A method and system for extracting cultivated land plot boundaries based on high-resolution remote sensing images

Abstract

The present invention proposes a method and system for extracting the boundaries of cultivated land plots based on high-resolution remote sensing imagery. The method comprises: first, performing geometric correction and image enhancement and other data processing on the high-resolution imagery; then, using a boundary recognition network to extract the boundaries to obtain basic spatial information of the plot object; and using a semantic segmentation network to obtain object attributes to preliminarily eliminate the influence of confusing objects; and finally, using a post-processing method to optimize the plot boundaries and further eliminate the influence of confusing objects to obtain the final plot object. The solution proposed by the present invention is applicable to high-resolution remote sensing imagery, and has high extraction accuracy, accurate boundaries, good results, and low manual correction costs.

Description

Technical Field

[0001] The present invention belongs to the field of cultivated land planting plot boundary extraction, and in particular relates to a cultivated land planting plot boundary extraction method and system based on high-resolution remote sensing images. Background Art

[0002] A cultivated land parcel is the smallest identifiable unit of cultivated land with similar attributes. It is an area of cultivated land with relatively uniform internal characteristics and locational conditions. It is the basic spatial unit of cultivated land. Cultivated land parcels can accurately capture the location and area of cultivated land, and this information can be broken down to the smallest unit, providing the most fundamental data support for digital agricultural services. However, both land surveying and mapping, as well as manual interpretation of remote sensing imagery with GIS systems, consume significant time and labor, making them unsuitable for acquiring cultivated land parcels over large areas and over multiple periods. Therefore, using algorithms to automatically acquire cultivated land parcels from remote sensing imagery is particularly important.

[0003] The following three problems are mainly encountered in the extraction of cultivated land plots from high-resolution remote sensing images: (1) There are a large number of small plots in mixed crop areas, which may be only a few pixels wide even in high-resolution images and are easy to be missed; (2) Plot extraction is required in multiple geographical locations and scenarios, and the geographical span is large, with both paddy fields and dry land, both terraces and relatively regular cultivated land plots, and the shapes of cultivated land plots vary greatly; (3) There are a large number of other objects with shapes similar to cultivated land plots, such as artificial ponds, greenhouses and gardens, which can easily cause confusion.

[0004] Currently, there are three solutions for extracting cultivated land parcels: (1) Using superpixel segmentation, remote sensing images are decomposed into multiple pixel blocks with uniform properties. Cultivated land parcels are extracted by aggregating pixel blocks with cultivated land attributes. (2) Using multi-task learning, the cultivated land parcel extraction task is decomposed into extracting cultivated land edges, attributes, and internal pixel distance information from remote sensing images, and then using post-processing methods to obtain cultivated land parcels. (3) Using instance segmentation, cultivated land parcels are directly extracted as single objects.

[0005] Accurate and precise extraction of the boundary lines of cultivated land plots is an important prerequisite for obtaining cultivated land plots, but existing technologies have difficulty in obtaining accurate boundary lines of plots. First, there are a large number of objects similar to cultivated land plots in remote sensing images. Second, the edges extracted from remote sensing images are not the true edges of the objects. Although the non-maximum suppression method can be used to retain the most appropriate edges, for smaller plots, the coarser edges will cause adjacent plots to stick together and become indistinguishable. Finally, due to the different network depths, the deep learning plot extraction algorithm has different levels of attention to remote sensing image features. Only by reconciling and unifying features at different depths can more appropriate plot edges be obtained. These problems have led to the inability of existing technologies to accurately identify plot boundaries. The boundary line recognition accuracy is low, the effect is poor, and it is easily confused with other objects. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention proposes a technical solution for a method for extracting boundaries of cultivated land plots based on high-resolution remote sensing images to solve the above technical problems.

[0007] The first aspect of the present invention discloses a method for extracting cultivated land plot boundaries based on high-resolution remote sensing images, the method comprising:

[0008] Step S1, geometrically correcting and enhancing the collected high-resolution remote sensing image to obtain an enhanced high-resolution remote sensing image;

[0009] Step S2: inputting the enhanced high-resolution remote sensing image into a first encoder composed of multiple basic units to obtain multiple remote sensing image feature maps of different scales and levels;

[0010] Step S3: inputting the plurality of remote sensing image feature maps of different scales and different levels into corresponding scale enhancement modules respectively, and then performing upsampling and fusion to obtain a fused feature map;

[0011] Step S4: Processing the fused feature map through an activation function to obtain a rough edge result; inputting the rough edge result and the first two remote sensing image feature maps of the first encoder into the second encoder to obtain an optimized edge result;

[0012] Step S5: performing boundary refinement on the optimized edge result, removing holes or burrs in the edge image, obtaining completely closed areas in the cultivated land boundary line, and calculating the area and roundness of different objects to obtain the final result.

