connected to the output layer through the fusion layer.

[0036] Optionally, calculating the loss function of the deep learning model includes: calculating normal loss, texture loss and boundary loss, assigning weighted values ​​to the normal loss, texture loss and boundary loss and summing them to obtain the loss value of the loss function; using weighted binary cross entropy loss to calculate the difference between the prediction and the label to calculate the normal loss; smoothing the image through a convolution kernel, and calculating the texture loss between the prediction and the true label; calculating the difference between the boundary of the predicted multi-scale feature image and the image of the label raster data to obtain the boundary loss.

[0037] Optionally, the dilation algorithm in the image morphological operation is used to expand the edge, and then a closing operation is used to connect the gaps between the field boundaries, and an erosion operation is used to restore the edge position information.

[0038] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for fine extraction of farmland plots based on GF-7 images, characterized in that: include: Collect multispectral remote sensing data and perform panchromatic and multispectral orthorectification and image fusion processing; The ridges in the natural boundaries of farmland are used as the boundaries of field distribution, and field sample vector data is established through manual processing. The field sample vector data is converted into raster data with the same size as the farmland area in the multispectral remote sensing data, and the farmland area and non-farmland data in the raster data are defined to obtain the label raster data. The multispectral remote sensing data and the label grid data are superimposed and the clips are cut to obtain the sample data; Construct a deep learning model, input sample data into the deep learning model to train the deep learning model, calculate the loss function of the deep learning model, use the Adam optimization method to optimize the deep learning model, and obtain the trained field boundary recognition model; Collect current multispectral remote sensing data and input it into the trained field boundary recognition model to obtain the field boundary recognition result; The field boundary is expanded by dilation method, and the gaps between field boundaries are connected by closing operation method. The field boundary recognition result is obtained by increasing the continuity of field boundaries and removing surface objects.

2. The method for fine extraction of farmland plots based on GF-7 images according to claim 1, characterized in that: The multispectral remote sensing data and the label grid data are superimposed and then the slices are cut, including: setting the slice image size, retaining the set proportion of overlapping pixels for rectangular sliding window slicing, setting is the number of image slice rows, is the number of image slice columns, is the image length, is the image width, is the pixel length of the image slice, is the pixel length of the image slice, is the horizontal length of overlapping pixels of adjacent image slices, is the longitudinal length of overlapping pixels of adjacent image slices, and the image slice formula is: ； 。 3. The method for fine extraction of farmland plots based on GF-7 images according to claim 2 is characterized in that: By adding reverse sliding window slices to preserve data integrity, is the number of rows in the final slice, is the number of columns of the final slice, is the floor function, is the number of image slice rows, is the number of image slice columns, then the calculation formula for expanded slices is: ； 。 4. The method for fine extraction of farmland plots based on GF-7 images according to claim 1, characterized in that: The deep learning model is trained to obtain a trained field boundary recognition model, including: the deep learning model performs feature extraction and feature fusion on the cropped input sample data through a single-layer convolution module and a double-layer convolution module to obtain a multi-scale feature map, uses a single-layer convolution module and a densely connected block of a dense layer to reuse inter-layer features, and restores the spatial resolution of the multi-scale feature map through jump connections and upsampling blocks, and finally fuses the multi-scale features through a custom fusion layer to output a field boundary recognition result to obtain a field boundary recognition model.

5. The method for fine extraction of farmland plots based on GF-7 images according to claim 1, characterized in that: Multispectral remote sensing data includes spectral information of infrared band, red band, green band and blue band; after panchromatic and multispectral orthorectification of GF-7 image data, the texture information of true color image is enhanced through image fusion.

6. The method for fine extraction of farmland plots based on GF-7 images according to claim 1, characterized in that: The deep learning model includes a base layer, a dense connection block, an upsampling block, a fusion layer and an output layer; the base layer includes a first double-layer convolution module and a single-layer convolution module, which are used to extract basic semantic information and compress input image information; the dense connection block includes several repeatedly superimposed convolution layers; the upsampling block uses a transposed convolution method to gradually increase the size of the feature map; the fusion layer uses a convolution method to extract attention and then performs comprehensive processing with the feature map to complete the fusion of different features before outputting the image; the first output end of the base layer is connected to the input end of the dense connection block; the second output end of the base layer and the output end of the dense connection block are connected to the input end of the upsampling block; the output end of the upsampling block is connected to the output layer through a fusion layer.

7. The method for fine extraction of farmland plots based on GF-7 images according to claim 1, characterized in that: Calculating the loss function of the deep learning model includes: calculating the normal loss, texture loss and boundary loss, assigning weights to the normal loss, texture loss and boundary loss and summing them up to get the loss value of the loss function; using weighted binary cross entropy loss to calculate the difference between the prediction and the label to calculate the normal loss; smoothing the image through the convolution kernel, calculating the texture loss between the prediction and the true label; calculating the difference between the boundary of the predicted multi-scale feature image and the image of the label raster data to get the boundary loss.

8. The method for fine extraction of farmland plots based on GF-7 images according to claim 1, characterized in that: The dilation algorithm in image morphological operations is used to expand the edge lines, and then the closing operation is used to connect the gaps between the field boundaries, and the erosion operation is used to restore the edge line position information.

9. The farmland plot fine extraction system based on GF-7 image is characterized by: It includes a first collection unit, a conversion unit, a sample data construction unit, a model construction and training unit, a second collection unit, and a data processing and output unit; The first collection unit is used to collect multispectral remote sensing data and perform panchromatic and multispectral orthorectification and image fusion processing; The conversion unit is used to use the ridge in the natural boundary of the farmland as the field distribution boundary, establish the field sample vector data through manual processing, convert the field sample vector data into raster data with the same size as the farmland area in the multispectral remote sensing data, and define the farmland area and non-farmland data in the raster data to obtain the label raster data; A sample data construction unit is used to superimpose the multispectral remote sensing data and the label grid data and then cut out the clip to obtain the sample data; The model building and training unit is used to build a deep learning model, input sample data into the deep learning model to train the deep learning model, calculate the loss function of the deep learning model, optimize the deep learning model using the Adam optimization method, and obtain a trained field boundary recognition model; The second collection unit is used to collect the current multispectral remote sensing data and input it into the trained field boundary recognition model to obtain the field boundary recognition result; The data processing and output unit is used to expand the boundary of the field by the dilation method, connect the gaps between the field boundaries by the closing operation method, and obtain the field boundary recognition result by increasing the continuity of the field boundary and removing the surface objects.

Images