Remote sensing data-based plot boundary identification method and system

Abstract

The invention discloses a remote sensing data-based plot boundary identification method and system, and relates to the technical field of plot boundary identification, optical and SAR data are complementarily fused, the limitation of a single data source in an extreme illumination or texture fuzzy scene is broken through, the boundary detection rate in a complex environment is significantly enhanced, and the fusion perception capability is effectively improved; self-supervised pre-training is adopted, mass unlabeled data is utilized to learn universal edge features, manual labeling requirements are reduced, the method is particularly suitable for regions with deficient labeling resources, and labeling dependency is reduced; in addition, a gradient direction constraint is adopted to force the model to learn an edge trend conforming to a physical rule, the problems of fracture and burr of a traditional segmentation algorithm in a weak boundary region such as a farmland in a growing period are solved, and geometric continuity is optimized.

Description

Technical Field

[0001] This invention relates to the field of land parcel boundary identification technology, and in particular to a method and system for land parcel boundary identification based on remote sensing data. Background Technology

[0002] In the fields of precision agriculture and land survey, remote sensing land boundary identification technology is gradually replacing traditional manual delineation; especially during the crop growth period, such as the rice heading stage and wheat grain filling stage, farmland exhibits a wide range of homogeneous spectral characteristics, and the boundaries between plots are stuck due to high crop density and blurred furrows.

[0003] Current mainstream solutions rely on high-resolution optical satellite imagery, such as Sentinel-2 / GF-7, combined with deep learning semantic segmentation models for plot extraction. While these techniques improve automation and enhance contextual awareness by expanding the receptive field (e.g., dilated convolution), they still suffer from spectral confusion and edge weakening issues. Adjacent plots often have highly similar crop canopy spectra, making it easy for models to misclassify multiple plots as a single connected region, especially in large-scale planting areas in plains. Crop cover causes the disappearance of field ridge texture, and edge detection algorithms, such as RCF, are insufficiently responsive to low-contrast boundaries, resulting in fragmented or jagged images. Some solutions attempt to introduce superpixel segmentation, such as SLIC pre-cutting image blocks, but this over-reliance on color consistency exacerbates over-segmentation during crop color change when yellow and green hues intermingle. Therefore, a plot boundary recognition solution based on remote sensing data is urgently needed to address these problems. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] This invention provides a method and system for identifying land parcel boundaries based on remote sensing data. This solves the problems of existing methods relying on single optical data and fixed models, which are difficult to overcome due to farmland adhesion during the growing season, shadow interference, and differences in terrain generalization, and rely on manual correction.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a method for identifying land parcel boundaries based on remote sensing data, comprising,

[0008] Step S1: Acquire optical remote sensing images and synthetic aperture radar images;

[0009] Step S2: The two types of images are fused by an adaptive weight gating unit to generate a fused feature map. Its dynamic weight is jointly determined by the near-infrared band reflectivity and radar scattering intensity.

[0010] Step S3: Input the fused feature map into the self-supervised pre-trained boundary recognition model and output the initial boundary probability map;

[0011] Step S4: Apply gradient direction consistency constraints to the initial boundary probability map and calculate the gradient direction cosine similarity loss between the predicted boundary and the true boundary.

[0012] Step S5: Based on the slope data extracted from the digital elevation model, dynamically adjust the convolution kernel size for morphological post-processing;

[0013] Step S6: Output the optimized parcel boundary vector data.

[0014] As a preferred embodiment of the land parcel boundary identification method based on remote sensing data described in this invention, wherein: the adaptive weight gating unit in step S2 executes:

[0015] The shadow area is defined based on the near-infrared reflectance threshold of the optical image.

[0016] In shadowed areas, the synthetic aperture radar imagery is assigned a weight increment, while in unshaded areas, the optical imagery is assigned a weight increment.