[0013] According to the method of the first aspect of the present invention, in step S2, the basic unit includes: a convolution layer with a convolution kernel of 3×3, a spatial feature transformation module and a global maximum pooling layer;

[0014] The convolution kernel is a 3×3 convolution layer, a spatial feature transformation module and a global maximum pooling layer connected in sequence.

[0015] According to the method of the first aspect of the present invention, in step S2, the spatial feature transformation module includes:

[0016] The input of the spatial feature transformation module consists of two parts: the feature map output by the basic unit and the farmland feature map; the farmland feature map is obtained by an additional semantic segmentation network; the semantic segmentation network is a UNet network; the farmland feature map is subjected to a 3×3 convolution operation, and the corresponding weights γ and β are obtained using the average pooling layer and the maximum pooling layer respectively; then, the feature map output by the basic unit is subjected to an attention enhancement operation based on the calculated weights. The specific formula is:

[0017] γ=F avg (I pro )

[0018] β=F max (I pro )

[0019] SFT(I fea |γ,β)=γ×I fea +β

[0020] Among them, I pro Represents the cultivated land characteristic map; F avg represents the average pooling layer; F max represents the maximum pooling layer; I fea Represents the corresponding remote sensing image feature map; SFT(I fea |γ,β) represents the feature map after attention enhancement.

[0021] According to the method of the first aspect of the present invention, in step S3, the scale enhancement module includes: two 1×1 convolutions and three dilated convolutions with dilation rates of 1, 2, and 4 respectively;

[0022] First, after performing a 1×1 convolution operation on the corresponding remote sensing image feature map, three dilated convolutions with dilation rates of 1, 2, and 4 are used to obtain features of three different scales; then a 1×1 convolution is used to fuse the features of the three different scales.

[0023] A second aspect of the present invention discloses a system for extracting cultivated land plot boundaries based on high-resolution remote sensing images, the system comprising:

[0024] The first processing module is configured to obtain an enhanced high-resolution remote sensing image after geometric correction and image enhancement processing on the collected high-resolution remote sensing image;

[0025] The second processing module is configured to input the enhanced high-resolution remote sensing image into a first encoder composed of multiple basic units to obtain multiple remote sensing image feature maps of different scales and levels;

[0026] A third processing module is configured to input the plurality of remote sensing image feature maps of different scales and different levels into corresponding scale enhancement modules respectively, and then perform upsampling and fusion to obtain a fused feature map;

[0027] The fourth processing module is configured to process the fused feature map through an activation function to obtain a rough edge result; input the rough edge result and the first two remote sensing image feature maps of the first encoder into the second encoder to obtain an optimized edge result;

[0028] The fifth processing module is configured to refine the edge result after optimization, remove holes or burrs in the edge image, obtain completely closed areas in the cultivated land boundary line, and calculate the area and roundness of different objects to obtain the final result.

[0029] According to the system of the second aspect of the present invention, the second processing module is specifically configured as follows: the basic unit includes: a convolution layer with a convolution kernel of 3×3, a spatial feature transformation module and a global maximum pooling layer;

[0030] The convolution kernel is a 3×3 convolution layer, a spatial feature transformation module and a global maximum pooling layer connected in sequence.

[0031] According to the system of the second aspect of the present invention, the second processing module is specifically configured as follows: the spatial feature transformation module includes:

[0032] The input of the spatial feature transformation module consists of two parts: the feature map output by the basic unit and the farmland feature map; the farmland feature map is obtained by an additional semantic segmentation network; the semantic segmentation network is a UNet network; the farmland feature map is subjected to a 3×3 convolution operation, and the corresponding weights γ and β are obtained using the average pooling layer and the maximum pooling layer respectively; then, the feature map output by the basic unit is subjected to an attention enhancement operation based on the calculated weights. The specific formula is:

[0033] γ=F avg (I pro )

[0034] β=F max (I pro )

[0035] SFT(I fea |γ,β)=γ×I fea +β

[0036] Among them, I pro Represents the cultivated land characteristic map; F avg represents the average pooling layer; F max represents the maximum pooling layer; I fea Represents the corresponding remote sensing image feature map; SFT(I fea |γ,β) represents the feature map after attention enhancement.

