A power transmission line tower foundation land disturbance automatic extraction method based on a full convolutional neural network

Abstract

The present application relates to the technical field of power line geographical environment monitoring, and particularly relates to a transmission line tower foundation land disturbance automatic extraction method based on a full convolutional neural network, comprising the following steps: S1, acquiring multi-temporal remote sensing images of a transmission line tower foundation, combining ground object difference and boundary continuity constraints, extracting a tower foundation region satisfying a change threshold, and generating a structure guide candidate graph; S2, extracting two-phase boundary features using a double-flow network, constructing a disturbance boundary response graph, and generating a disturbance probability graph in combination with the candidate graph and a current image; S3, based on the disturbance probability graph, removing misjudgment regions, combining edge compensation to repair missing boundaries, and outputting an optimized disturbance region vector result. The present application, by fusing tower foundation structure priori and boundary response enhancement mechanisms, realizes high-precision extraction of a disturbance region and effective suppression of misjudgment regions, and significantly improves the integrity and spatial consistency of a disturbance boundary.

Description

Technical Field

[0001] This invention relates to the field of power line geographical environment monitoring technology, and in particular to an automatic extraction method for land disturbance of transmission line tower foundations based on a fully convolutional neural network. Background Technology

[0002] With the large-scale construction of transmission lines and the continuous evolution of the natural environment, the automatic monitoring and change identification of land disturbance around the tower base has become a key task in power grid operation and maintenance management. The widespread availability of high-resolution remote sensing images has provided a rich data foundation for the extraction of land disturbance areas, while image segmentation methods based on computer vision have gradually become the mainstream technical path for land disturbance detection. Especially in areas with significant structural features, such as transmission tower bases, accurate identification of disturbance patches caused by construction is of great significance for environmental protection, operation and maintenance inspections, and patch updates.

[0003] However, existing methods generally rely on texture or spectral information from single-phase images, lacking prior modeling of the tower base structure. This makes it difficult to distinguish disturbed areas from other undisturbed features with high reflectivity or boundary characteristics. Meanwhile, some dual-phase change detection methods, while incorporating temporal information, do not consider the spatial structural stability of the tower base area, leading to a tendency for the model to misjudge occlusion, high-contrast buildings, or noisy areas. The disturbance extraction results have blurred boundaries and poor spatial consistency, making it difficult to meet the needs of refined power facility supervision. Summary of the Invention

[0004] This invention provides an automatic method for extracting land disturbances at transmission line tower foundations based on a fully convolutional neural network. It integrates tower foundation structure priors, dual-temporal boundary response enhancement, and disturbance probability suppression mechanisms to construct a structure-guided dual-flow feature extraction and spatial consistency optimization framework. This effectively achieves accurate extraction of disturbed areas and removal of pseudo-disturbances, improves the completeness of disturbance boundary representation, and enhances the vector representation capability of positioning results. It is applicable to automatic patch updating and intelligent inspection applications during the construction and operation phases of transmission line projects.

[0005] An automatic method for extracting land disturbance of transmission line tower foundations based on a fully convolutional neural network includes the following steps: S1. Acquire multi-temporal remote sensing image sequences including the base of the transmission line tower, and based on the differences in ground features before and after the construction of the tower base, combined with the structural boundary constraint rules, extract the tower base structural regions that meet the conditions of structural boundary continuity and ground feature change threshold, and construct a candidate region map of structure-guided disturbance. S2 employs a cross-temporal boundary enhancement mechanism and a dual-stream feature extraction network to extract the boundary response features of the perturbation region in the current temporal phase and the reference temporal phase, respectively, and performs structure alignment and fusion to construct a time-sensitive perturbation boundary response map. It also jointly models the structure-guided perturbation candidate region map with the current temporal remote sensing image to output a perturbation probability map. S3, based on the perturbation probability map, uses a pseudo-perturbation region suppression mechanism to remove misjudged regions caused by the tower base structure edge and high-reflectivity ground objects, and repairs the missing perturbation boundary through a high-frequency edge compensation strategy, outputting the perturbation region vector result after spatial consistency optimization.