[0017] A fused feature map is generated by weighted summation;

[0018] The weight increment is calculated as follows:

[0019] The weight of synthetic aperture radar images in shadowed areas is set to 1.5-2 times that of optical images;

[0020] The weight of optical images in non-shaded areas is set to 1.5-2 times that of synthetic aperture radar images.

[0021] As a preferred embodiment of the land parcel boundary identification method based on remote sensing data described in this invention, the self-supervised pre-training in step S3 includes:

[0022] Randomly crop unlabeled remote sensing images to generate positive sample pairs, and crop across plots to generate negative sample pairs;

[0023] By comparing the loss functions, the distance between positive sample pairs and features is narrowed, while the distance between negative sample pairs and features is widened.

[0024] As a preferred embodiment of the land parcel boundary identification method based on remote sensing data described in this invention, the gradient direction consistency constraint in step S4 is implemented in the following manner:

[0025] The Sobel operator is used to calculate the gradient vectors of the predicted boundary and the true boundary, and the direction cosine similarity loss is calculated.

[0026] As a preferred embodiment of the land parcel boundary identification method based on remote sensing data described in this invention, in step S4, after the initial boundary probability map is generated, a gradient direction consistency constraint is applied. This constraint is calculated by plotting the predicted boundary probability map P. pred With the truth boundary T true The gradient-direction cosine similarity loss is implemented, and the specific execution process includes:

[0027] Calculate the bidirectional gradient components using the Sobel operator:

[0028] in,

[0029] The Sobel convolution kernel is in the x-direction.

[0030] The Sobel convolution kernel is in the y-direction.

[0031] P pred (i,j) represents the predicted boundary probability value at position (i,j), g px (i,j) and g py (i,j) represent the gradient components in the x and y directions of the predicted boundary probability map, respectively.

[0032] Calculate the true boundary gradient:

[0033]

[0034] Among them, T true (i,j) represents the truth boundary at position (i,j), g tx (i,j) and g ty (i,j) represent the gradient components in the x and y directions of the truth boundary, respectively;

[0035] Calculate the magnitude of the true gradient vector as a measure of significance:

[0036]

[0037] Among them, |g t (i,j)| represents the magnitude of the true gradient vector at position (i,j);

[0038] Set the gradient amplitude threshold τ = 0.1, determined based on image dynamic range normalization, when |g t Skip the loss calculation for pixel (i,j)|<τ;

[0039] For effective pixels |g t Calculate the direction cosines of the predicted and true gradients for (i,j)|≥τ:

[0040]

[0041] Where cosθ(i,j) represents the direction cosine similarity at position (i,j), and the numerator is the predicted gradient vector g. p (i,j)=(g px (i,j),g py (i,j) and the true gradient vector g t (i,j)=(g tx (i,j),g ty The dot product of (i,j) is such that the denominator contains the product of the predicted gradient magnitude and the true gradient magnitude, ∈=10 -8 It is a minimal constant used to avoid division by zero errors;

[0042] Introduce an adaptive weighting function based on gradient significance:

[0043] w(i,j)=tanh(|g t (i,j)|),

[0044] Where w(i,j) is the adaptive weight at position (i,j), and tanh(·) is the hyperbolic tangent function that maps the gradient magnitude to the interval [0,1).

[0045] Construct pixel-level orientation loss:

[0046]

[0047] in, The direction cosine similarity loss is represented at position (i,j);

[0048] The global loss is obtained by averaging the effective pixels:

[0049]

[0050] in, For global directional consistency loss, Ω represents the loss that satisfies |g t The set of pixels (i,j)|≥τ, where N=|Ω| is the total number of valid pixels.

[0051] As a preferred embodiment of the land parcel boundary identification method based on remote sensing data described in this invention, step S5, which involves dynamically adjusting the morphological convolution kernel size, includes:

[0052] Establish a negative correlation between the absolute value of the slope and the kernel size to determine the kernel size under the current slope.