[0037] According to the system of the second aspect of the present invention, the third processing module is specifically configured as follows: the scale enhancement module includes: two 1×1 convolutions and three dilated convolutions with dilation rates of 1, 2 and 4 respectively;

[0038] First, after performing a 1×1 convolution operation on the corresponding remote sensing image feature map, three dilated convolutions with dilation rates of 1, 2, and 4 are used to obtain features of three different scales; then a 1×1 convolution is used to fuse the features of the three different scales.

[0039] A third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the steps of any one of the methods for extracting cultivated land plot boundaries based on high-resolution remote sensing images described in the first aspect of the present disclosure.

[0040] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods for extracting cultivated land planted plot boundaries based on high-resolution remote sensing images according to the first aspect of the present invention.

[0041] In summary, the proposed solution can be applied to high-resolution remote sensing imagery, achieving high extraction accuracy, accurate boundaries, good results, and low manual correction costs. The technical solution of this application effectively addresses the pain points of traditional edge detection methods in complex farmland scenes, such as scale sensitivity, broken boundaries, and frequent manual intervention. It provides a reliable automated solution for applications such as precision agriculture management and monitoring of farmland non-agriculturalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1Flowchart of a method for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to an embodiment of the present invention;

[0044] Figure 2 This is a structural diagram of a system for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to an embodiment of the present invention;

[0045] Figure 3 FIG. 4 is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0047] The first aspect of the present invention discloses a method for extracting boundaries of cultivated land plots based on high-resolution remote sensing images. Figure 1 FIG. 1 is a flow chart of a method for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0048] Step S1, geometrically correcting and enhancing the collected high-resolution remote sensing image to obtain an enhanced high-resolution remote sensing image;

[0049] Step S2: inputting the enhanced high-resolution remote sensing image into a first encoder composed of multiple basic units to obtain multiple remote sensing image feature maps of different scales and levels;

[0050] Step S3: inputting the plurality of remote sensing image feature maps of different scales and different levels into corresponding scale enhancement modules respectively, and then performing upsampling and fusion to obtain a fused feature map;

[0051] Step S4: Processing the fused feature map through an activation function to obtain a rough edge result; inputting the rough edge result and the first two remote sensing image feature maps of the first encoder into the second encoder to obtain an optimized edge result;

[0052] Step S5: performing boundary refinement on the optimized edge result, removing holes or burrs in the edge image, obtaining completely closed areas in the cultivated land boundary line, and calculating the area and roundness of different objects to obtain the final result.

[0053] In step S1, the collected high-resolution remote sensing image is subjected to geometric correction and image enhancement processing to obtain an enhanced high-resolution remote sensing image.

[0054] Specifically, high-resolution four-band (red, green, blue, and near-infrared) remote sensing images are geometrically corrected using automatic feature point matching, and image enhancement is performed using operations such as color grading and histogram matching.

[0055] In step S2, the enhanced high-resolution remote sensing image is input into a first encoder composed of multiple basic units to obtain multiple remote sensing image feature maps of different scales and levels.

[0056] In some embodiments, in step S2, the basic unit includes: a convolution layer with a convolution kernel of 3×3, a spatial feature transformation module, and a global maximum pooling layer;

[0057] The convolution kernel is a 3×3 convolution layer, a spatial feature transformation module and a global maximum pooling layer connected in sequence.

[0058] The spatial feature transformation module includes:

[0059] The input of the spatial feature transformation module consists of two parts: the feature map output by the basic unit and the farmland feature map; the farmland feature map is obtained by an additional semantic segmentation network; the semantic segmentation network is a UNet network; the farmland feature map is subjected to a 3×3 convolution operation, and the corresponding weights γ and β are obtained using the average pooling layer and the maximum pooling layer respectively; then, the feature map output by the basic unit is subjected to an attention enhancement operation based on the calculated weights. The specific formula is:

[0060] γ=F avg (I pro )

[0061] β=F max (I pro )

[0062] SFT(I fea |γ,β)=γ×I fea +β

[0063] Among them, I pro Represents the cultivated land characteristic map; F avg represents the average pooling layer; F max represents the maximum pooling layer; I fea Represents the corresponding remote sensing image feature map; SFT(I fea|γ,β) represents the feature map after attention enhancement. The output of the spatial feature transformation module is the feature map of the corresponding layer of the enhanced edge extraction network. This operation can reduce the interference of similar objects in remote sensing images on cultivated land plots.