[0006] Optionally, S1 includes: S11, acquire a multi-temporal remote sensing image sequence covering the same transmission line tower foundation area, including temporal images before tower foundation construction. Phase images after tower foundation construction And the time-lapse images before the construction of the tower foundation. Phase images after tower foundation construction Preprocessing is performed, including geometric correction and radiometric normalization, to unify resolution and registration reference, resulting in registered image pairs. ; S12, after obtaining the registered image pair Then, by calculating the difference in ground feature changes of corresponding pixels in the temporal images before and after construction, the change range of each pixel is quantified, and a ground feature change difference map is generated. S13. Based on the geometric morphological features of the tower base structure in the image (highlight, rectangle, high reflectance hardened surface), combined with the historical tower location coordinate point set, local windows are extracted from the preprocessed tower base construction phase image. Edge detection (Canny operator) is performed on each window and morphological constraints are applied to generate a candidate set of tower base structure boundaries. S14. On the land feature change difference map, apply a land feature change difference threshold constraint to the candidate set of tower base structure boundaries, and extract the tower base structure regions that satisfy both land feature change difference and boundary continuity. ; S15 will be the tower base structure area that satisfies the differences in ground features and the continuity of boundaries. Converted into a candidate region map of structure-guided perturbation ,in, For a set of nodes, It is an edge set.

[0007] Optionally, S11 includes: S111, Obtain temporal images of the tower foundation before construction, covering the same tower foundation area. Phase images after tower foundation construction ; S112, regarding the acquired temporal images prior to tower foundation construction. Phase images after tower foundation construction Perform geometric corrections to place them in the same map projection coordinate system; S113, using the geometrically corrected post-construction phase image of the tower base as a reference, perform radiometric normalization processing on the geometrically corrected pre-construction phase image of the tower base. S114, using the radiometrically normalized post-construction temporal image of the tower foundation as a reference, performs sub-pixel-level registration on the radiometrically normalized pre-construction temporal image of the tower foundation, and resamples it to a uniform spatial resolution, outputting the registered image pair. .

[0008] Optionally, S13 includes: S131, the preprocessed time-phase image after the tower foundation construction is recorded as Let the set of coordinate points of the historical tower sites be... ,in, For the first Two-dimensional coordinate vectors of the base points of the tower. The x-axis is... The vertical axis is , The total number of historical tower location coordinates, in Center-trimmed window that adapts to resolution ; S132, in each window Within the process, the image grayscale values ​​are normalized and then processed according to a preset high reflectance threshold. Extract image grayscale values ​​that meet the normalization criteria and are greater than or equal to The region, thereby generating a high-reflectivity mask. ; S133, Apply Gaussian smoothing to the preprocessed time-phase image of the tower base construction within the window. The gradient is then calculated, and the edge set is obtained through nonmaximum suppression and double-threshold hysteresis connection. And calculate the edge connectivity. ; S134, in a high-reflectivity mask With edge set Extract candidate contour sets from the intersection of Calculate the area for each candidate contour. Minimum circumscribed rectangle size ( ),perimeter And define morphological constraints, including aspect ratio. Rectangularity Edge filling rate Candidate contours that simultaneously satisfy the candidate conditions are retained as the candidate set of the tower base structure boundary. The candidate conditions are expressed as follows: ; in, For the first The first window One candidate contour, , , , , These are the corresponding morphology and continuity discrimination thresholds, This represents the candidate set of tower base structure boundaries constrained by morphological structure.

[0009] Optionally, S14 includes: S141, in each candidate set of tower base structure boundaries Within, the mean change range of its value on the land cover change difference map is statistically analyzed. ; S142, Set the threshold for ground feature change intensity Preserve the mean change range Greater than or equal to the ground feature change intensity threshold The region, as a set of regions constrained by morphology and intensity of change. ; S143, sets the regions that have passed the morphological and change intensity constraints. Apply boundary connectivity constraints again, extract the effective tower base structure region, and finally output the tower base structure region. , is represented as: .

[0010] Optionally, S2 includes: S21, the registered tower base construction phase image Phase images before tower foundation construction Each input is fed into a boundary feature extraction network with shared structure to extract the corresponding boundary response feature maps. , ; S22, through the , A boundary difference map is constructed by performing pixel-by-pixel interpolation to identify the spatial variation boundary of the tower base disturbance area; S23 takes the boundary difference map, the structure-guided perturbation candidate region map, and the current temporal remote sensing image as input, performs joint feature fusion and prediction, and outputs a perturbation probability map.

[0011] Optionally, S21 includes: S211, Time-phase images of the registered tower base after construction Phase images before tower foundation construction Normalization processing is performed to obtain normalized time-phase images after tower base construction and before tower base construction; S212, the normalized post-construction phase image and pre-construction phase image of the tower base are input into the structure-shared boundary feature extraction network to obtain the boundary response feature map.

