A multi-source remote sensing data fusion large-scale farmland boundary recognition method and system

Abstract

The present application relates to the technical field of object recognition, and discloses a large-scale farmland boundary recognition method and system based on multi-source remote sensing data fusion, the method comprising: adaptively fusing multi-spectral remote sensing images to obtain a spectral fusion remote sensing image, and mapping pixel values of pixels in the spectral fusion remote sensing image to three-dimensional coordinate points on the surface of a three-dimensional terrain structure model; performing multi-modal gating fusion on the mapped multi-source remote sensing image to obtain a boundary probability image; extracting a coarse-grained farmland boundary from the boundary probability image, and combining slope information of pixels in a slope image to constrain and optimize the coarse-grained farmland boundary. The present application constructs a farmland three-dimensional terrain structure model, fuses SAR polarization characteristics and spectral indexes, generates a boundary probability image through multi-modal gating, constructs a coarse-grained boundary, combines terrain slope information, introduces slope driving constraints, optimizes the physical consistency of the coarse-grained boundary, and recognizes continuous and real farmland boundaries.

Description

Technical Field

[0001] This invention relates to the field of object recognition technology, and more particularly to the field of farmland boundary recognition technology, specifically a method and system for large-scale farmland boundary recognition based on multi-source remote sensing data fusion. Background Technology

[0002] In agricultural production, boundary identification is not only crucial for monitoring farmland area, analyzing crop growth, and assessing disasters, but also plays a fundamental role in the implementation of precision agriculture. Remote sensing data, due to its efficiency and wide-area monitoring capabilities, has become an important data source for farmland boundary identification research.

[0003] In the field of agricultural remote sensing, high-precision identification of farmland boundaries is of great significance for farmland management, plot segmentation, and agricultural informatization. In existing research, CN114648704A provides a method and system for high-precision extraction of farmland boundaries, including: acquiring remote sensing images of the area to be extracted in multiple bands; inputting the remote sensing image of each band into a recurrent residual convolutional neural network to obtain multiple coarse farmland boundary maps; fusing all the coarse farmland boundary maps and calculating the boundary strength in the fused map to obtain a primary farmland boundary map; and using a watershed transform algorithm to close the gaps in the primary farmland boundary map to obtain the final farmland boundary. However, this method does not incorporate three-dimensional terrain structure, cannot perceive slope / undulation, and is prone to boundary jumps or fragmentation. Furthermore, the recurrent residual convolutional structure lacks a fusion mechanism for physical features (such as slope, solar altitude, etc.), resulting in poor generalization ability under different terrain and remote sensing conditions, and a tendency to overfit small sample scenarios.

[0004] However, traditional methods mostly rely on single remote sensing images or only two-dimensional images for boundary extraction, making it difficult to handle complex terrain (such as mountain terraces), mixed vegetation, and boundary fracturing in large-scale areas. Furthermore, existing methods lack the ability to perceive topographic relief information, failing to effectively distinguish between true plot boundaries and false edges caused by shadows, soil moisture, or reflection interference. With the improvement of remote sensing sensor capabilities, fusing multispectral remote sensing images, synthetic aperture radar (SAR) images, and digital elevation models (DEMs) has become an effective direction for improving boundary recognition accuracy. However, issues such as resolution inconsistencies, complex spatial registration, and feature inconsistencies among multi-source data still pose significant challenges to fusion modeling. Therefore, a new method for farmland boundary recognition that integrates spectral, structural, and radar information is urgently needed to enhance boundary continuity, stability, and topographic plausibility. Summary of the Invention

[0005] In view of this, the present invention provides a large-scale farmland boundary identification method and system based on multi-source remote sensing data fusion. A three-dimensional terrain structure model is constructed based on multispectral images, SAR images and digital elevation models. Then, spectral index and SAR polarization features are extracted and fused to generate a boundary probability map. By extracting coarse-grained farmland boundaries and introducing slope information to optimize the curve shape, fine boundary identification in complex terrain areas can be achieved.

[0006] To achieve the above objectives, this invention provides a large-scale farmland boundary identification method based on multi-source remote sensing data fusion, comprising the following steps:

[0007] S1: Collect multi-source remote sensing data of farmland areas, including multispectral remote sensing images, SAR images and digital elevation models. Use the digital elevation model of farmland areas to perform terrain perception of farmland areas and generate a three-dimensional terrain structure model of farmland areas.

[0008] S2: Adaptive spectral fusion is performed on the multispectral remote sensing image to obtain a spectral fusion remote sensing image. Based on the collinearity equation correction principle, the pixel values ​​of the pixels in the spectral fusion remote sensing image are mapped to the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model.

[0009] S3: Extract the SAR polarization feature image and spectral index fusion image of the surface of the mapped 3D terrain structure model, perform multimodal gating fusion on the extracted images, and obtain the boundary probability image of the surface of the 3D terrain structure model.

[0010] S4: Extract coarse-grained farmland boundaries from the boundary probability image and extract slope images from the 3D terrain structure model. Combine the slope information of pixels in the slope image to constrain and optimize the coarse-grained farmland boundaries, and obtain smooth farmland boundary recognition results.

[0011] As a further improvement of the present invention:

[0012] Optionally, multi-source remote sensing data of farmland areas are collected, and digital elevation models of the farmland areas are used to perform terrain perception of the farmland areas, generating a three-dimensional terrain structure model of the farmland areas, including:

[0013] A digital elevation model of the farmland area is extracted. The digital elevation model records the elevation values ​​of some two-dimensional geographic coordinates in the farmland area. The two-dimensional geographic coordinates and their corresponding elevation values ​​are used as point cloud data of the farmland area. Specifically, the two-dimensional geographic coordinates are latitude and longitude coordinates.

[0014] The point cloud data is filtered using median filtering.

[0015] For each point cloud data, search for the K nearest neighbor point cloud data to form a local plane, and calculate the normal vector of the local plane as the normal vector of the point cloud data;

[0016] The normal vectors of all point cloud data are extracted to form a normal vector field. The normal vector field is converted into the divergence value of the gradient field by the divergence calculation method. The Poisson equation of the three-dimensional surface of the farmland area is constructed. The conjugate gradient algorithm is used to iteratively solve the Poisson equation to obtain the implicit representation of the three-dimensional surface of the farmland area.

[0017] The implicit representation of the three-dimensional surface of farmland area obtained by solving the Poisson equation is converted into an explicit triangular mesh model using a meshing method.

[0018] Based on the triangular mesh model, all two-dimensional geographic coordinates and their corresponding elevation values ​​of the farmland area are extracted to form a complete three-dimensional structural model of the farmland area. The three-dimensional structural model of the farmland area describes the elevation values ​​of the farmland area at different two-dimensional geographic coordinates, and the two-dimensional geographic coordinates and their corresponding elevation values ​​are used as three-dimensional coordinate points on the surface of the three-dimensional structural model.