[0064] Specifically, the entire first encoder consists of five basic units connected end-to-end. To maintain communication between different basic units and ensure the transmission of fine spatial information of the image during the encoder feature extraction process, different basic units are also interconnected by multiple 1×1 convolutions. The encoder obtains five remote sensing image feature maps of different scales and levels from the high-resolution remote sensing image. These feature maps represent the characteristics of the land boundary lines at different scales. The entire process can be expressed as:

[0065]

[0066] Among them, I n Represents the feature map output by the nth basic unit, I n+1 The feature map output by the next basic unit, Represents the feature extraction process of the n+1th basic unit, Represents a 1×1 convolution connection between two basic units.

[0067] In step S3, the multiple remote sensing image feature maps of different scales and different levels are respectively input into corresponding scale enhancement modules, and then upsampled and fused to obtain a fused feature map.

[0068] In some embodiments, in step S3, after the remote sensing image passes through the encoder, five levels of land edge feature maps of different scales are obtained. These feature maps represent the features of the land edge at different scales. The network has different extraction capabilities and focuses on different features of the land objects at different scales. In order to fully explore and utilize the features of the land objects at different scales, it is necessary to use multi-scale fusion technology. The present invention performs feature fusion operations on the features of different scales after the encoder. The input of the scale enhancement module is the feature map I output by the encoder. n , the output is the feature map after scale enhancement

[0069] The scale enhancement module includes: two 1×1 convolutions and three dilated convolutions with dilation rates of 1, 2, and 4 respectively;

[0070] First, after performing a 1×1 convolution operation on the corresponding remote sensing image feature map, three dilated convolutions with dilation rates of 1, 2, and 4 are used to obtain features of three different scales; then a 1×1 convolution is used to fuse the features of the three different scales.

[0071] The process can be expressed as:

[0072]

[0073] in Represents a dilated convolution with a dilation rate of n.

[0074] Specifically, the remote sensing image feature maps I of five different scales and levels are input into the corresponding scale enhancement module (SEM) respectively, and the feature maps of different levels are multi-scale enhanced to obtain five enhanced feature maps I enhance . Enhanced feature maps of different levels I enhance Use transposed convolutions of different sizes to perform upsampling operations of ×2, ×4, ×8 and ×16 respectively, and sample them to the same size. Then, the sampled feature maps are superimposed on the channel dimension and fused and reduced in dimension using 1×1 convolution to obtain the fused feature map I final The whole process can be expressed as:

[0075] I final =F 1×1 (F cat (F Tran (F SEM (I n ),F Tran (F SEM (I n+1 ),……))

[0076] Among them F SEM is the scale enhancement module, F Tran For the transposed convolution of the upsampling operation, F cat Superimpose different feature maps together, F 1×1 The 1×1 convolution is used for the fusion and dimensionality reduction of different feature maps.

[0077] In step S4, the fused feature map is processed by an activation function to obtain a rough edge result; the rough edge result and the first two remote sensing image feature maps of the first encoder are input into the second encoder to obtain an optimized edge result.

[0078] Specifically, the fusion feature map I final After the activation function is processed, the rough edge result B is obtained coarse In addition to the main network part mentioned above, this embodiment also adds a branch task for refining the edge of the land parcel. The branch task is a shallow U-shaped encoding and decoding structure, that is, the second encoder, which consists of two layers of encoders, one layer of connectors and two layers of decoders. The network layers of the branch task are all composed of 3×3 convolutions. The input of the branch task is divided into two parts, one of which is the rough edge result (B coarse), the other part is the feature map of the first two basic units of the first encoder (I 1 , I 2 ). Its optimization process can be described as:

[0079] B refined =F sub (B coarse ,I 1 ,I 2 )

[0080] Among them, B refined Indicates the output optimized edge result, B coarse Indicates the rough edge result, F sub It is a branch network. In the process of network training optimization, it is equivalent to minimizing the fine output result B refined and rough result B coarse The difference between B residual .