[0053] As a preferred embodiment of the land parcel boundary identification method based on remote sensing data described in this invention, in step S5, during the process of dynamically adjusting the morphological convolution kernel size, after obtaining the initial boundary probability map, the morphological post-processing convolution kernel size is dynamically adjusted based on the slope data extracted from the digital elevation model (DEM). The steps include:

[0054] Read the raw slope value θ from the DEM data raw (i,j), normalize it to the interval [0,1]:

[0055]

[0056] Where, θ raw (i,j) represents the original slope angle at position (i,j), in degrees, θ. min =0° and θ max = 45° are the minimum and maximum effective slope angles preset by the system, respectively, θ n (i,j) is the normalized slope value, which is dimensionless;

[0057] Establish a negative correlation between slope and kernel size:

[0058]

[0059] Among them, K b (i,j) represents the size of the base convolution kernel at position (i,j), in pixels, K min and K max These are the preset minimum and maximum convolutional kernel sizes, in pixels. λ = 3.0 is the attenuation coefficient, dimensionless, which controls the intensity of the effect of slope on kernel size and is calibrated through the slope-kernel size response curve. The exponential function ensures that the kernel size decreases smoothly as the slope increases.

[0060] Introducing a terrain curvature compensation factor enhances detail preservation in steep slope areas:

[0061]

[0062] K c (i,j)=K b (i,j)·(1-α·tanh(β·C(i,j))),

[0063] Where C(i,j) represents the terrain curvature at position (i,j), in meters. -1 The calculation is performed using the second derivative of the DEM elevation data h(i,j); α = 0.3 is the curvature influence weight, dimensionless, β = 0.5 is the curvature scaling factor, dimensionless, and K... c(i,j) is the compensated convolution kernel size in pixels; the hyperbolic tangent function tanh(·) restricts the curvature effect to the interval [0,α);

[0064] Convert contiguous kernel sizes to odd pixel sizes:

[0065]

[0066] Among them, K final (i,j) represents the final convolution kernel size at position (i,j), in pixels. This is the floor function;

[0067] Constraints: When K c (i,j) <K min Forced K final (i,j)=K min When K c (i,j)>K max Forced K final (i,j)=K max .

[0068] Secondly, the present invention provides a land parcel boundary identification system based on remote sensing data, comprising,

[0069] Multi-source data input interface, simultaneously receiving optical images and synthetic aperture radar images;

[0070] Adaptive fusion processor with built-in weighted gating logic circuitry;

[0071] Boundary recognition engine, loading a self-supervised pre-trained convolutional neural network model;

[0072] Gradient optimizer, configured to perform gradient direction cosine similarity calculation;

[0073] The dynamic post-processing unit connects to the slope database and configures adjustable convolution kernels.

[0074] As a preferred embodiment of the land parcel boundary identification system based on remote sensing data described in this invention, the dynamic post-processing unit comprises:

[0075] The slope analysis module extracts pixel-level slope values ​​from digital elevation model data.

[0076] Kernel size mapper, outputs convolution kernel size parameters in real time according to slope value;

[0077] The kernel size mapper includes:

[0078] The slope classification unit discretizes the slope values ​​into three levels: flat, gentle, and steep.

[0079] The kernel size matching unit assigns the largest convolution kernel to flat regions and the smallest convolution kernel to steep regions.

[0080] As a preferred embodiment of the land parcel boundary identification system based on remote sensing data described in this invention, the boundary identification engine comprises:

[0081] The illumination sensing branch outputs a probability map of shadow areas.

[0082] Geometric structure branch, output boundary geometric feature map;

[0083] A feature decoupler uses a shadow probability map to suppress shadow interference regions in the geometric feature map;

[0084] The feature decoupler performs the following:

[0085] shadow probability Figure 2 Values ​​are converted into a mask matrix;

[0086] The feature response values ​​of the masked regions in the geometric feature map are suppressed by element-wise multiplication.