[0012] Optionally, S22 includes: S221, Calculate the boundary response difference between the boundary response feature map of the time-phase image after tower foundation construction and the boundary response feature map of the time-phase image before tower foundation construction at each pixel location, and obtain the boundary difference map. ; S222, Based on whether the pixel values ​​in the boundary difference map exceed the boundary disturbance intensity threshold. Generate a preliminary perturbation region mask. , is represented as: ; in, This represents the candidate boundary disturbance region. S223, Masking of the initial disturbance region All connected regions are filtered by size and shape, retaining those with an area greater than [missing information]. And the boundary continuity is greater than The region is used as the final disturbance boundary region. , is represented as: ; in, This is the set of all connected regions in the boundary perturbation mask. Connected region The number of pixels (area). For the region Edge connectivity, The minimum disturbance area threshold. This is the minimum edge connectivity threshold.

[0013] Optionally, S23 includes: S231, the boundary difference map, the candidate region map of structure-guided perturbation and the current time-phase remote sensing image (time-phase image after tower base construction) are stitched together in the channel dimension into a unified feature tensor, and the fused features are extracted by a convolutional encoder; S232 uses a decoder to upsample the fused features and predict the perturbation distribution, outputting a perturbation probability map.

[0014] Optionally, S3 includes: S31, using the perturbation probability map and local structural strength, eliminates the spurious perturbation areas caused by the tower base edge or high-reflectivity background, as shown below: ; in, To provide a mask for the real perturbation region retained after removing false perturbations. For the pixels in the output perturbation probability map The probability value of the disturbance. As a probability threshold, Boundary difference map medium pixel The local grayscale standard deviation, This represents the suppression threshold for the intensity of local structural disturbances. S32, through high-frequency edge detection of the image, compensates for the broken and missing parts of the boundary of the perturbation region, and obtains the mask of the missing region of the perturbation boundary. ; S33, a mask that merges the actual perturbation region mask with a mask that compensates for the missing perturbation boundary region. Generate vector perturbation results after spatial consistency optimization. .

[0015] The beneficial effects of this invention are: This invention, by introducing a dual mechanism of prior knowledge of the tower base structure and constraint of differences in ground features, can efficiently screen candidate regions of tower base disturbance with continuous boundaries and significant variation amplitude from multi-temporal remote sensing images, construct a structure-guided disturbance candidate region map, realize error constraint on spatial guidance and semantic segmentation of disturbance location, effectively reduce background noise and interference from non-target regions, and improve the accuracy and controllability of candidate region extraction.

[0016] This invention, by designing a cross-temporal boundary enhancement mechanism and a dual-stream feature extraction network for structural sharing, extracts and aligns boundary response features before and after construction, achieving time-sensitive disturbance boundary modeling. By combining structural maps with current temporal images for feature fusion, it can fully capture the semantic differences and spatial structural changes of tower base disturbances, improve the accurate positioning capability of disturbance areas, and effectively characterize the disturbance confidence distribution through a disturbance probability map, thereby enhancing the robustness and adaptability of the model.

[0017] This invention, by setting up a pseudo-disturbance region suppression mechanism and a high-frequency edge compensation strategy, and combining multi-source information such as disturbance probability, local structural strength, and high-frequency response, dynamically eliminates misjudged regions caused by tower base edge reflection or complex background, while repairing boundary breaks and missing parts, achieving spatial consistency optimization, and finally outputting vectorized disturbance results with clear boundaries, accurate positioning, and good structural representation characteristics, meeting the application requirements of intelligent monitoring of transmission line tower base disturbances and automatic patch updating. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the extraction method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the construction of a candidate region map for structure-guided perturbation according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] like Figures 1-2 As shown, an automatic method for extracting land disturbance of transmission line tower foundations based on a fully convolutional neural network includes the following steps: S1. Acquire multi-temporal remote sensing image sequences including the base of transmission line towers. Based on the differences in ground features before and after the construction of the tower base, and combined with the structural boundary constraint rules, extract the tower base structural regions that meet the conditions of structural boundary continuity and ground feature change threshold. Construct a candidate region map of structural guided disturbance, which is used for semantic segmentation region guidance and error constraint. S2 employs a cross-temporal boundary enhancement mechanism and a dual-stream feature extraction network to extract the boundary response features of the perturbation region in the current temporal phase and the reference temporal phase, respectively, and performs structure alignment and fusion to construct a time-sensitive perturbation boundary response map. It also jointly models the structure-guided perturbation candidate region map with the current temporal remote sensing image to output a perturbation probability map. S3, based on the perturbation probability map, uses a pseudo-perturbation region suppression mechanism to remove misjudged regions caused by the tower base structure edge and high-reflectivity ground objects, and repairs the missing perturbation boundary through a high-frequency edge compensation strategy, outputting the perturbation region vector result after spatial consistency optimization.