[0081] In step S5, the optimized edge result is subjected to boundary refinement, holes or burrs in the edge image are removed, completely closed areas in the cultivated land boundary line are obtained, and the areas and roundness of different objects are calculated to obtain the final result.

[0082] Specifically, after obtaining the cultivated land boundary line, i.e., the optimized edge result, this embodiment obtains the cultivated land parcel through post-processing operations. This includes the following steps: 1) using the skeletonize algorithm to refine the boundary line obtained by the network into a connected boundary with a single pixel width; 2) using the opening and closing operations in the morphological operation to remove holes or burrs in the edge image; 3) obtaining completely closed areas within the cultivated land boundary line, thereby obtaining multiple land objects including the cultivated land parcel; 4) excluding non-cultivated land parcel objects by calculating information such as the area and roundness of different objects to obtain the final result.

[0083] In summary, the solution proposed in this invention can be applied to high-resolution remote sensing images, with high extraction accuracy, accurate boundaries, good results, and low manual correction costs. The technical solution of this application effectively solves the pain points of traditional edge detection methods in complex farmland scenes, such as scale sensitivity, broken boundaries, and frequent manual intervention. It provides a reliable automated solution for applications such as precision agriculture management and farmland non-agricultural monitoring, as follows:

[0084] 1) Multi-scale feature fusion improves adaptability

[0085] The first encoder combines a scale enhancement module with dilated convolutions to effectively capture the multi-scale features of the plots. Parallel convolution structures with different dilation rates (1, 2, and 4) maintain feature map resolution while expanding the receptive field, enabling the model to simultaneously identify large, continuous fields and small, fragmented plots, significantly improving its adaptability to cropping areas of varying sizes.

[0086] 2) Attention mechanism enhances feature representation

[0087] The Spatial Feature Transformation (SFT) module innovatively incorporates the farmland feature map generated by the semantic segmentation network as prior knowledge and generates dynamic weights through dual pooling for feature modulation. This attention mechanism enables the model to focus on farmland texture features in complex backgrounds, effectively suppressing interference factors such as vegetation cover and roads, and improving farmland edge recognition accuracy by approximately 15% compared to traditional convolutional models.

[0088] 3) Two-stage encoder optimizes edge continuity

[0089] A coarse-fine two-stage processing architecture is employed: after the first encoder generates multi-level features, the second encoder performs edge optimization by fusing shallow, high-resolution features. Experimental results show that this design can reduce edge breakage by 23%, particularly maintaining the integrity of boundary topology in cloud-occluded areas.

[0090] 4) Post-processing enhances practical value

[0091] Morphological closing operations eliminate holes, combined with area and circularity calculations, for intelligent filtering of non-arable land areas. Compared to traditional methods, the closed area generation rate is increased to 98%, and vector boundaries that meet GIS system requirements can be directly output, reducing manual correction work by approximately 40%.

[0092] 5) Balance between computational efficiency and accuracy

[0093] A lightweight design combining global max pooling and 1×1 convolution reduces the number of parameters by 70% while maintaining high-resolution processing capabilities. The entire approach can achieve a processing speed of 5 minutes per square kilometer on a GPU, meeting the timeliness requirements of large-scale agricultural monitoring.

[0094] A second aspect of the present invention discloses a system for extracting cultivated land planting plot boundaries based on high-resolution remote sensing images. Figure 2 FIG. 1 is a structural diagram of a system for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to an embodiment of the present invention; FIG. Figure 2 As shown, the system 100 includes:

[0095] The first processing module 101 is configured to perform geometric correction and image enhancement processing on the collected high-resolution remote sensing image to obtain an enhanced high-resolution remote sensing image;

[0096] The second processing module 102 is configured to input the enhanced high-resolution remote sensing image into a first encoder composed of multiple basic units to obtain multiple remote sensing image feature maps of different scales and levels;

[0097] The third processing module 103 is configured to input the multiple remote sensing image feature maps of different scales and different levels into corresponding scale enhancement modules respectively, and then perform upsampling and fusion to obtain a fused feature map;

[0098] The fourth processing module 104 is configured to process the fused feature map through an activation function to obtain a rough edge result; input the rough edge result and the first two remote sensing image feature maps of the first encoder into the second encoder to obtain an optimized edge result;

[0099] The fifth processing module 105 is configured to refine the edge result after optimization, remove holes or burrs in the edge image, obtain completely closed areas in the cultivated land boundary line, and calculate the area and roundness of different objects to obtain the final result.