[0087] The beneficial effects of this invention are as follows: The complementary fusion of optical and SAR data overcomes the limitations of single data sources in extreme lighting (shadow) or texture blurry scenes, significantly enhancing the boundary detection rate in complex environments and effectively improving fusion perception capabilities; the self-supervised pre-training method utilizes massive amounts of unlabeled data to learn general edge features, reducing the need for manual annotation, especially suitable for areas with scarce annotation resources, thus reducing annotation dependence; in addition, the gradient direction constraint forces the model to learn edge directions that conform to physical laws, solving the problem of fractures and burrs in weak boundary areas, such as farmland during the growing season, and optimizing geometric continuity.

[0088] Meanwhile, the slope-driven dynamic morphological post-processing of this invention achieves a balance between denoising in flat areas and preserving shape on steep slopes, avoiding over-smoothing or under-processing of fixed parameters in complex terrain scenarios, thus enhancing terrain adaptability. Attached Figure Description

[0089] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0090] Figure 1 This is a flowchart illustrating the land parcel boundary identification method based on remote sensing data in Example 1.

[0091] Figure 2This is a schematic diagram of the framework of the land parcel boundary identification system based on remote sensing data in Example 1. Detailed Implementation

[0092] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0093] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0094] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0095] Example 1, referring to Figure 1 and Figure 2 This embodiment provides a method for identifying land parcel boundaries based on remote sensing data, including the following steps:

[0096] Step S1: Acquire optical remote sensing images and synthetic aperture radar images;

[0097] Step S2: The two types of images are fused by an adaptive weight gating unit to generate a fused feature map. Its dynamic weight is jointly determined by the near-infrared band reflectivity and radar scattering intensity.

[0098] In step S2, the adaptive weight gating unit is executed as follows:

[0099] The shadow area is defined based on the near-infrared reflectance threshold of the optical image.

[0100] In shadowed areas, the synthetic aperture radar imagery is assigned a weight increment, while in unshaded areas, the optical imagery is assigned a weight increment.

[0101] A fused feature map is generated by weighted summation;

[0102] The weight increment is calculated as follows:

[0103] The weight of synthetic aperture radar images in shadowed areas is set to 1.5-2 times that of optical images;

[0104] The weight of optical images in non-shadowed areas is set to 1.5-2 times that of synthetic aperture radar images;

[0105] Step S3: Input the fused feature map into the self-supervised pre-trained boundary recognition model and output the initial boundary probability map;

[0106] Step S3, self-supervised pre-training, includes:

[0107] Randomly crop unlabeled remote sensing images to generate positive sample pairs, and crop across plots to generate negative sample pairs;

[0108] By comparing the loss functions, the distance between positive sample pairs and features is narrowed, while the distance between negative sample pairs and features is increased.

[0109] Step S4: Apply gradient direction consistency constraints to the initial boundary probability map and calculate the gradient direction cosine similarity loss between the predicted boundary and the true boundary.

[0110] The gradient direction consistency constraint in step S4 is implemented in the following way:

[0111] The Sobel operator is used to calculate the gradient vectors of the predicted boundary and the true boundary respectively, and the direction cosine similarity loss is calculated.

[0112] In step S4, after the initial boundary probability map is generated, a gradient direction consistency constraint is applied. This constraint is calculated by plotting the predicted boundary probability map P. pred With the truth boundary T true The gradient-direction cosine similarity loss is implemented, and the specific execution process includes:

[0113] Calculate the bidirectional gradient components using the Sobel operator:

[0114]

[0115] in,

[0116] The Sobel convolution kernel is in the x-direction.

[0117] The Sobel convolution kernel is in the y-direction.

[0118] P pred (i,j) represents the predicted boundary probability value at position (i,j), g px (i,j) and g py (i,j) represent the gradient components in the x and y directions of the predicted boundary probability map, respectively.