[0022] S1 includes: S11, acquire a multi-temporal remote sensing image sequence covering the same transmission line tower foundation area, including temporal images before tower foundation construction. Phase images after tower foundation construction And the time-lapse images before the construction of the tower foundation. Phase images after tower foundation construction Preprocessing is performed, including geometric correction and radiometric normalization, to unify resolution and registration reference, resulting in registered image pairs. ; S12, after obtaining the registered image pair Then, by calculating the difference in ground feature changes between corresponding pixels in the temporal images before and after construction, the change amplitude of each pixel is quantified, and a ground feature change difference map is output, represented as follows: ; in, For pixels The difference in land cover changes, , These are pre-processed temporal images before and after tower foundation construction. To prevent tiny constants with a denominator of zero; S13. Based on the geometric morphological features of the tower base structure in the image (highlight, rectangle, high reflectance hardened surface), combined with the historical tower location coordinate point set, local windows are extracted from the preprocessed tower base construction phase image. Edge detection (Canny operator) is performed on each window and morphological constraints are applied to generate a candidate set of tower base structure boundaries. S14. On the land feature change difference map, apply a land feature change difference threshold constraint to the candidate set of tower base structure boundaries, and extract the tower base structure regions that satisfy both land feature change difference and boundary continuity. ; S15 will be the tower base structure area that satisfies the differences in ground features and the continuity of boundaries. Converted into a candidate region map of structure-guided perturbation ,in, For a set of nodes, each node This represents a candidate perturbation block. For edge set, edge It indicates the spatial similarity or boundary connection between adjacent regions.

[0023] S11 includes: S111, Obtain temporal images of the tower foundation before construction, covering the same tower foundation area. Phase images after tower foundation construction ; S112, regarding the acquired temporal images prior to tower foundation construction. Phase images after tower foundation construction Perform geometric correction to place them in the same map projection coordinate system, as shown below: ; in, Let be the homogeneous coordinates of the pixels on the original image plane. These are the geometrically corrected homogeneous coordinates of the pixels. This is the homography / projection transformation matrix; S113, using the geometrically corrected post-construction temporal imagery of the tower base as a reference, performs radiometric normalization processing on the geometrically corrected pre-construction temporal imagery of the tower base, as follows: ; ; ; ; in, , These are the radiometrically normalized images of the tower foundation before construction, taken at the [number]th [year]. The pixel value of the band and the time-phase image after the construction of the tower base in the [number]th [year] Pixel values ​​of the band, , These are the geometrically corrected temporal images of the tower base before construction, in the [number]th [year]. The pixel values ​​of the band and the temporal images after the construction of the tower base are in the first... Pixel values ​​of the band, , The first The normalization coefficient of the band, , These are the temporal images of the tower base before construction within the ROI. The mean and standard deviation of the band. , These are the temporal images of the tower base after construction within the ROI. Mean and standard deviation of the band; S114, using the radiometrically normalized post-construction temporal image of the tower foundation as a reference, performs sub-pixel-level registration on the radiometrically normalized pre-construction temporal image of the tower foundation, and resamples it to a uniform spatial resolution, outputting the registered image pair. Specifically, it includes: (1) Solving the registration model: Obtain the set of corresponding points through feature point matching. Estimate the parameters of the registration model such as affine / thin plate splines. , is represented as: ; in The first phase image of the tower foundation before construction, after radiation normalization processing. Feature points The first phase image of the tower base after construction, after radiation normalization processing. One feature point, , The linear transformation and translation parameters to be estimated; (2) Resampling unification: Set the target pixel size Interpolation is performed on the same grid for two images, as shown below: ; ; in, Based on Map the pre-construction imagery to the coordinates of the reference imagery. To be based on target resolution Resampling; (3) Output: Registered image pairs with uniform projection, uniform radiometrics, uniform resolution, and strict pixel alignment. .