[0019] Specifically, the SAR image is a synthetic aperture radar image. The SAR radar transmits horizontally polarized electromagnetic waves and vertically polarized electromagnetic waves to the two-dimensional geographic coordinates of the farmland area, and receives the echo signals in both horizontal and vertical modes. This yields horizontally polarized echo signals, vertically polarized echo signals, and cross-polarized echo signals for the two-dimensional geographic coordinates. The horizontally polarized echo signals are horizontally transmitted and horizontally received echo signals, the vertically polarized echo signals are vertically transmitted and vertically received echo signals, and the cross-polarized echo signals are horizontally transmitted and vertically received echo signals, as well as vertically transmitted and horizontally received echo signals. The intensity of the echo signals is calculated to form a scattering matrix for the two-dimensional geographic coordinates. The two-dimensional geographic coordinates are then mapped to three-dimensional coordinate points on the surface of a three-dimensional structural model of the farmland area. The scattering matrix is ​​used as the pixel values ​​of the three-dimensional coordinate points in the SAR image.

[0020] Optionally, adaptive spectral fusion is performed on the multispectral remote sensing image to obtain a spectrally fused remote sensing image, including:

[0021] The multispectral remote sensing image is composed of spectral remote sensing images of multiple bands, including bands B3, B4, B8 and B11; specifically, bands B3, B4, B8 and B11 belong to the green light band, red light band, near-infrared band and short-infrared band respectively.

[0022] Adaptive spatial attention calculation is performed on spectral remote sensing images of different bands to obtain the spatial attention weights of pixels in the spectral remote sensing images. The pixel values ​​of the pixels in the spectral remote sensing images are then weighted to obtain the weighted pixel values ​​of pixels in different bands in the multispectral remote sensing image. The vegetation index, water index, and soil moisture index of different pixels in the multispectral remote sensing image are calculated. These indices are then used as the pixel values ​​of pixels in the multispectral remote sensing image to construct a spectral fusion remote sensing image. Specifically, the spectral remote sensing image is a pixel matrix S of row A1 and column A2, and the pixel value of each pixel in the spectral remote sensing image is the reflectance of the pixel in the corresponding band. The adaptive spatial attention calculation method is as follows:

[0023] Att S =δ(Conv 7×7 (Concat(AvgPool(S),MaxPool(S))));

[0024] Among them, Att S Let S represent the spatial attention weight matrix corresponding to the pixel matrix S. AvgPool((S)) represents the average pooling process performed on the pixel matrix S, and MaxPool(S) represents the maximum pooling process performed on the pixel matrix S. The result of the average pooling process reflects the background smoothness features of the pixel matrix, and the result of the maximum pooling process reflects the edges and high-frequency pixel regions of the pixel matrix.

[0025] Concat(AvgPool(S),MaxPool(S)) represents concatenating the average pooling result AvgPool(S) and the maximum pooling result MaxPool((S)) along the channel dimension. The Concat(AvgPool(S),MaxPool((S)) is a 2-channel feature map in the form of row A1 and column A2.

[0026] Conv 7×7 (·) represents a 2D convolution operation with a kernel size of 7×7, δ(·) represents the Sigmoid activation function, and the Att S Let A1 be the spatial attention weight matrix and A2 be the spatial attention weight matrix. Att S Any element in the matrix S is between 0 and 1, corresponding to the spatial attention weight of the pixel in the pixel matrix S.

[0027] The formulas for calculating the vegetation index, water index, and soil moisture index of different pixels in the multispectral remote sensing image are as follows:

[0028]

[0029] Where S1(i,j), S2(i,j), and S3(i,j) represent the vegetation index, water index, and soil moisture index of the pixel in the i-th row and j-th column of the multispectral remote sensing image, respectively; B3(i,j), B4(i,j), B8(i,j), and B11(i,j) represent the weighted pixel values ​​of the pixel in the i-th row and j-th column of the multispectral remote sensing image under the B3 band, B4 band, B8 band, and B11 band, respectively; i∈[1,A1], j∈[1,A2];

[0030] Specifically, healthy vegetation in farmland areas reflects strongly in the near-infrared band, while withered vegetation or bare soil reflects strongly in the red band. Therefore, using the B4 and B8 bands to calculate the vegetation index can accurately identify crop-growing areas and surrounding unplanted areas within farmland. Water bodies reflect weakly in the near-infrared band but strongly in the green band. Therefore, using the B3 and B8 bands to calculate the water body index can identify irrigation water bodies, moist areas, and dry soil, helping to delineate the boundaries between farmland and water bodies. The short-wave infrared band is very sensitive to water absorption. Moist soil reflects weakly in this band, while dry soil reflects strongly. The difference in reflection between moist and dry soil in the near-infrared band is small, but vegetation reflects strongly. Combining the B11 and B8 bands to calculate the soil moisture index serves as the measure of soil moisture, monitoring soil moisture levels and distinguishing between moist farmland and arid areas.

[0031] Before calculating each index, spatial attention weights of pixels in different bands are calculated using spatial attention. Based on the characteristics of different regions, the weights of each pixel in different bands are dynamically adjusted. The weights reflect the relative importance of different regions in the image, thereby improving the ability to express spatial information in the image. This improves the accuracy of important regions (such as farmland, water bodies, or wetlands) in the image when calculating vegetation index, water body index, and soil moisture index, and thus improves the accuracy of farmland boundary recognition.

[0032] Optionally, mapping the pixel values ​​of pixels in the spectral fusion remote sensing image to three-dimensional coordinate points on the surface of the three-dimensional terrain structure model includes:

[0033] Specifically, the collinearity equation correction principle is as follows: a certain three-dimensional coordinate point in space is collinear with its pixel point on the sensor imaging plane and the three-dimensional coordinate point of the spatial imaging center, satisfying the projection geometric relationship;

[0034] The three-dimensional coordinate points on the surface of the three-dimensional terrain structure model are normalized to obtain normalized three-dimensional coordinate points.

[0035] The normalized three-dimensional coordinate points are mapped using the RPC model to obtain the pixel coordinates of the normalized three-dimensional coordinate points in the spectral fusion remote sensing image. The pixel values ​​of the pixel coordinates in the spectral fusion remote sensing image are extracted as the spectral index fusion values ​​of the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model associated with the normalized three-dimensional coordinate points.

[0036] Specifically, the RPC model is used to describe the imaging geometry of remote sensing images, and is in the form of a rational polynomial, provided by the provider of multi-source remote sensing data;

[0037] As a preferred algorithm of this invention, the solar elevation angle at the time of multi-source remote sensing data acquisition and the scattering matrix of the three-dimensional coordinate points in the SAR image are combined to generate the shadow compensation coefficient of the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model. The spectral index fusion value is then weighted and compensated. The shadow compensation coefficient of the three-dimensional coordinate point ((x,y,z)) is:

[0038] in, The shadow compensation coefficient represents the 3D coordinate point (x,y,z) on the surface of the 3D terrain structure model. λ(x,y,z) represents the adaptive weight. Specifically, λ(x,y,z) = 0.2 - 0.2·min{0.5,θ(x,y,z) / θ th )},θ th This represents the preset incident angle threshold, where θ(x,y,z) represents the emission angle of the horizontally polarized electromagnetic wave emitted by the SAR towards the two-dimensional geographic coordinates corresponding to the three-dimensional coordinate point (x,y,z). * ,y * ,z * HV(x,y,z) represents the location of the SAR, HV(x,y,z) represents the horizontal and vertical cross divergence in the divergence matrix of the three-dimensional coordinate point (x,y,z), and HH((x,y,z) represents the horizontal divergence in the divergence matrix of the three-dimensional coordinate point (x,y,z).