[0119] Calculate the true boundary gradient:

[0120]

[0121] Among them, T true(i,j) represents the truth boundary at position (i,j), g tx (i,j) and g ty (i,j) represent the gradient components in the x and y directions of the truth boundary, respectively;

[0122] Calculate the magnitude of the true gradient vector as a measure of significance:

[0123]

[0124] Among them, |g t (i,j)| represents the magnitude of the true gradient vector at position (i,j);

[0125] Set the gradient amplitude threshold τ = 0.1, determined based on image dynamic range normalization, when |g t Skip the loss calculation for pixel (i,j)|<τ;

[0126] For effective pixels |g t Calculate the direction cosines of the predicted and true gradients for (i,j)|≥τ:

[0127]

[0128] Where cosθ(i,j) represents the direction cosine similarity at position (i,j), and the numerator is the predicted gradient vector g. p (i,j)=(g px (i,j),g py (i,j) and the true gradient vector g t (i,j)=(g tx (i,j),g ty The dot product of (i,j) is such that the denominator contains the product of the predicted gradient magnitude and the true gradient magnitude, ∈=10 -8 It is a minimal constant used to avoid division by zero errors;

[0129] Introduce an adaptive weighting function based on gradient significance:

[0130] w(i,j)=tanh(|g t (i,j)|),

[0131] Where w(i,j) is the adaptive weight at position (i,j), and tanh(·) is the hyperbolic tangent function, which maps the gradient magnitude to the interval [0,1). The hyperbolic tangent provides gradient-sensitive saturation characteristics to avoid oversaturation of weights in the low gradient region.

[0132] Construct pixel-level orientation loss:

[0133]

[0134] in, The direction cosine similarity loss is represented at position (i,j);

[0135] The global loss is obtained by averaging the effective pixels:

[0136]

[0137] in, For global directional consistency loss, Ω represents the loss that satisfies |g t The set of pixels (i,j)|≥τ, where N=|Ω| is the total number of valid pixels;

[0138] Specifically, the gradient direction consistency constraint significantly improves the geometric accuracy of boundary recognition by forcing the local geometry of the predicted boundary to align with that of the true boundary. In low-light or shadow areas, traditional methods are susceptible to noise interference that can lead to boundary breaks. This constraint utilizes direction cosine similarity loss to effectively suppress false boundaries with inconsistent gradient directions, especially enhancing continuity in blurred edge regions. The adaptive weighting mechanism prioritizes optimization of high-gradient regions, making the main boundaries sharper while reducing the risk of overfitting in flat regions. The added gradient magnitude threshold eliminates the influence of noise points, improving the robustness of the loss function.

[0139] Step S5: Based on the slope data extracted from the digital elevation model, dynamically adjust the convolution kernel size for morphological post-processing;

[0140] Step S5, which involves dynamically adjusting the morphological convolution kernel size, includes:

[0141] Establish a negative correlation between the absolute value of the slope and the kernel size to determine the kernel size under the current slope;

[0142] In step S5, during the process of dynamically adjusting the morphological convolution kernel size, after obtaining the initial boundary probability map, the morphological post-processing convolution kernel size is dynamically adjusted based on the slope data extracted from the digital elevation model (DEM). The steps include:

[0143] Read the raw slope value θ from the DEM data raw (i,j), normalize it to the interval [0,1]:

[0144]

[0145] Where, θ raw (i,j) represents the original slope angle at position (i,j), in degrees, θ. min =0° and θ max = 45° are the minimum and maximum effective slope angles preset by the system, respectively, θ n (i,j) is the normalized slope value, which is dimensionless;

[0146] Establish a negative correlation between slope and kernel size:

[0147]

[0148] Among them, K b (i,j) represents the size of the base convolution kernel at position (i,j), in pixels, K min and K max These are the preset minimum and maximum convolutional kernel sizes, in pixels. λ = 3.0 is the attenuation coefficient, dimensionless, which controls the intensity of the effect of slope on kernel size and is calibrated through the slope-kernel size response curve. The exponential function ensures that the kernel size decreases smoothly as the slope increases.