[0024] S13 includes: S131, the preprocessed time-phase image after the tower foundation construction is recorded as Let the set of coordinate points of the historical tower sites be... ,in, For the first Two-dimensional coordinate vectors of the base points of the tower. The x-axis is... The vertical axis is , The total number of historical tower location coordinates, in Center-trimmed window that adapts to resolution , is represented as: ; in, These are the cell coordinates within the local window. The horizontal pixel position. Vertical pixel position, , This represents half the window size (in pixels), corresponding to the horizontal and vertical ranges respectively. S132, in each window Within the process, the image grayscale values ​​are normalized and then processed according to a preset high reflectance threshold. Extract image grayscale values ​​that meet the normalization criteria and are greater than or equal to The region, thereby generating a high-reflectivity mask. , is represented as: ; ; in, For normalized image grayscale values, The image grayscale values ​​are the preprocessed images of the tower base after construction. , They are respectively in Minimum and maximum pixel values ​​within, To avoid tiny constants with a denominator of zero; High reflectivity threshold Click window To configure, it is represented as: (1) Collect normalized grayscale samples: for the window Normalized image gray values ​​of internal pixels Form a set, represented as: ; in, For window All A set; (2) Robust statistic estimation: Calculating the median Absolute deviation from median , is represented as: ; ; (3) High quantile protection: Calculation 95th percentile , is represented as: ; (4) Threshold synthesis: ; in, For the reason The robust standard deviation obtained from the conversion This is the magnification factor; S133, Apply Gaussian smoothing to the preprocessed time-phase image of the tower base construction within the window. The gradient is then calculated, and the edge set is obtained through nonmaximum suppression and double-threshold hysteresis connection. And calculate the edge connectivity. , is represented as: ; ; ; ; ; in, For pixels within the window The gradient component in the horizontal direction. For pixels within the window The gradient component in the vertical direction, For pixels gradient magnitude, For pixels gradient direction, , These represent the low and high thresholds for the Canny operator, respectively. The standard deviation of the Gaussian kernel. For edge set The first in Connected components For connected components The number of pixels; Low threshold of the Canny operator High threshold An adaptive quantile method is used to set the quantile value in each local window. First, calculate the gradient magnitude map. Then, based on its distribution, a dual threshold is adaptively set, expressed as: ; ; in, This is the set of gradient magnitudes for all pixels within the window. , These are the corresponding quantiles; S134, in a high-reflectivity mask With edge set Extract candidate contour sets from the intersection of Calculate the area for each candidate contour. Minimum circumscribed rectangle size ( ),perimeter And define morphological constraints, including aspect ratio. Rectangularity Edge filling rate Candidate profiles that simultaneously satisfy the candidate conditions are retained as the candidate set of the tower base structure boundary. The candidate conditions are expressed as follows: ; in, For the first The first window One candidate contour, , , , , These are the corresponding morphology and continuity discrimination thresholds, This is a candidate set of tower base structure boundaries constrained by morphological structure. ; ; .

[0025] S14 includes: S141, in each candidate set of tower base structure boundaries Within, the mean change range of its value on the land cover change difference map is statistically analyzed. , is represented as: ; in, The number of pixels within the region; S142, Set the threshold for ground feature change intensity Preserve the mean change range Greater than or equal to the ground feature change intensity threshold The region, as a set of regions constrained by morphology and intensity of change. ; Land feature change intensity threshold The setting is performed using the local fractional adaptive method, expressed as follows: ; in, This is the quantile threshold; S143, sets the regions that have passed the morphological and change intensity constraints. Apply boundary connectivity constraints again, extract the effective tower base structure region, and finally output the tower base structure region. , is represented as: .

[0026] S2 includes: S21, the registered tower base construction phase image Phase images before tower foundation construction Each input is fed into a boundary feature extraction network with shared structure to extract the corresponding boundary response feature maps. , ; S22, through the , A boundary difference map is constructed by performing pixel-by-pixel interpolation to identify the spatial variation boundary of the tower base disturbance area; S23 takes the boundary difference map, the structure-guided perturbation candidate region map, and the current temporal remote sensing image as input, performs joint feature fusion and prediction, and outputs a perturbation probability map.

[0027] S21 includes: S211, Time-phase images of the registered tower base after construction Phase images before tower foundation construction After normalization processing, the normalized temporal images after tower foundation construction and before tower foundation construction are obtained, represented as follows: ; ; in, , These are normalized temporal images after tower base construction and temporal images before tower base construction. This is the normalization function; normalization function Represented as: ; in, The normalized pixel value The original image in spatial position Band Channel The pixel value on For the first The average of each channel, For the first Standard deviation of each channel; S212, the normalized post-construction and pre-construction temporal images of the tower base are input into the structure-shared boundary feature extraction network to obtain the boundary response feature map, represented as: ; ; in, This is a boundary response feature map of the temporal image after the construction of the tower base. Boundary response feature map of pre-construction temporal imagery. For boundary feature extraction network; Boundary feature extraction network Specifically, it includes: (1) Backbone network feature extraction: The backbone network uses a lightweight CNN to extract multi-scale feature maps, represented as follows: ; ; in, , These are shallow and mid-layer features extracted from the backbone network, respectively. , These are the corresponding convolutional kernel weights. The input image is after normalization. , These are the corresponding convolution bias terms. It is the ReLU activation function; (2) Boundary Attention Enhancement Module: In To enhance boundary response using edge attention mechanisms, first compute the channel attention map. Spatial boundary weight map The final boundary response feature map is obtained, represented as: ; ; ; in, , These are the image height and image width, respectively. For the first in spatial attention fusion The weighting coefficients of the channels, This is the bias term in the spatial attention branch. This represents the number of channels.