[0039] θ sun The solar altitude angle represents the moment when multi-source remote sensing data is acquired;

[0040] Due to topographical obstruction, some farmland areas receive less solar radiation, resulting in localized shadows or insufficient illumination. The spectral values ​​in these areas are artificially underestimated, causing the index calculations to underestimate vegetation, water levels, or humidity. Therefore, by combining the HV / HH ratio in SAR images, the scattering response of ground features to different polarizations can be estimated, revealing "obstructed" vegetation or structural signals and compensating for the spectral index at three-dimensional coordinate points. HV is generally more sensitive to vegetation, while HH is more sensitive to changes in water bodies and topography. Furthermore, [the text abruptly ends here, likely due to an incomplete sentence or missing information]. It corrects the differences in solar irradiance caused by terrain tilt, enhances surface reflection in the same direction as the light, compensates for the attenuation on the shaded side, makes the boundaries of farmland areas near hillsides and shady areas clearer, which is beneficial for farmland boundary extraction, and uses adaptive weights to control the size of the coefficients to avoid the shadow compensation coefficient being too large.

[0041] Optionally, SAR polarization feature images and spectral index fusion images of the surface of the mapped 3D terrain structure model are extracted, including:

[0042] Using a vertical top-down projection method, the three-dimensional coordinate points on the surface of the mapped three-dimensional terrain structure model are projected along the Z-axis to generate an index image in a two-dimensional planar view. The SAR polarization features and spectral indices of the three-dimensional coordinate points are extracted and used as the pixel values ​​of the projected pixels in the index image, respectively, to construct a SAR polarization feature image and a spectral index fusion image of the surface of the mapped three-dimensional terrain structure model.

[0043] Specifically, the polarization characteristics of the three-dimensional coordinate point (x,y,z) are HV(x,y,z) / HH(x,y,z), θ((x,y,z), where θ(x,y,z) is the emission angle of the horizontally polarized electromagnetic wave emitted by the SAR towards the two-dimensional geographic coordinate position corresponding to the three-dimensional coordinate point (x,y,z). The spectral indices of the three-dimensional coordinate point (x,y,z) are S1((x,y,z), S2((x,y,z), S3((x,y,z)).

[0044] S1(x,y,z), S2(x,y,z), and S3(x,y,z) are the vegetation index, water index, and soil moisture index of the three-dimensional coordinate point (x,y,z) respectively.

[0045] The pixel position of the three-dimensional coordinate point (x, y, z) in the indexed image is ((x, y).

[0046] Optionally, the extracted images are subjected to multimodal gating fusion to obtain a boundary probability image of the surface of the three-dimensional terrain structure model, including:

[0047] Calculate the fusion gating factor of the SAR polarization feature image and the spectral index fused image. The formula for calculating the fusion gating factor is as follows:

[0048] g=δ(W SAR ·I SAR +W Spec ·I Spec );

[0049] Where g represents the fusion gating factor, I SAR Represents SAR polarization feature image, I Spec W represents the spectral index fusion image. SARW represents the SAR weight matrix. Spec δ(·) represents the spectral index weight matrix, and δ(·) represents the Sigmoid activation function.

[0050] The fusion gating factor is used to fuse the SAR polarization feature image and the spectral index fusion image to obtain fused image I:

[0051] I=g⊙tanh(W SAR ·I SAR )+(Ug)⊙tanh(W Spec ·I Spec );

[0052] Where U represents a unit element matrix with all elements being 1 and the same scale as the indexed image, and ⊙ represents element-wise multiplication;

[0053] The fused image I is input into a convolution classifier to generate the boundary probability of each pixel in the fused image I. The boundary probability is used as the pixel value to construct the edge probability image.

[0054] Specifically, the fusion gating factor includes learnable dynamic weighting coefficients for each pixel. The fusion gating factor is a matrix of A rows and B columns, where A represents the number of 3D coordinate points of the 3D terrain structure model in the east-west direction, and B represents the number of 3D coordinate points of the 3D terrain structure model in the north-south direction. Each matrix element in the fusion gating factor is a dynamic weighting coefficient of the pixel position associated with the matrix element position. For each pixel position, the dynamic weighting coefficient is used to determine whether it relies more on SAR polarization information (structure, texture) or spectral index information (vegetation, water, soil). The closer the dynamic weighting coefficient is to 1, the more SAR features dominate the pixel position, and the pixel is likely located in a boundary or texture abrupt change region. The closer the dynamic weighting coefficient is to 0, the more spectral indexes dominate the pixel position, and the pixel is likely located in a vegetation or water body region.

[0055] Optionally, coarse-grained farmland boundaries are extracted from the boundary probability image, including:

[0056] Set a boundary probability threshold, convert the boundary probability image into a binary image, perform morphological dilation-erosion operation on the binary image, and remove isolated pixels and small breaks;

[0057] Connectivity analysis is performed on the binary image after morphological operations. All interconnected boundary regions are marked and connected to obtain multiple connected regions. The area of ​​each connected region is calculated, and the connected region with the largest area is selected as a coarse-grained farmland boundary candidate region. The edge point sequence in the coarse-grained farmland boundary candidate region is extracted using a contour tracking algorithm to form a closed curve. The closed curve is used as the coarse-grained farmland boundary.

[0058] Specifically, the pixel value of a pixel in the boundary probability map reflects the probability that the location is a farmland boundary. Through thresholding, low-probability noise is filtered out, and only areas with clear structures and well-defined boundaries are retained. Furthermore, through dilation, broken edge segments are connected, and local noise or isolated response points are removed by erosion, thereby enhancing the structural integrity and topological connectivity of the boundary. Geometric enhancement modeling is performed on continuous structures such as farmland ridges, and the area of ​​connected regions is calculated, retaining the main region or regions larger than the area threshold, effectively avoiding interference from small-area abnormal boundaries on the final recognition result.

[0059] Optionally, slope images are extracted from the three-dimensional terrain structure model, and slope information of pixels in the slope images is used to perform constraint optimization on the coarse-grained farmland boundary, including:

[0060] Calculate the slope of the three-dimensional coordinate points in the three-dimensional terrain structure model, and map the slope of the three-dimensional coordinate points to an index image to form a slope image. The index image is a two-dimensional planar view of the three-dimensional terrain structure model, and the image size of the slope image is consistent with that of the boundary probability image.

[0061] A farmland boundary energy function based on slope constraints is constructed. The farmland boundary energy function takes the farmland boundary as input and the farmland boundary energy as output. The farmland boundary energy includes internal energy characterizing the smoothness and continuity of the farmland boundary, boundary image energy driving the farmland boundary to move closer to the boundary probability position, and slope energy penalizing the crossing of steep slope areas.

[0062] The coarse-grained farmland boundary is used as the input to the farmland boundary energy function. The gradient descent algorithm is used to evolve the coarse-grained farmland boundary until the boundary evolution result meets the convergence condition. The current boundary evolution result is output as the smooth farmland boundary recognition result, and the farmland boundary recognition result is mapped to the two-dimensional planar view of the three-dimensional terrain structure model.