[0149] Introducing a terrain curvature compensation factor enhances detail preservation in steep slope areas:

[0150]

[0151] K c (i,j)=K b (i,j)·(1-α·tanh(β·C(i,j))),

[0152] Where C(i,j) represents the terrain curvature at position (i,j), in meters. -1 The calculation is performed using the second derivative of the DEM elevation data h(i,j); α = 0.3 is the curvature influence weight, dimensionless, β = 0.5 is the curvature scaling factor, dimensionless, and K... c (i,j) is the compensated convolution kernel size in pixels; the hyperbolic tangent function tanh(·) restricts the curvature effect to the interval [0,α);

[0153] Convert contiguous kernel sizes to odd pixel sizes:

[0154]

[0155] Among them, K final (i,j) represents the final convolution kernel size at position (i,j), in pixels. This is the floor function;

[0156] Constraints: When K c (i,j) <K min Forced K final (i,j)=K min When K c (i,j)>K max Forced K final (i,j)=K max ;

[0157] Specifically, the dynamic convolution kernel adjustment mechanism significantly improves the boundary recognition accuracy under complex terrain; in flat areas, a larger convolution kernel is used to effectively smooth noise and eliminate broken boundaries caused by vegetation cover or shadows; in steep slope areas, the kernel size is automatically reduced to successfully preserve the edges of micro-topographic features such as terraces and ridgelines; the nonlinear mapping function avoids the jagged edges of the boundaries caused by abrupt size changes, and terrain curvature compensation further enhances the ability to capture details in ravine areas;

[0158] Step S6: Output the optimized parcel boundary vector data.

[0159] This embodiment also provides a land parcel boundary identification system based on remote sensing data, including:

[0160] Multi-source data input interface, simultaneously receiving optical images and synthetic aperture radar images;

[0161] Adaptive fusion processor with built-in weighted gating logic circuitry;

[0162] Boundary recognition engine, loading a self-supervised pre-trained convolutional neural network model;

[0163] Gradient optimizer, configured to perform gradient direction cosine similarity calculation;

[0164] A dynamic post-processing unit connects to a slope database and is configured with adjustable convolutional kernels;

[0165] The dynamic post-processing unit includes:

[0166] The slope analysis module extracts pixel-level slope values ​​from digital elevation model data;

[0167] Kernel size mapper, outputs convolution kernel size parameters in real time according to slope value;

[0168] The kernel size mapper includes:

[0169] The slope classification unit discretizes the slope values ​​into three levels: flat (0°-5°), gentle slope (5°-15°), and steep slope (>15°).

[0170] Kernel size matching unit: the largest convolution kernel is assigned to flat areas, and the smallest convolution kernel is assigned to steep slope areas;

[0171] The boundary recognition engine includes:

[0172] The illumination sensing branch outputs a probability map of shadow areas.

[0173] Geometric structure branch, output boundary geometric feature map;

[0174] A feature decoupler uses a shadow probability map to suppress shadow interference regions in the geometric feature map;

[0175] Feature decoupler execution:

[0176] shadow probability Figure 2 Values ​​are converted into a mask matrix;

[0177] The feature response values ​​of the masked regions in the geometric feature map are suppressed by element-wise multiplication.

[0178] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying land parcel boundaries based on remote sensing data, characterized in that, include, Step S1: Acquire optical remote sensing images and synthetic aperture radar images; Step S2: The two types of images are fused by an adaptive weight gating unit to generate a fused feature map. Its dynamic weight is jointly determined by the near-infrared band reflectivity and radar scattering intensity. Step S3: Input the fused feature map into the self-supervised pre-trained boundary recognition model and output the initial boundary probability map; Step S4: Apply gradient direction consistency constraints to the initial boundary probability map and calculate the gradient direction cosine similarity loss between the predicted boundary and the true boundary. Step S5: Based on the slope data extracted from the digital elevation model, dynamically adjust the convolution kernel size for morphological post-processing; Step S6: Output the optimized parcel boundary vector data.