[0028] S22 includes: S221, Calculate the boundary response difference between the boundary response feature map of the time-phase image after tower foundation construction and the boundary response feature map of the time-phase image before tower foundation construction at each pixel location, and obtain the boundary difference map. , is represented as: ; in, Boundary response feature map of temporal image after tower base construction at pixel location The channel vector, Boundary response feature map of temporal image before tower foundation construction at pixel location The channel vector; S222, Based on whether the pixel values ​​in the boundary difference map exceed the boundary disturbance intensity threshold. Generate a preliminary perturbation region mask. , is represented as: ; in, This represents the candidate boundary disturbance region. Boundary disturbance strength threshold Adaptive percentile setting is used, expressed as: ; in, For quantile parameters; S223, Masking of the initial disturbance region All connected regions are filtered by size and shape, retaining those with an area greater than [missing information]. And the boundary continuity is greater than The region is used as the final disturbance boundary region. , is represented as: ; in, This is the set of all connected regions in the boundary perturbation mask. Connected region The number of pixels (area). For the region Edge connectivity, The minimum disturbance area threshold. The minimum edge connectivity threshold; Minimum perturbation area threshold Represented as: ; in, The minimum identifiable area for tower base disturbance. Image spatial resolution; Minimum edge connectivity threshold Represented as: ;; ; in, Connected region Edge connectivity, For the region The number of boundary pixels, , For all candidate regions The mean and standard deviation, To adjust the parameters.

[0029] S23 includes: S231, the boundary difference map, the candidate region map of structure-guided perturbation, and the current temporal remote sensing image (temporal image after tower base construction) are stitched together in the channel dimension into a unified feature tensor, and the fused features are extracted through a convolutional encoder, represented as: ; ; in, This is a channel-level splicing operation. For a shared convolutional encoder used for feature extraction, This is the unified feature tensor after splicing. Features of fusion; The channel dimension concatenation operation is represented as: ; in, This represents the number of channels in the current time-phase remote sensing image; The shared convolutional encoder used for feature extraction is represented as: ; ; ; in, , These are the feature maps output from the first and second convolutional layers, respectively. To modify the activation function of the linear unit, This is a Batch Normalization operation. This is a two-dimensional convolution operation with a kernel size of 3×3; S232, the decoder is used to upsample and predict the perturbation distribution of the fused features, and the output perturbation probability map is represented as follows: ; in, This is a probability diagram of the perturbation. For the corresponding decoding network, It is the Sigmoid activation function. This is a convolution operation with a compressed channel of 1.

[0030] S3 includes: S31, using the perturbation probability map and local structural strength, eliminates the spurious perturbation areas caused by the tower base edge or high-reflectivity background, as shown below: ; in, To provide a mask for the real perturbation region retained after removing false perturbations. For the pixels in the output perturbation probability map The probability value of the disturbance. As a probability threshold, Boundary difference map medium pixel The local grayscale standard deviation, This represents the suppression threshold for the intensity of local structural disturbances. ; ; in, For The set of pixels within a local window centered on the pixel. This represents the number of pixels within the local window. This represents the average grayscale value of the pixels within the window. probability threshold Represented as: ; in, The median of the probability of perturbation across the entire graph. This is an empirical offset value; Suppression threshold of local structural disturbance intensity Represented as: ; in, The mean of the local standard deviations of the entire image. This represents the standard deviation of the local standard deviation of the entire graph. To control the adjustment factor for the rejection intensity; S32, through high-frequency edge detection of the image, compensates for the broken and missing parts of the boundary of the perturbation region, and obtains the mask of the missing region of the perturbation boundary. To improve boundary continuity, specifically including: (1) High-frequency response extraction: ; in, This is a high-frequency edge map. For the Laplace operator response; The Laplace operator response is expressed as: ; in, For high-frequency edge maps at pixels The high-frequency edge response value, For pixels grayscale value at that location , , , These are the pixel values ​​at the four neighboring positions of the pixel: top, bottom, left, and right. (2) Compensation for missing boundary areas: ; in, This is a mask used to compensate for missing regions at the perturbation boundary. The threshold for edge compensation intensity; Edge compensation strength threshold Represented as: ; in, , High-frequency edge maps The mean and standard deviation of the entire graph. These are empirical parameters; S33, a mask that merges the actual perturbation region mask with a mask that compensates for the missing perturbation boundary region. Generate vector perturbation results after spatial consistency optimization. , is represented as: ; ; in, The perturbation mask optimized for spatial consistency. To extract connected components and trace contours from the input perturbation mask; Connectivity extraction and contour tracing are performed on the input perturbation mask. Specifically, it includes: (1) Boundary extraction: ; in, The set of contours for all perturbation candidate regions. This is a contour detection function; (2) Polygon fitting: ; in, To outline Fit to a closed polygon. The area of ​​the outline region.