[0063] Specifically, the expression for the coarse-grained farmland boundary Q0 extracted from the boundary probability image is:

[0064]

[0065] Where Q0(r) represents the pixel coordinates of the r-th segment in the coarse-grained farmland boundary Q0 in the boundary probability image. r∈[0,1], where r represents the parameter for normalizing the length of farmland boundary;

[0066] The expression for the farmland boundary energy function is:

[0067] E(Q t )=w1·E1((Q t )+w2·E2(Q t,P)+w3·E2(Q t ,β);

[0068]

[0069] Among them, Q t E(·) represents the t-th boundary evolution result of the coarse-grained farmland boundary, and E(Q) represents the farmland boundary energy function. t ) represents Q t The farmland boundary energy, E1(Q) t ) represents Q t The internal energy, E2((Q) t P) represents Q t Boundary image energy, E2(Q t ,β) represents Q t The slope energy is P, which represents the boundary probability image, and β represents the slope image; w1, w2, and w3 represent the weights of the internal energy, boundary image energy, and slope energy, respectively; the boundary image energy is preceded by a negative sign to maximize the boundary probability, that is, to encourage the boundary to approach the region with a strong boundary probability response.

[0070] Q t (r) represents Q t The pixel coordinates of the r-th segment in the boundary probability image. Q represents t The first-order term of (r), Q represents t The second-order term of (r);

[0071] The first-order term is the elasticity term, which represents the sum of the squares of the line segment lengths between boundary points, corresponding to the degree of stretching or compression of the boundary. When this term is small, it encourages equidistant distribution and avoids excessive dispersion between points, which is conducive to forming a boundary with a compact structure and good contour closure. The second-order term is the smoothness term, which reflects the local curvature of each boundary point relative to its neighbors, encouraging the geometric continuity and smoothness of the boundary, and helping to avoid high-frequency oscillations in noisy regions.

[0072] P(Q t (r) represents the pixel coordinates Q in the boundary probability image. t (boundary probability at (r);

[0073] β(Q t (r) represents the pixel coordinates Q in the slope image. t The slope at (r);

[0074] τ represents the slope suppression coefficient. Denotes the activator, if β(Q) t (r)) is greater than 25 degrees, then otherwise dr represents the differential representation of parameter r.

[0075] To address the aforementioned problems, this invention provides a large-scale farmland boundary recognition system. The system includes a server and a data acquisition device. The server comprises a multimodal gating fusion module and a farmland boundary recognition module.

[0076] The multimodal gating fusion module is used to extract SAR polarization feature images and spectral index fusion images of the surface of the mapped three-dimensional terrain structure model, and to perform multimodal gating fusion on the extracted images to obtain the boundary probability image of the surface of the three-dimensional terrain structure model.

[0077] The farmland boundary recognition module is used to extract coarse-grained farmland boundaries from the boundary probability image and extract slope images from the three-dimensional terrain structure model. Combining the slope information of the pixels in the slope image, the coarse-grained farmland boundaries are constrained and optimized to obtain smooth farmland boundary recognition results.

[0078] The data acquisition device is used to collect multi-source remote sensing data of farmland areas, use the digital elevation model of farmland areas to perform terrain perception of farmland areas, generate a three-dimensional terrain structure model of farmland areas, perform adaptive spectral fusion on multispectral remote sensing images to obtain spectral fusion remote sensing images, and map the pixel values ​​of pixels in the spectral fusion remote sensing images to three-dimensional coordinate points on the surface of the three-dimensional terrain structure model based on the collinearity equation correction principle.

[0079] To achieve large-scale farmland boundary identification methods such as the multi-source remote sensing data fusion mentioned above.

[0080] To address the above problems, the present invention provides an electronic device, the electronic device comprising:

[0081] Memory, storing at least one instruction;

[0082] Communication interfaces enable communication between electronic devices; and

[0083] The processor executes the instructions stored in the memory to implement the large-scale farmland boundary identification method described above, which involves the fusion of multi-source remote sensing data.

[0084] To address the aforementioned problems, the present invention also provides a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the large-scale farmland boundary identification method based on multi-source remote sensing data fusion described above.

[0085] Compared with existing technologies, this invention proposes a method and system for large-scale farmland boundary identification through multi-source remote sensing data fusion. This technology has the following advantages:

[0086] First, this application combines the solar elevation angle at the time of multi-source remote sensing data acquisition with the scattering matrix of the three-dimensional coordinate points in the SAR image to generate a shadow compensation coefficient for the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model, and then performs weighted compensation on the fused spectral index value. Specifically, due to the existence of terrain shading, some farmland areas cannot receive sufficient solar radiation, easily forming local shadows or insufficient irradiance, which leads to lower spectral values ​​in the remote sensing image of this area, causing parameters such as vegetation index, water index, or soil moisture index to be underestimated. To overcome this problem, this application fuses polarization information in the SAR image, especially using the HV / HH polarization scattering ratio to estimate the response capability of ground objects to different polarization waves, thereby revealing the true reflection characteristics of vegetation or surface structures that still exist in the shadowed areas. Among them, HV polarization is more sensitive to vegetation structure and can supplement the vegetation signal lost due to spectral attenuation by shadows; HH polarization is more sensitive to water body boundaries and undulating terrain, which helps to characterize the reflection differences caused by terrain. Meanwhile, an adaptive weighting mechanism is used to dynamically adjust the compensation coefficients to prevent the introduction of noise or artifacts due to overcompensation, ensuring that the compensation effect has physical consistency and regional adaptability.

[0087] Meanwhile, the farmland boundary curve energy function proposed in this application integrates multimodal information from boundary probability images and slope images. By introducing a gradient adjustment factor, it suppresses boundary jumps across high-slope areas. Simultaneously, it achieves geometric continuity control of the boundary through elasticity and curvature terms, thus improving the overall physical rationality and spatial structural smoothness of the boundary curve. This is particularly suitable for fine boundary extraction scenarios in undulating terrain areas or multiple farmland structures. Specifically, by guiding contour evolution through multiple factors, it integrates internal energy (smoothness), boundary guidance, and slope suppression to achieve dual-mode constraints of spectral and topographical characteristics. This avoids false boundaries caused by spectral perturbations such as contours crossing shadows and high-reflectivity areas. Especially for areas with strong undulations such as terraces, it solves the problems of "broken boundaries" and "folded boundaries." A slope term is introduced for constraint, corresponding to the topographical continuity assumption: farmland boundaries should not cross suddenly steep slopes, avoiding non-physical edges. By introducing first-order and second-order terms and fusing them with boundary probability images and slope images, the contour adaptively adjusts its shape according to the strength of the image boundary. Attached Figure Description

[0088] Figure 1 This is a flowchart illustrating a large-scale farmland boundary identification method based on multi-source remote sensing data fusion, as provided in an embodiment of the present invention.

[0089] Figure 2 This is a spectral index fusion image based on multi-source remote sensing data fusion provided in an embodiment of the present invention.

[0090] Figure 3This is a comparison of the changes in vegetation index experimental indicators before and after shadow compensation, provided in an embodiment of the present invention.

[0091] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0092] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0093] This application provides a method for large-scale farmland boundary identification through multi-source remote sensing data fusion. The executing entity of this method includes, but is not limited to, at least one electronic device configured to execute the method provided in this application, such as a server or a terminal. In other words, the method can be executed by software or hardware installed on a terminal or server device, where the software may be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0094] Reference Figure 1 Embodiment 1 of the present invention is as follows:

[0095] A method for large-scale farmland boundary identification through multi-source remote sensing data fusion includes the following steps:

[0096] S1: Collect multi-source remote sensing data of farmland areas, including multispectral remote sensing images, SAR images and digital elevation models. Use the digital elevation model of farmland areas to perform terrain perception of farmland areas and generate a three-dimensional terrain structure model of farmland areas.