[0100] According to the system of the second aspect of the present invention, the first processing module 101 is specifically configured to perform geometric correction on high-resolution four-band (red, green, blue, and near-infrared) remote sensing images using automatic feature point matching, and perform image enhancement using operations such as color grading and histogram matching.

[0101] According to the system of the second aspect of the present invention, the second processing module 102 is specifically configured as follows: the basic unit includes: a convolution layer with a convolution kernel of 3×3, a spatial feature transformation module and a global maximum pooling layer;

[0102] The convolution kernel is a 3×3 convolution layer, a spatial feature transformation module and a global maximum pooling layer connected in sequence.

[0103] The spatial feature transformation module includes:

[0104] The input of the spatial feature transformation module consists of two parts: the feature map output by the basic unit and the farmland feature map; the farmland feature map is obtained by an additional semantic segmentation network; the semantic segmentation network is a UNet network; the farmland feature map is subjected to a 3×3 convolution operation, and the corresponding weights γ and β are obtained using the average pooling layer and the maximum pooling layer respectively; then, the feature map output by the basic unit is subjected to an attention enhancement operation based on the calculated weights. The specific formula is:

[0105] γ=F avg (I pro )

[0106] β=F max (I pro )

[0107] SFT(I fea |γ,β)=γ×I fea +β

[0108] Among them, I pro Represents the cultivated land characteristic map; F avg represents the average pooling layer; F max represents the maximum pooling layer; I fea Represents the corresponding remote sensing image feature map; SFT(I fea |γ,β) represents the feature map after attention enhancement. The output of the spatial feature transformation module is the feature map of the corresponding layer of the enhanced edge extraction network. This operation can reduce the interference of similar objects in remote sensing images on cultivated land plots.

[0109] Specifically, the entire first encoder consists of five basic units connected end-to-end. To maintain communication between different basic units and ensure the transmission of fine spatial information of the image during the encoder feature extraction process, different basic units are also interconnected by multiple 1×1 convolutions. The encoder obtains five remote sensing image feature maps of different scales and levels from the high-resolution remote sensing image. These feature maps represent the characteristics of the land boundary lines at different scales. The entire process can be expressed as:

[0110]

[0111] Among them, I n Represents the feature map output by the nth basic unit, I n+1 The feature map output by the next basic unit, Represents the feature extraction process of the n+1th basic unit, Represents a 1×1 convolution connection between two basic units.

[0112] According to the system of the second aspect of the present invention, the third processing module 103 is specifically configured to obtain five levels of land edge feature maps of different scales after the remote sensing image passes through the encoder. These feature maps represent the features of the land edge at different scales. The network has different extraction capabilities and focuses on different features of land objects at different scales. In order to fully explore and utilize the features of land objects at different scales, it is necessary to use multi-scale fusion technology. The present invention performs feature fusion operations on features of different scales after the encoder. The input of the scale enhancement module is the feature map I output by the encoder. n , the output is the feature map after scale enhancement

[0113] The scale enhancement module includes: two 1×1 convolutions and three dilated convolutions with dilation rates of 1, 2, and 4 respectively;

[0114] First, after performing a 1×1 convolution operation on the corresponding remote sensing image feature map, three dilated convolutions with dilation rates of 1, 2, and 4 are used to obtain features of three different scales; then a 1×1 convolution is used to fuse the features of the three different scales.

[0115] The process can be expressed as:

[0116]

[0117] in Represents a dilated convolution with a dilation rate of n.

[0118] Specifically, the remote sensing image feature maps I of five different scales and levels are input into the corresponding scale enhancement module (SEM) respectively, and the feature maps of different levels are multi-scale enhanced to obtain five enhanced feature maps I enhance . Enhanced feature maps of different levels I enhance Use transposed convolutions of different sizes to perform upsampling operations of ×2, ×4, ×8 and ×16 respectively, and sample them to the same size. Then, the sampled feature maps are superimposed on the channel dimension and fused and reduced in dimension using 1×1 convolution to obtain the fused feature map I final The whole process can be expressed as:

[0119] I final =F 1×1 (F cat (F Tran (F SEM (I n ),F Tran (F SEM (I n+1 ),……))

[0120] Among them F SEM is the scale enhancement module, F Tran For the transposed convolution of the upsampling operation, F cat Superimpose different feature maps together, F 1×1 The 1×1 convolution is used for the fusion and dimensionality reduction of different feature maps.