[0031] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0032] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatically extracting land disturbance of transmission line tower foundations based on a fully convolutional neural network, characterized in that, Includes the following steps: S1. Acquire multi-temporal remote sensing image sequences including the transmission line tower base, and based on the differences in ground features before and after the tower base construction, combined with structural boundary constraint rules, extract the tower base structural regions that meet the conditions of structural boundary continuity and ground feature change threshold, and construct a candidate region map for structure-guided disturbance, specifically including: S11, acquire a multi-temporal remote sensing image sequence covering the same transmission line tower foundation area, including temporal images before tower foundation construction. Phase images after tower foundation construction And the time-lapse images before the construction of the tower foundation. Phase images after tower foundation construction Preprocessing is performed, including geometric correction and radiometric normalization, to unify resolution and registration reference, resulting in registered image pairs. ; S12, after obtaining the registered image pair Then, by calculating the difference in ground feature changes of corresponding pixels in the temporal images before and after construction, the change range of each pixel is quantified, and a ground feature change difference map is generated. S13. Based on the geometric morphological features of the tower base structure in the image and combined with the historical tower location coordinate point set, local windows are extracted from the preprocessed post-construction phase image of the tower base. Edge detection is performed on each window and morphological constraints are applied to generate a candidate set of tower base structure boundaries. S14. On the land feature change difference map, apply a land feature change difference threshold constraint to the candidate set of tower base structure boundaries, and extract the tower base structure regions that satisfy both land feature change difference and boundary continuity. ; S15 will be the tower base structure area that satisfies the differences in ground features and the continuity of boundaries. Converted into a candidate region map of structure-guided perturbation ,in, For a set of nodes, It is an edge set; S2 employs a cross-temporal boundary enhancement mechanism and a dual-stream feature extraction network to extract boundary response features of perturbation regions in the current and reference temporal phases, respectively. These features are then structure-aligned and fused to construct a time-sensitive perturbation boundary response map. Furthermore, a joint modeling process is performed using a structure-guided perturbation candidate region map and the current temporal remote sensing image, outputting a perturbation probability map, specifically including: S21, the registered tower base construction phase image Phase images before tower foundation construction Each input is fed into a boundary feature extraction network with shared structure to extract the corresponding boundary response feature maps. , ; S22, through the , A boundary difference map is constructed by performing pixel-by-pixel interpolation to identify the spatial variation boundary of the tower base disturbance area; S23 takes the boundary difference map, the structure-guided perturbation candidate region map and the current time-phase remote sensing image as input, performs joint feature fusion and prediction, and outputs a perturbation probability map; S3, based on the perturbation probability map, uses a pseudo-perturbation region suppression mechanism to remove misjudged regions caused by the tower base structure edge and high-reflectivity ground objects, and repairs the missing perturbation boundary through a high-frequency edge compensation strategy, outputting the perturbation region vector result after spatial consistency optimization.

2. The method for automatic extraction of land disturbance of transmission line tower foundation based on a fully convolutional neural network according to claim 1, characterized in that, S11 includes: S111, Obtain temporal images of the tower foundation before construction, covering the same tower foundation area. Phase images after tower foundation construction ; S112, regarding the acquired temporal images prior to tower foundation construction. Phase images after tower foundation construction Perform geometric corrections to place them in the same map projection coordinate system; S113, using the geometrically corrected post-construction phase image of the tower base as a reference, perform radiometric normalization processing on the geometrically corrected pre-construction phase image of the tower base. S114, using the radiometrically normalized post-construction temporal image of the tower foundation as a reference, performs sub-pixel-level registration on the radiometrically normalized pre-construction temporal image of the tower foundation, and resamples it to a uniform spatial resolution, outputting the registered image pair. .