2. The method for identifying land parcel boundaries based on remote sensing data as described in claim 1, characterized in that, The adaptive weight gating unit described in step S2 executes as follows: The shadow area is defined based on the near-infrared reflectance threshold of the optical image. In shadowed areas, the synthetic aperture radar imagery is assigned a weight increment, while in unshaded areas, the optical imagery is assigned a weight increment. A fused feature map is generated by weighted summation; The weight increment is calculated as follows: The weight of synthetic aperture radar images in shadowed areas is set to 1.5-2 times that of optical images; The weight of optical images in non-shaded areas is set to 1.5-2 times that of synthetic aperture radar images.

3. The method for identifying land parcel boundaries based on remote sensing data as described in claim 1, characterized in that, The self-supervised pre-training mentioned in step S3 includes: Randomly crop unlabeled remote sensing images to generate positive sample pairs, and crop across plots to generate negative sample pairs; By comparing the loss functions, the distance between positive sample pairs and features is narrowed, while the distance between negative sample pairs and features is widened.

4. The method for identifying land parcel boundaries based on remote sensing data as described in claim 1, characterized in that, The gradient direction consistency constraint mentioned in step S4 is implemented in the following way: The Sobel operator is used to calculate the gradient vectors of the predicted boundary and the true boundary, and the direction cosine similarity loss is calculated.

5. The method for identifying land parcel boundaries based on remote sensing data as described in claim 4, characterized in that, In step S4, after the initial boundary probability map is generated, a gradient direction consistency constraint is applied. This constraint is calculated by plotting the predicted boundary probability map P. pred With the truth boundary T true The gradient-direction cosine similarity loss is implemented, and the specific execution process includes: Calculate the bidirectional gradient components using the Sobel operator: in, The Sobel convolution kernel is in the x-direction. The Sobel convolution kernel is in the y-direction. P pred (i,j) represents the predicted boundary probability value at position (i,j), g px (i,j) and g py (i,j) represent the gradient components in the x and y directions of the predicted boundary probability map, respectively. Calculate the true boundary gradient: Among them, T true (i,j) represents the truth boundary at position (i,j), g tx (i,j) and g ty (i,j) represent the gradient components in the x and y directions of the truth boundary, respectively; Calculate the magnitude of the true gradient vector as a measure of significance: Among them, |g t (i,j)|| represents the magnitude of the true gradient vector at position (i,j); Set the gradient amplitude threshold τ = 0.1, determined based on image dynamic range normalization, when |g t Skip the loss calculation for pixel (i,j)|<τ; For effective pixels |g t Calculate the direction cosines of the predicted and true gradients for (i,j)|≥τ: Where cosθ(i,j) represents the direction cosine similarity at position (i,j), and the numerator is the predicted gradient vector g. p (i,j)=(g px (i,j),g py (i,j) and the true gradient vector g t (i,j)=(g tx (i,j),g ty The dot product of (i,j) is such that the denominator contains the product of the predicted gradient magnitude and the true gradient magnitude, ∈=10 -8 It is a minimal constant used to avoid division by zero errors; Introduce an adaptive weighting function based on gradient significance: w(i,j)=tanh(|g t (i,j)|), Where w(i,j) is the adaptive weight at position (i,j), and tanh(·) is the hyperbolic tangent function that maps the gradient magnitude to the interval [0,1). Construct pixel-level orientation loss: in, The direction cosine similarity loss is represented at position (i,j); The global loss is obtained by averaging the effective pixels: in, For global directional consistency loss, Ω represents the loss that satisfies |g t The set of pixels (i,j)|≥τ, where N=|Ω| is the total number of valid pixels.