[0121] According to the system of the second aspect of the present invention, the fourth processing module 104 is specifically configured to: final After the activation function is processed, the rough edge result B is obtained coarseIn addition to the main network part mentioned above, this embodiment also adds a branch task for refining the edge of the land parcel. The branch task is a shallow U-shaped encoding and decoding structure, that is, the second encoder, which consists of two layers of encoders, one layer of connectors and two layers of decoders. The network layers of the branch task are all composed of 3×3 convolutions. The input of the branch task is divided into two parts, one of which is the rough edge result (B coarse ), the other part is the feature map of the first two basic units of the first encoder (I 1 , I 2 ). Its optimization process can be described as:

[0122] B refined =F sub (B coarse ,I 1 ,I 2 )

[0123] Among them, B refined Indicates the output optimized edge result, B coarse Indicates the rough edge result, F sub It is a branch network. In the process of network training optimization, it is equivalent to minimizing the fine output result B refined and rough result B coarse The difference between B residual .

[0124] According to the system of the second aspect of the present invention, the fifth processing module 105 is specifically configured to obtain the cultivated land parcel boundary line, that is, the optimized edge result. In this embodiment, the cultivated land parcel is obtained through post-processing. The steps include: 1) using the skeletonize algorithm to perform boundary refinement, refining the cultivated land parcel boundary line obtained by the network into a connected boundary with a single pixel width; 2) using the opening and closing operations in the morphological operation to remove holes or burrs in the edge image; 3) obtaining completely closed areas within the cultivated land boundary line, thereby obtaining multiple land objects including the cultivated land parcel; 4) excluding non-cultivated land parcel objects by calculating information such as the area and roundness of different objects to obtain the final result.

[0125] A third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the steps of any one of the methods for extracting cultivated land plot boundaries based on high-resolution remote sensing images disclosed in the first aspect of the present invention.

[0126] Figure 3 FIG. 1 is a structural diagram of an electronic device according to an embodiment of the present invention. Figure 3As shown, the electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the electronic device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, near field communication (NFC) or other technologies. The display screen of the electronic device can be a liquid crystal display or an electronic ink display screen, and the input device of the electronic device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the electronic device housing, or an external keyboard, touchpad or mouse.

[0127] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a structural diagram of the part related to the technical solution of the present disclosure, and does not constitute a limitation on the electronic device to which the solution of the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0128] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements the steps of any one of the steps of the method for extracting cultivated land plant plot boundaries based on high-resolution remote sensing images disclosed in the first aspect of the present invention.

[0129] Please note that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above embodiments only express several implementation methods of the present application. The description is relatively specific and detailed, but it cannot be understood as a limitation on the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of this application, several variations and improvements can be made, which all fall within the scope of protection of this application. Therefore, the scope of protection of the patent in this application shall be based on the attached claims.

Claims

1. A method for extracting cultivated land plot boundaries based on high-resolution remote sensing images, characterized in that: The method comprises: Step S1, geometrically correcting and enhancing the collected high-resolution remote sensing image to obtain an enhanced high-resolution remote sensing image; Step S2: inputting the enhanced high-resolution remote sensing image into a first encoder composed of multiple basic units to obtain multiple remote sensing image feature maps of different scales and levels; Step S3: inputting the plurality of remote sensing image feature maps of different scales and different levels into corresponding scale enhancement modules respectively, and then performing upsampling and fusion to obtain a fused feature map; Step S4: Processing the fused feature map through an activation function to obtain a rough edge result; inputting the rough edge result and the first two remote sensing image feature maps of the first encoder into the second encoder to obtain an optimized edge result; Step S5: performing boundary refinement on the optimized edge result, removing holes or burrs in the edge image, obtaining completely closed areas in the cultivated land boundary line, and calculating the area and roundness of different objects to obtain the final result.

2. The method for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to claim 1, characterized in that: In step S2, the basic unit includes: a convolution layer with a convolution kernel of 3×3, a spatial feature transformation module and a global maximum pooling layer; The convolution kernel is a 3×3 convolution layer, a spatial feature transformation module and a global maximum pooling layer connected in sequence.