[0097] Multi-source remote sensing data of farmland areas are collected, and digital elevation models of farmland areas are used to perform terrain perception of the farmland areas, generating a three-dimensional terrain structure model of the farmland areas, including:

[0098] A digital elevation model of the farmland area is extracted. The digital elevation model records the elevation values ​​of some two-dimensional geographic coordinates in the farmland area. The two-dimensional geographic coordinates and their corresponding elevation values ​​are used as point cloud data of the farmland area. Specifically, the two-dimensional geographic coordinates are latitude and longitude coordinates.

[0099] The point cloud data is filtered using median filtering.

[0100] For each point cloud data, search for the K nearest neighbor point cloud data to form a local plane, and calculate the normal vector of the local plane as the normal vector of the point cloud data;

[0101] The normal vectors of all point cloud data are extracted to form a normal vector field. The normal vector field is converted into the divergence value of the gradient field by the divergence calculation method. The Poisson equation of the three-dimensional surface of the farmland area is constructed. The conjugate gradient algorithm is used to iteratively solve the Poisson equation to obtain the implicit representation of the three-dimensional surface of the farmland area.

[0102] The three-dimensional surface of farmland area, which is implicitly represented by the Poisson equation, is converted into an explicit triangular mesh model by a meshing method. As an embodiment of the present invention, the meshing method adopts the Delaunay triangulation algorithm or the Marching Cubes algorithm.

[0103] Based on the triangular mesh model, all two-dimensional geographic coordinates and their corresponding elevation values ​​of the farmland area are extracted to form a complete three-dimensional structural model of the farmland area. The three-dimensional structural model of the farmland area describes the elevation values ​​of the farmland area at different two-dimensional geographic coordinates, and the two-dimensional geographic coordinates and their corresponding elevation values ​​are used as three-dimensional coordinate points on the surface of the three-dimensional structural model.

[0104] S2: Adaptive spectral fusion is performed on the multispectral remote sensing image to obtain a spectral fusion remote sensing image. Based on the collinearity equation correction principle, the pixel values ​​of the pixels in the spectral fusion remote sensing image are mapped to the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model.

[0105] Adaptive spectral fusion is performed on multispectral remote sensing images to obtain spectrally fused remote sensing images, including:

[0106] The multispectral remote sensing image is composed of spectral remote sensing images of multiple bands, including bands B3, B4, B8 and B11; specifically, bands B3, B4, B8 and B11 belong to the green light band, red light band, near-infrared band and short-infrared band respectively.

[0107] Adaptive spatial attention calculation is performed on spectral remote sensing images of different bands to obtain the spatial attention weights of pixels in the spectral remote sensing images. The pixel values ​​of the pixels in the spectral remote sensing images are then weighted to obtain the weighted pixel values ​​of pixels in different bands in the multispectral remote sensing image. The vegetation index, water index, and soil moisture index of different pixels in the multispectral remote sensing image are calculated. These indices are then used as the pixel values ​​of pixels in the multispectral remote sensing image to construct a spectral fusion remote sensing image. Specifically, the spectral remote sensing image is a pixel matrix S of row A1 and column A2, and the pixel value of each pixel in the spectral remote sensing image is the reflectance of the pixel in the corresponding band. The adaptive spatial attention calculation method is as follows:

[0108] Att S =δ(Conv 7×7(Concat(AvgPool(S),MaxPool(S))));

[0109] Among them, Att S Let S represent the spatial attention weight matrix corresponding to the pixel matrix S. AvgPool((S)) represents the average pooling process performed on the pixel matrix S, and MaxPool(S) represents the maximum pooling process performed on the pixel matrix S. The result of the average pooling process reflects the background smoothness features of the pixel matrix, and the result of the maximum pooling process reflects the edges and high-frequency pixel regions of the pixel matrix.

[0110] Concat(AvgPool(S),MaxPool(S)) represents concatenating the average pooling result AvgPool(S) and the maximum pooling result MaxPool((S)) along the channel dimension. The Concat(AvgPool(S),MaxPool((S)) is a 2-channel feature map in the form of row A1 and column A2.

[0111] Conv 7×7 (·) represents a 2D convolution operation with a kernel size of 7×7, δ(·) represents the Sigmoid activation function, and the Att S Let A1 be the spatial attention weight matrix and A2 be the spatial attention weight matrix. Att S Any element in the matrix S is between 0 and 1, corresponding to the spatial attention weight of the pixel in the pixel matrix S.

[0112] The formulas for calculating the vegetation index, water index, and soil moisture index of different pixels in the multispectral remote sensing image are as follows:

[0113]

[0114] Where S1(i,j), S2(i,j), and S3(i,j) represent the vegetation index, water index, and soil moisture index of the pixel in the i-th row and j-th column of the multispectral remote sensing image, respectively, and B3(i,j), B4(i,j), B8(i,j), and B11(i,j) represent the weighted pixel values ​​of the pixel in the i-th row and j-th column of the spectral remote sensing image under the B3 band, B4 band, B8 band, and B11 band, respectively, i∈[1,A1],j∈[1,A2].

[0115] Mapping pixel values ​​from the spectral fusion remote sensing image to three-dimensional coordinates on the surface of a three-dimensional terrain structure model includes:

[0116] The three-dimensional coordinate points on the surface of the three-dimensional terrain structure model are normalized to obtain normalized three-dimensional coordinate points.

[0117] The normalized three-dimensional coordinate points are mapped using the RPC model to obtain the pixel coordinates of the normalized three-dimensional coordinate points in the spectral fusion remote sensing image. The pixel values ​​of the pixel coordinates in the spectral fusion remote sensing image are extracted as the spectral index fusion values ​​of the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model associated with the normalized three-dimensional coordinate points.

[0118] As a preferred algorithm of this invention, the solar elevation angle at the time of multi-source remote sensing data acquisition and the scattering matrix of the three-dimensional coordinate points in the SAR image are combined to generate the shadow compensation coefficient of the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model. The spectral index fusion value is then weighted and compensated. The shadow compensation coefficient of the three-dimensional coordinate point ((x,y,z)) is:

[0119]

[0120] in, The shadow compensation coefficient represents the 3D coordinate point (x,y,z) on the surface of the 3D terrain structure model. λ(x,y,z) represents the adaptive weight. Specifically, λ(x,y,z) = 0.2 - 0.2·min{0.5,θ(x,y,z) / θ th )},θ th This represents the preset incident angle threshold, where θ(x,y,z) represents the emission angle of the horizontally polarized electromagnetic wave emitted by the SAR towards the two-dimensional geographic coordinates corresponding to the three-dimensional coordinate point (x,y,z). * ,y * ,z * HV(x,y,z) represents the location of the SAR, HV(x,y,z) represents the horizontal and vertical cross divergence in the divergence matrix of the three-dimensional coordinate point (x,y,z), and HH((x,y,z) represents the horizontal divergence in the divergence matrix of the three-dimensional coordinate point (x,y,z).