6. The method for identifying land parcel boundaries based on remote sensing data as described in claim 1, characterized in that, The dynamic adjustment of morphological convolution kernel size in step S5 includes: Establish a negative correlation between the absolute value of the slope and the kernel size to determine the kernel size under the current slope.

7. The method for identifying land parcel boundaries based on remote sensing data as described in claim 6, characterized in that, In step S5, during the process of dynamically adjusting the morphological convolution kernel size, after obtaining the initial boundary probability map, the morphological post-processing convolution kernel size is dynamically adjusted based on the slope data extracted from the digital elevation model (DEM). The steps include: Read the raw slope value θ from the DEM data raw (i,j), normalize it to the interval [0,1]: Where, θ raw (i,j) represents the original slope angle at position (i,j), in degrees, θ. min =0° and θ max = 45° are the minimum and maximum effective slope angles preset by the system, respectively, θ n (i,j) is the normalized slope value, which is dimensionless; Establish a negative correlation between slope and kernel size: Among them, K b (i,j) represents the size of the base convolution kernel at position (i,j), in pixels, K min and K max These are the preset minimum and maximum convolutional kernel sizes, in pixels. λ = 3.0 is the attenuation coefficient, dimensionless, which controls the intensity of the effect of slope on kernel size and is calibrated through the slope-kernel size response curve. The exponential function ensures that the kernel size decreases smoothly as the slope increases. Introducing a terrain curvature compensation factor enhances detail preservation in steep slope areas: K c (i,j)=K b (i,j)·(1-α·tanh(β·C(i,j))), Where C(i,j) represents the terrain curvature at position (i,j), in meters. -1 The calculation is performed using the second derivative of the DEM elevation data h(i,j); α = 0.3 is the curvature influence weight, dimensionless, β = 0.5 is the curvature scaling factor, dimensionless, and K... c (i,j) is the compensated convolution kernel size in pixels; the hyperbolic tangent function tanh(·) restricts the curvature effect to the interval [0,α); Convert contiguous kernel sizes to odd pixel sizes: Among them, K final (i,j) represents the final convolution kernel size at position (i,j), in pixels. This is the floor function; Constraints: When K c (i,j) <K min Forced K final (i,j)=K min When K c (i,j)>K max Forced K final (i,j)=K max .

8. A land parcel boundary identification system based on remote sensing data, based on the land parcel boundary identification method based on remote sensing data according to any one of claims 1 to 7, characterized in that, include: Multi-source data input interface, simultaneously receiving optical images and synthetic aperture radar images; Adaptive fusion processor with built-in weighted gating logic circuitry; Boundary recognition engine, loading a self-supervised pre-trained convolutional neural network model; Gradient optimizer, configured to perform gradient direction cosine similarity calculation; The dynamic post-processing unit connects to the slope database and configures adjustable convolution kernels.

9. A land parcel boundary identification system based on remote sensing data as described in claim 8, characterized in that, The dynamic post-processing unit includes: The slope analysis module extracts pixel-level slope values ​​from digital elevation model data. Kernel size mapper, outputs convolution kernel size parameters in real time according to slope value; The kernel size mapper includes: The slope classification unit discretizes the slope values ​​into three levels: flat, gentle, and steep. The kernel size matching unit assigns the largest convolution kernel to flat regions and the smallest convolution kernel to steep regions.

10. A land parcel boundary identification system based on remote sensing data as described in claim 8, characterized in that, The boundary recognition engine includes: The illumination sensing branch outputs a probability map of shadow areas. Geometric structure branch, output boundary geometric feature map; A feature decoupler uses a shadow probability map to suppress shadow interference regions in the geometric feature map; The feature decoupler performs the following: Binarize the shadow probability map into a mask matrix; The feature response values ​​of the masked regions in the geometric feature map are suppressed by element-wise multiplication.

Images