3. The method for automatic extraction of land disturbance of transmission line tower foundation based on a fully convolutional neural network according to claim 2, characterized in that, S13 includes: S131, the preprocessed time-phase image after the tower foundation construction is recorded as Let the set of coordinate points of the historical tower sites be... ,in, For the first Two-dimensional coordinate vectors of the base points of the tower. The x-axis is... The vertical axis is , The total number of historical tower location coordinates, in Center-trimmed window that adapts to resolution ; S132, in each window Within the process, the image grayscale values ​​are normalized and then processed according to a preset high reflectance threshold. Extract image grayscale values ​​that meet the normalization criteria and are greater than or equal to The region, thereby generating a high-reflectivity mask. ; S133, Apply Gaussian smoothing to the preprocessed time-phase image of the tower base construction within the window. The gradient is then calculated, and the edge set is obtained through nonmaximum suppression and double-threshold hysteresis connection. And calculate the edge connectivity. ; S134, in a high-reflectivity mask With edge set Extract candidate contour sets from the intersection of Calculate the area for each candidate contour. Minimum circumscribed rectangle size ,perimeter And define morphological constraints, including aspect ratio. Rectangularity Edge filling rate Candidate contours that simultaneously satisfy the candidate conditions are retained as the candidate set of the tower base structure boundary. The candidate conditions are expressed as follows: ； in, For the first The first window One candidate contour, , , , , These are the corresponding morphology and continuity discrimination thresholds, This represents the candidate set of tower base structure boundaries constrained by morphological structure.

4. The method for automatic extraction of land disturbance of transmission line tower foundation based on a fully convolutional neural network according to claim 3, characterized in that, S14 includes: S141, in each candidate set of tower base structure boundaries Within, the mean change range of its value on the land cover change difference map is statistically analyzed. ; S142, Set the threshold for ground feature change intensity Preserve the mean change range Greater than or equal to the ground feature change intensity threshold The region, as a set of regions constrained by morphology and intensity of change. ; S143, sets the regions that have passed the morphological and change intensity constraints. Apply boundary connectivity constraints again, extract the effective tower base structure region, and finally output the tower base structure region. , is represented as: 。 5. The method for automatic extraction of land disturbance of transmission line tower foundation based on a fully convolutional neural network according to claim 4, characterized in that, S21 includes: S211, Time-phase images of the registered tower base after construction Phase images before tower foundation construction Normalization processing is performed to obtain normalized time-phase images after tower base construction and before tower base construction; S212, the normalized post-construction phase image and pre-construction phase image of the tower base are input into the structure-shared boundary feature extraction network to obtain the boundary response feature map.

6. The method for automatic extraction of land disturbance of transmission line tower foundation based on a fully convolutional neural network according to claim 5, characterized in that, S22 includes: S221, Calculate the boundary response difference between the boundary response feature map of the time-phase image after tower foundation construction and the boundary response feature map of the time-phase image before tower foundation construction at each pixel location, and obtain the boundary difference map. ; S222, Based on whether the pixel values ​​in the boundary difference map exceed the boundary disturbance intensity threshold. Generate a preliminary perturbation region mask. , is represented as: ； in, This represents the candidate boundary disturbance region. S223, Masking of the initial disturbance region All connected regions are filtered by size and shape, retaining those with an area greater than [missing information]. And the boundary continuity is greater than The region is used as the final disturbance boundary region. , is represented as: ； in, This is the set of all connected regions in the boundary perturbation mask. Connected region The number of pixels, For the region Edge connectivity, The minimum disturbance area threshold. This is the minimum edge connectivity threshold.

7. The method for automatic extraction of land disturbance of transmission line tower foundation based on a fully convolutional neural network according to claim 6, characterized in that, S23 includes: S231, the boundary difference map, the candidate region map of structure-guided perturbation and the current time-phase remote sensing image are stitched together into a unified feature tensor in the channel dimension, and the fused features are extracted by a convolutional encoder; S232 uses a decoder to upsample the fused features and predict the perturbation distribution, outputting a perturbation probability map.

8. The method for automatic extraction of land disturbance of transmission line tower foundation based on a fully convolutional neural network according to claim 7, characterized in that, S3 includes: S31, using the perturbation probability map and local structural strength, eliminates the spurious perturbation areas caused by the tower base edge or high-reflectivity background, as shown below: ； in, To provide a mask for the real perturbation region retained after removing false perturbations. For the pixels in the output perturbation probability map The probability value of the disturbance. As a probability threshold, Boundary difference map medium pixel The local grayscale standard deviation, This represents the suppression threshold for the intensity of local structural disturbances. S32, through high-frequency edge detection of the image, compensates for the broken and missing parts of the boundary of the perturbation region, and obtains the mask of the missing region of the perturbation boundary. ; S33, a mask that merges the actual perturbation region mask with a mask that compensates for the missing perturbation boundary region. Generate vector perturbation results after spatial consistency optimization. .

Images