[0121] θ sun The solar altitude angle represents the moment when multi-source remote sensing data is acquired;

[0122] As an embodiment of the present invention, if the pixel coordinates corresponding to the normalized three-dimensional coordinate points do not have pixel values, then bilinear interpolation is used in the spectral fusion remote sensing image to obtain the pixel values ​​of the pixel coordinates.

[0123] S3: Extract the SAR polarization feature image and spectral index fusion image of the surface of the mapped 3D terrain structure model, perform multimodal gating fusion on the extracted images, and obtain the boundary probability image of the surface of the 3D terrain structure model.

[0124] Extract SAR polarization feature images and spectral index fusion images from the surface of the mapped 3D terrain structure model, including:

[0125] Using a vertical top-down projection method, the three-dimensional coordinate points on the surface of the mapped three-dimensional terrain structure model are projected along the Z-axis to generate an index image in a two-dimensional planar view. The SAR polarization features and spectral indices of the three-dimensional coordinate points are extracted and used as the pixel values ​​of the projected pixels in the index image, respectively, to construct a SAR polarization feature image and a spectral index fusion image of the surface of the mapped three-dimensional terrain structure model.

[0126] Specifically, the polarization characteristics of the three-dimensional coordinate point (x,y,z) are HV(x,y,z) / HH(x,y,z), θ((x,y,z), where θ(x,y,z) is the emission angle of the horizontally polarized electromagnetic wave emitted by the SAR towards the two-dimensional geographic coordinate position corresponding to the three-dimensional coordinate point (x,y,z). The spectral indices of the three-dimensional coordinate point (x,y,z) are S1((x,y,z), S2((x,y,z), S3((x,y,z)).

[0127] S1(x,y,z), S2(x,y,z), and S3(x,y,z) are the vegetation index, water index, and soil moisture index of the three-dimensional coordinate point (x,y,z) respectively.

[0128] The pixel position of the three-dimensional coordinate point (x, y, z) in the indexed image is ((x, y).

[0129] Multimodal gating fusion is performed on the extracted images to obtain boundary probability images of the surface of the 3D terrain structure model, including:

[0130] Calculate the fusion gating factor of the SAR polarization feature image and the spectral index fused image. The formula for calculating the fusion gating factor is as follows:

[0131] g=δ(W SAR ·I SAR +W Spec ·I Spec );

[0132] Where g represents the fusion gating factor, I SAR Represents SAR polarization feature image, I Spec W represents the spectral index fusion image. SAR W represents the SAR weight matrix. Spec δ(·) represents the spectral index weight matrix, and δ(·) represents the Sigmoid activation function.

[0133] The fusion gating factor is used to fuse the SAR polarization feature image and the spectral index fusion image to obtain fused image I:

[0134] I=g⊙tanh(W SAR ·I SAR)+(Ug)⊙tanh(W Spec ·I Spec );

[0135] Where U represents a unit element matrix with all elements being 1 and the same scale as the indexed image, and ⊙ represents element-wise multiplication;

[0136] The fused image I is input into a convolution classifier to generate the boundary probability of each pixel in the fused image I. The boundary probability is used as the pixel value to construct the edge probability image.

[0137] As an embodiment of the present invention, the index image is in the form of a pixel matrix with A rows and B columns, and the SAR polarization feature image I SAR The SAR polarization feature image I is a multi-channel pixel matrix in the form of a 2-channel A-row B-column matrix. SAR The pixel value of the middle pixel is a two-dimensional polarization feature, and the spectral index is fused into the image I. Spec It is a multi-channel pixel matrix with 3 channels, A rows and B columns, where the 3 channels correspond to 3 spectral indices, W SAR The weight matrix is ​​in the form of an A-row, B-column matrix with two channels, W. Spec If W is a 3-channel A-row B-column weight matrix, then... SAR ·I SAR W Spec ·I Spec It is a single-channel A-row B-column pixel matrix, where A represents the number of 3D coordinate points of the 3D terrain structure model in the east-west direction, and B represents the number of 3D coordinate points of the 3D terrain structure model in the north-south direction.

[0138] S4: Extract coarse-grained farmland boundaries from the boundary probability image and extract slope images from the 3D terrain structure model. Combine the slope information of pixels in the slope image to constrain and optimize the coarse-grained farmland boundaries, and obtain smooth farmland boundary recognition results.

[0139] Extracting coarse-grained farmland boundaries from the boundary probability image includes:

[0140] Set a boundary probability threshold, convert the boundary probability image into a binary image, perform morphological dilation-erosion operation on the binary image, and remove isolated pixels and small breaks;

[0141] Connectivity analysis is performed on the binary image after morphological operations. All interconnected boundary regions are marked and connected to obtain multiple connected regions. The area of ​​each connected region is calculated, and the connected region with the largest area is selected as a coarse-grained farmland boundary candidate region. The edge point sequence in the coarse-grained farmland boundary candidate region is extracted using a contour tracking algorithm to form a closed curve. The closed curve is used as the coarse-grained farmland boundary.

[0142] As an embodiment of the present invention, if there are multiple field structures in the farmland area, connected regions with an area greater than a preset area threshold are selected, and the coarse-grained farmland boundary of each connected region is extracted. The contour tracking algorithm is the Suzuki boundary tracking algorithm.

[0143] Slope images are extracted from the 3D terrain structure model. Based on the slope information of the pixels in the slope images, constraint optimization is performed on the coarse-grained farmland boundary, including:

[0144] Calculate the slope of the three-dimensional coordinate points in the three-dimensional terrain structure model, and map the slope of the three-dimensional coordinate points to an index image to form a slope image. The index image is a two-dimensional planar view of the three-dimensional terrain structure model, and the image size of the slope image is consistent with that of the boundary probability image.

[0145] In this embodiment of the invention, the slope of the three-dimensional coordinate point (x, y, z) is:

[0146]

[0147] Where β(x,y,z) represents the slope of the three-dimensional coordinate point (x,y,z), Δ1 represents the width of the coordinate points of adjacent three-dimensional coordinate points in the longitude direction in the three-dimensional terrain structure model, Δ2 represents the width of the coordinate points of adjacent three-dimensional coordinate points in the dimensional direction in the three-dimensional terrain structure model, and Z(x+1,y) represents the elevation value of the two-dimensional geographic coordinate ((x+1,y)) in the three-dimensional terrain structure model.

[0148] The pixel coordinates of the three-dimensional coordinate point (x, y, z) in the slope image are ((x, y), and the pixel value is β(x, y, z);

[0149] A farmland boundary energy function based on slope constraints is constructed. The farmland boundary energy function takes the farmland boundary as input and the farmland boundary energy as output. The farmland boundary energy includes internal energy characterizing the smoothness and continuity of the farmland boundary, boundary image energy driving the farmland boundary to move closer to the boundary probability position, and slope energy penalizing the crossing of steep slope areas.

[0150] The coarse-grained farmland boundary is used as the input to the farmland boundary energy function. A gradient descent algorithm is employed to evolve the coarse-grained farmland boundary until the boundary evolution result meets the convergence condition. The current boundary evolution result is then output as a smoothed farmland boundary identification result, and this result is mapped onto a two-dimensional planar view of the three-dimensional terrain structure model. In one embodiment of the invention, the convergence condition is that the change in farmland boundary energy is less than a preset energy change threshold.