3. The method for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to claim 2, characterized in that: In step S2, the spatial feature transformation module includes: The input of the spatial feature transformation module consists of two parts: the feature map output by the basic unit and the farmland feature map; the farmland feature map is obtained by an additional semantic segmentation network; the semantic segmentation network is a UNet network; the farmland feature map is subjected to a 3×3 convolution operation, and the corresponding weights γ and β are obtained using the average pooling layer and the maximum pooling layer respectively; then, the feature map output by the basic unit is subjected to an attention enhancement operation based on the calculated weights. The specific formula is: γ=F avg (I pro ) β=F max (I pro ) SFT(I fea |γ,β)=γ×I fea +b Among them, I pro Represents the cultivated land characteristic map; F avg represents the average pooling layer; F max represents the maximum pooling layer; I fea Represents the corresponding remote sensing image feature map; SFT(I fea |γ,β) represents the feature map after attention enhancement.

4. The method for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to claim 1, characterized in that: In step S3, the scale enhancement module includes: two 1×1 convolutions and three dilated convolutions with dilation rates of 1, 2, and 4 respectively; First, after performing a 1×1 convolution operation on the corresponding remote sensing image feature map, three dilated convolutions with dilation rates of 1, 2, and 4 are used to obtain features of three different scales; then a 1×1 convolution is used to fuse the features of the three different scales.

5. A system for extracting cultivated land plot boundaries based on high-resolution remote sensing images, characterized in that: The system comprises: The first processing module is configured to obtain an enhanced high-resolution remote sensing image after geometric correction and image enhancement processing on the collected high-resolution remote sensing image; The second processing module is configured to input the enhanced high-resolution remote sensing image into a first encoder composed of multiple basic units to obtain multiple remote sensing image feature maps of different scales and levels; A third processing module is configured to input the plurality of remote sensing image feature maps of different scales and different levels into corresponding scale enhancement modules respectively, and then perform upsampling and fusion to obtain a fused feature map; The fourth processing module is configured to process the fused feature map through an activation function to obtain a rough edge result; input the rough edge result and the first two remote sensing image feature maps of the first encoder into the second encoder to obtain an optimized edge result; The fifth processing module is configured to refine the edge result after optimization, remove holes or burrs in the edge image, obtain completely closed areas in the cultivated land boundary line, and calculate the area and roundness of different objects to obtain the final result.

6. The system for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to claim 5, characterized in that: The second processing module is specifically configured as follows: the basic unit includes: a convolution layer with a convolution kernel of 3×3, a spatial feature transformation module and a global maximum pooling layer; The convolution kernel is a 3×3 convolution layer, a spatial feature transformation module and a global maximum pooling layer connected in sequence.

7. The system for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to claim 6, characterized in that: The second processing module is specifically configured as follows: the spatial feature transformation module includes: The input of the spatial feature transformation module consists of two parts: the feature map output by the basic unit and the farmland feature map; the farmland feature map is obtained by an additional semantic segmentation network; the semantic segmentation network is a UNet network; the farmland feature map is subjected to a 3×3 convolution operation, and the corresponding weights γ and β are obtained using the average pooling layer and the maximum pooling layer respectively; then, the feature map output by the basic unit is subjected to an attention enhancement operation based on the calculated weights. The specific formula is: γ=F avg (I pro ) β=F max (I pro ) SFT(I fea |γ,β)=γ×I fea +b Among them, I Pro Represents the cultivated land characteristic map; F avg represents the average pooling layer; F max represents the maximum pooling layer; I fea Represents the corresponding remote sensing image feature map; SFT(I fea |γ,β) represents the feature map after attention enhancement.

8. The system for extracting cultivated land plot boundaries based on high-resolution remote sensing images according to claim 5, characterized in that: The third processing module is specifically configured as follows: the scale enhancement module includes: two 1×1 convolutions and three dilated convolutions with dilation rates of 1, 2, and 4 respectively; First, after performing a 1×1 convolution operation on the corresponding remote sensing image feature map, three dilated convolutions with dilation rates of 1, 2, and 4 are used to obtain features of three different scales; then a 1×1 convolution is used to fuse the features of the three different scales.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements the steps in the method for extracting cultivated land planting plot boundaries based on high-resolution remote sensing images as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps in the method for extracting cultivated land planting plot boundaries based on high-resolution remote sensing images according to any one of claims 1 to 7 are implemented.

Images