[0151] Specifically, the expression for the coarse-grained farmland boundary Q0 extracted from the boundary probability image is:

[0152]

[0153] Where Q0(r) represents the pixel coordinates of the r-th segment in the coarse-grained farmland boundary Q0 in the boundary probability image. r represents the parameter for normalizing the length of farmland boundaries;

[0154] The expression for the farmland boundary energy function is:

[0155] E(Q t )=w1·E1((Q t )+w2·E2(Q t ,P)+w3·E2(Q t ,β);

[0156]

[0157] Among them, Q t E(·) represents the t-th boundary evolution result of the coarse-grained farmland boundary, and E(Q) represents the farmland boundary energy function. t ) represents Q t The farmland boundary energy, E1(Q) t ) represents Q t The internal energy, E2((Q) t P) represents Q t Boundary image energy, E2(Q t ,β) represents Q t The slope energy is P, which represents the boundary probability image, and β represents the slope image; w1, w2, and w3 represent the weights of the internal energy, boundary image energy, and slope energy, respectively; the boundary image energy is preceded by a negative sign to maximize the boundary probability, that is, to encourage the boundary to approach the region with a strong boundary probability response.

[0158] Q t (r) represents Q t The pixel coordinates of the r-th segment in the boundary probability image. Q represents t The first-order term of (r), Q represents t The second-order term of (r);

[0159] The first-order term is the elasticity term, which represents the sum of the squares of the line segment lengths between boundary points, corresponding to the degree of stretching or compression of the boundary. When this term is small, it encourages equidistant distribution and avoids excessive dispersion between points, which is conducive to forming a boundary with a compact structure and good contour closure. The second-order term is the smoothness term, which reflects the local curvature of each boundary point relative to its neighbors, encouraging the geometric continuity and smoothness of the boundary, and helping to avoid high-frequency oscillations in noisy regions.

[0160] P(Q t (r) represents the pixel coordinates Q in the boundary probability image. t (boundary probability at (r);

[0161] β(Q t (r) represents the pixel coordinates Q in the slope image. t The slope at (r);

[0162] τ represents the slope suppression coefficient. Denotes the activator, if β(Q) t (r)) is greater than 25 degrees, then otherwise dr represents the differential representation of parameter r.

[0163] Example 2:

[0164] like Figure 2 The image shown is a spectral index fusion image based on multi-source remote sensing data fusion in the case of multiple farmlands.

[0165] Example 3:

[0166] Taking vegetation index as an example, such as Figure 3 The experiment shows a comparison of vegetation index changes before and after shading compensation. The experimental area is a typical mountainous and hilly farmland. The traditional uncompensated vegetation index method and the vegetation index calculation method with shading compensation coefficient proposed in this invention are compared. The comparison results show that:

[0167] 1. In the shaded area of ​​the slope, the local variance of the traditional vegetation index is 0.041, while it is reduced to 0.017 after compensation using the method of this invention, indicating that the vegetation index is smoother and has higher stability in areas with uneven light.

[0168] 2. The traditional vegetation index boundary gradient has a mean value of 0.11, while the compensation method increases it to 0.22, which enhances the contrast of land cover boundaries and facilitates farmland boundary identification.

[0169] 3. Compared with the measured vegetation index on the ground, the root mean square error between the traditional vegetation index and the measured vegetation index on the ground is 0.082, while the root mean square error after compensation by the present invention is 0.046, showing higher consistency in the field.

[0170] Example 4:

[0171] This embodiment provides a large-scale farmland boundary recognition system, which includes a server and a data acquisition device. The server includes a multimodal gating fusion module and a farmland boundary recognition module.

[0172] The multimodal gating fusion module is used to extract SAR polarization feature images and spectral index fusion images of the surface of the mapped three-dimensional terrain structure model, and to perform multimodal gating fusion on the extracted images to obtain the boundary probability image of the surface of the three-dimensional terrain structure model.

[0173] The farmland boundary recognition module is used to extract coarse-grained farmland boundaries from the boundary probability image and extract slope images from the three-dimensional terrain structure model. Combining the slope information of the pixels in the slope image, the coarse-grained farmland boundaries are constrained and optimized to obtain smooth farmland boundary recognition results.

[0174] The data acquisition device is used to collect multi-source remote sensing data of farmland areas, use the digital elevation model of farmland areas to perform terrain perception of farmland areas, generate a three-dimensional terrain structure model of farmland areas, perform adaptive spectral fusion on multispectral remote sensing images to obtain spectral fusion remote sensing images, and map the pixel values ​​of pixels in the spectral fusion remote sensing images to three-dimensional coordinate points on the surface of the three-dimensional terrain structure model based on the collinearity equation correction principle.

[0175] To implement the large-scale farmland boundary identification method of multi-source remote sensing data fusion in Example 1.

[0176] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0177] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0178] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0179] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for large-scale farmland boundary recognition based on multi-source remote sensing data fusion, characterized in that, The method comprises: S1: collecting multi-source remote sensing data of the farmland area, wherein the multi-source remote sensing data comprises multi-spectral remote sensing images, SAR images and a digital elevation model, performing terrain perception on the farmland area by using the digital elevation model of the farmland area, and generating a three-dimensional terrain structure model of the farmland area; The pixel coordinates of the SAR image are three-dimensional coordinate points on the surface of the three-dimensional terrain structure model, and the pixel value is a scattering matrix of the three-dimensional coordinate points; S2: adaptively fusing the multi-spectral remote sensing images to obtain a spectral fusion remote sensing image, and mapping the pixel value of the pixel in the spectral fusion remote sensing image to the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model based on the collinear equation correction principle; The pixel value of the pixel in the spectral fusion remote sensing image comprises a vegetation index, a water body index and a soil moisture index; S3: extracting a SAR polarization feature image and a spectral index fusion image on the surface of the mapped three-dimensional terrain structure model, performing multi-modal gated fusion on the extracted images to obtain a boundary probability image on the surface of the three-dimensional terrain structure model, and the pixel value of the pixel in the boundary probability image is the probability that the pixel is a farmland boundary; S4: extracting a coarse-grained farmland boundary from the boundary probability image, extracting a slope image in the three-dimensional terrain structure model, and combining the slope information of the pixel in the slope image to constrain and optimize the coarse-grained farmland boundary to obtain a smooth farmland boundary recognition result.

2. The method of claim 1, wherein the method is characterized by, Collecting multi-source remote sensing data of the farmland area, performing terrain perception on the farmland area by using the digital elevation model of the farmland area, and generating a three-dimensional terrain structure model of the farmland area, comprising: extracting a digital elevation model of the farmland area, wherein the digital elevation model records the elevation values of part of the two-dimensional geographic coordinates in the farmland area, and the two-dimensional geographic coordinates and the corresponding elevation values are taken as point cloud data of the farmland area; performing filtering processing on the point cloud data by using a median filtering method; searching for the nearest K neighbor point cloud data for each point cloud data to form a local plane, and calculating the normal vector of the local plane as the normal vector of the point cloud data; extracting the normal vectors of all point cloud data to form a normal vector field, converting the normal vector field into divergence values of a gradient field by using a divergence calculation method, constructing a Poisson equation of the three-dimensional surface of the farmland area, and iteratively solving the Poisson equation by using a conjugate gradient algorithm to obtain an implicit representation of the three-dimensional surface of the farmland area; converting the implicit representation of the three-dimensional surface of the farmland area obtained by solving the Poisson equation into an explicit representation of a triangular mesh model by using a gridding processing method; based on the triangular mesh model, extracting all two-dimensional geographic coordinates and corresponding elevation values of the farmland area to form a complete three-dimensional terrain structure model of the farmland area, wherein the three-dimensional terrain structure model of the farmland area describes the elevation values of the farmland area at different two-dimensional geographic coordinates, and the two-dimensional geographic coordinates and the corresponding elevation values are taken as three-dimensional coordinate points on the surface of the three-dimensional terrain structure model.

3. The method of claim 1, wherein the method is characterized by, Adaptively fusing the multi-spectral remote sensing images to obtain a spectral fusion remote sensing image, comprising: the multi-spectral remote sensing image is composed of spectral remote sensing images of multiple bands; The adaptive spatial attention calculation is performed on spectral remote sensing images of different wave bands to obtain spatial attention weights of pixels in the spectral remote sensing images, the pixel values of the pixels in the spectral remote sensing images are weighted to obtain weighted pixel values of pixels in a multispectral remote sensing image in different wave bands, vegetation indexes, water body indexes and soil moisture indexes of different pixels in the multispectral remote sensing image are calculated, the vegetation indexes, the water body indexes and the soil moisture indexes are taken as pixel values of pixels in the multispectral remote sensing image, and a spectral fusion remote sensing image is constructed.

4. The method of claim 3, wherein the method is characterized by, Mapping the pixel values of the pixels in the spectral fusion remote sensing image to three-dimensional coordinate points on a surface of a three-dimensional terrain structure model comprises: Normalizing the three-dimensional coordinate points on the surface of the three-dimensional terrain structure model to obtain normalized three-dimensional coordinate points; Mapping the normalized three-dimensional coordinate points by using an RPC model to obtain pixel coordinates of the normalized three-dimensional coordinate points in the spectral fusion remote sensing image, and extracting pixel values of the pixel coordinates as spectral index fusion values of the three-dimensional coordinate points associated with the surface of the three-dimensional terrain structure model.

5. The method of claim 1, wherein the method is a multi-source remote sensing data fusion based large-scale farmland boundary recognition method, characterized in that, Extracting a SAR polarization feature image and a spectral index fusion image of the mapped surface of the three-dimensional terrain structure model comprises: Projecting and mapping the three-dimensional coordinate points on the mapped surface of the three-dimensional terrain structure model along the Z-axis direction in a vertical top view projection mode to generate an index image in a two-dimensional plane view, and extracting SAR polarization features and spectral indexes of the three-dimensional coordinate points as pixel values of projected and mapped pixels in the index image to construct the SAR polarization feature image and the spectral index fusion image of the mapped surface of the three-dimensional terrain structure model.

6. The method of claim 5, wherein the method is a multi-source remote sensing data fusion based large-scale farmland boundary recognition method, characterized in that, Performing multi-modal gating fusion on the extracted images to obtain a boundary probability image of the surface of the three-dimensional terrain structure model comprises: Calculating a fusion gating factor of the SAR polarization feature image and the spectral index fusion image, and the calculation formula of the fusion gating factor is: ； wherein, denotes a fusion gating factor, denotes a SAR polarization feature image, denotes a spectral index fusion image, denotes a SAR weight matrix, denotes a spectral index weight matrix, denotes a Sigmoid activation function; Performing fusion calculation on the SAR polarization feature image and the spectral index fusion image by using the fusion gating factor to obtain a fusion image I: ； wherein U denotes a unit element matrix whose elements are all 1 and whose scale is consistent with the index image, denotes element-wise multiplication; Inputting the fusion image I into a convolution classifier to generate a boundary probability of each pixel in the fusion image I, taking the boundary probability as a pixel value, and constructing an edge probability image.

7. The method of claim 6, wherein the method is a multi-source remote sensing data fusion based large-scale farmland boundary recognition method, characterized in that, Extracting a coarse-grained farmland boundary from the boundary probability image comprises: Setting a boundary probability threshold, converting the boundary probability image into a binary image, performing morphological dilation-erosion operation on the binary image to remove isolated pixels and small broken sections; Performing connected region analysis on the binary image after morphological operation to label all connected boundary regions and connect them to obtain a plurality of connected regions, calculating the area of each connected region, screening the largest connected region as a coarse-grained farmland boundary candidate region, using a contour tracking algorithm to extract an edge point sequence in the coarse-grained farmland boundary candidate region to form a closed curve, and taking the closed curve as the coarse-grained farmland boundary.

8. The method of claim 7, wherein the method is a multi-source remote sensing data fusion based large-scale farmland boundary recognition method, characterized in that, Extracting a slope image in the three-dimensional terrain structure model and combining slope information of pixels in the slope image to constrain and optimize the coarse-grained farmland boundary comprises: The slope of a three-dimensional coordinate point in the three-dimensional terrain structure model is calculated, and the slope of the three-dimensional coordinate point is mapped to an index image to form a slope image, the index image being a two-dimensional planar view of the three-dimensional terrain structure model, and the image size of the slope image being consistent with the boundary probability image; A farmland boundary energy function based on slope constraint is constructed, the farmland boundary energy function taking a farmland boundary as input and outputting farmland boundary energy, the farmland boundary energy including internal energy representing the smoothness and continuity of the farmland boundary, boundary image energy driving the farmland boundary to be close to a position with high boundary probability, and slope energy punishing the crossing of an abrupt slope region; The coarse-grained farmland boundary is taken as input of the farmland boundary energy function, and a gradient descent algorithm is used to perform boundary evolution on the coarse-grained farmland boundary until the boundary evolution result meets a convergence condition, and the current boundary evolution result is output as a smooth farmland boundary recognition result, and the farmland boundary recognition result is mapped to a two-dimensional planar view of the three-dimensional terrain structure model.

9. A large-scale farmland boundary recognition system, characterized by, The large-scale farmland boundary recognition system includes a server and a data acquisition device, the server including a multi-modal gating fusion module and a farmland boundary recognition module: The multi-modal gating fusion module is configured to extract a SAR polarization feature image and a spectral index fusion image of the surface of the mapped three-dimensional terrain structure model, and perform multi-modal gating fusion on the extracted images to obtain a boundary probability image of the surface of the three-dimensional terrain structure model; The farmland boundary recognition module is configured to extract a coarse-grained farmland boundary from the boundary probability image, extract a slope image from the three-dimensional terrain structure model, and perform constraint optimization on the coarse-grained farmland boundary in combination with the slope information of the pixels in the slope image to obtain a smooth farmland boundary recognition result; The data acquisition device is configured to acquire multi-source remote sensing data of a farmland region, perform terrain perception on the farmland region by using a digital elevation model of the farmland region to generate a three-dimensional terrain structure model of the farmland region, perform adaptive spectral fusion on a multi-spectral remote sensing image to obtain a spectral fusion remote sensing image, and map the pixel value of a pixel in the spectral fusion remote sensing image to a three-dimensional coordinate point on the surface of the three-dimensional terrain structure model based on a collinearity equation correction principle; to realize a large-scale farmland boundary recognition method based on multi-source remote sensing data fusion according to any one of claims 1-8.

Images