Intelligent identification method for river and lake management boundary based on aerial survey image

Abstract

The application provides a kind of river and lake management boundary intelligent identification method based on aerial image, comprising: obtaining the aerial image of target river and lake area, camera attitude data, ground positioning data and ground elevation data, forming candidate boundary set after correction, registration and alongshore ground object identification;Virtual water rise, virtual water recession and local water accumulation image are constructed, and cross-image boundary response sequence is generated;According to the position, length, category and two-side adjacent ground object change of candidate boundary, water level response track is formed, and identity stable candidate boundary is screened;Isolation reconstruction and open detection reconstruction are respectively carried out on the two sides of candidate boundary, and the leakage degree of cross-border reconstruction is calculated;Identity stability degree and cross-border reconstruction leakage degree are combined to screen effective boundary section, and river and lake management boundary and sub-section confidence are output after boundary connection, topological verification, constrained smoothing and ground coordinate correction.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an intelligent identification method for river and lake management boundaries based on aerial survey images. Background Technology

[0002] River and lake management boundaries are crucial foundational data for river and lake spatial control, shoreline utilization supervision, and the review of river-related construction projects. Existing technologies typically utilize drones to acquire aerial survey images of river and lake areas, and then use image recognition methods to segment water bodies, bank slopes, embankments, roads, vegetation, and riverside structures. Edges are then extracted based on differences in feature outlines, colors, or textures to aid in determining river and lake management boundaries.

[0003] However, the boundaries of river and lake management do not necessarily coincide with the current water body edges in aerial survey images. Affected by water level fluctuations, localized water accumulation, and bank submersion, the water body edges change with the time of shooting, while boundaries of management significance such as levee tops, the outer edges of roads along levees, and flood control walls usually remain stable. Existing methods mostly identify boundaries directly based on the water-land boundary or prominent edges in a single temporal image, making it difficult to distinguish between temporary edges that change with the state of water cover and long-term stable boundaries.

[0004] Meanwhile, while internal road lines, abrupt slope texture lines, vegetation light-dark boundaries, and shadow edges are positionally stable, they do not represent administrative boundaries. Existing methods lack verification of whether the image structures on both sides of the candidate boundary are truly isolated from each other, easily misidentifying stable pseudo-edges within the same continuous feature as administrative boundaries. Furthermore, tree canopies, bridges, culverts, and structures can cause local breakages in the boundary. Existing connection methods, if relying solely on endpoint distance or direction, are prone to issues such as crossing main water bodies, connecting left and right banks, or repeatedly segmenting the same feature.

[0005] Therefore, this invention proposes an intelligent identification method for river and lake management boundaries based on aerial survey images. The information disclosed in the background section is only for enhancing understanding of the background of this disclosure and may therefore contain prior art information that is not common knowledge to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent identification method for river and lake management boundaries based on aerial survey images, thereby solving the technical problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The intelligent identification method for river and lake management boundaries based on aerial survey images includes the following steps: S1. Acquire aerial survey images, camera attitude data, ground positioning data, and ground elevation data of the target river and lake area. After correction, registration, and image recognition, obtain standard aerial survey images, coastal feature identification maps, and candidate boundary sets. S2. Based on ground elevation data, coastal feature identification map and candidate boundary set, determine the allowable and prohibited change areas of virtual flooding, virtual receding and local water accumulation. Under the condition that the main outline of stable features remains unchanged, generate a group of generated images of water body coverage status, and match the candidate boundaries in the generated image group to form a cross-image boundary response sequence. S3. Calculate the position, length, category and changes of adjacent objects on both sides of the candidate boundary based on the cross-image boundary response sequence to form a water level response trajectory. Eliminate water level response type candidate boundaries and perform segmented identity stability evaluation on the remaining candidate boundaries to obtain a set of identity stable candidate boundaries. S4. Establish a two-sided boundary analysis zone for the candidate boundary segments in the identity stable candidate boundary set, and perform isolation reconstruction and open detection reconstruction respectively. Calculate the degree of cross-boundary reconstruction leakage based on the reconstruction difference before and after opening the image features on the other side. S5. Select effective boundary segments based on the stability of identity and the degree of leakage in cross-boundary reconstruction. Establish connection relationships based on endpoint distance, extension direction, boundary category and adjacent objects on both sides. After topology verification, constrained smoothing and ground coordinate correction, output the confidence of river and lake management boundaries and sub-segments.

[0008] S1 specifically includes: acquiring original aerial survey images, camera attitude data, ground positioning data, camera intrinsic parameters, and ground elevation data; after correction, registration, pixel scale unification, and brightness equalization, a standard aerial survey image is obtained; an image recognition model is used to identify water bodies, banks, embankments, roads, vegetation, exposed surfaces, and coastal structures, extracting the region contours, color distribution, main texture direction, and spatial adjacency relationships to form a coastal feature identification map; based on the coastal feature identification map, the edges of water bodies, banks, embankment tops, roads, vegetation transitions, and structures are extracted, and connected according to endpoint distance, extension direction, boundary category, and adjacent features on both sides to form a candidate boundary set.

[0009] S2 specifically includes: determining the change distances, allowed change areas, prohibited change areas, and stable feature outlines for virtual flooding, virtual receding, and localized water accumulation based on ground elevation data, coastal feature identification maps, and candidate boundary sets, forming a counterfactual water body change constraint set; generating extended water body, synthetic receding, and low-lying water accumulation images based on the counterfactual water body change constraint set, and obtaining a generated image group that passes the stable feature outline displacement and category retention rate verification through edge transition and strategy optimization; performing image recognition and edge extraction on the generated image group, matching candidate boundaries based on distance, direction, boundary category, and adjacent features on both sides, recording the matching status according to water body change type and level, and forming a cross-image boundary response sequence.

[0010] S3 specifically includes: calculating the lateral position offset, continuous length change, boundary category change, and changes of adjacent objects on both sides of the candidate boundary based on the cross-image boundary response sequence to obtain water level response features; arranging the water level response features in the order of virtual receding water, original water body state, virtual rising water, and local water accumulation to form a water level response trajectory; distinguishing between water level response type candidate boundaries and initial stable candidate boundaries based on the synchronous movement of the candidate boundary and the water body edge, the coverage by the water body, and the category replacement; dividing the initial stable candidate boundaries into segments, calculating the degree of identity stability based on the position, length, category, retention of adjacent objects on both sides, and effective detection, and selecting a set of identity stable candidate boundaries.

[0011] S4 specifically includes: establishing a bilateral boundary analysis band for candidate boundary segments in the identity-stable candidate boundary set, setting masking windows located on both sides of the boundary but not crossing the boundary to form a bilateral reconstruction input image; using a bi-branch mask image reconstruction model, performing isolation reconstruction by cutting off the opposite side image features and open detection reconstruction by opening the opposite side image features for each masking window, calculating the cross-boundary dependency gain based on the error difference between the two reconstruction results, and generating a bilateral cross-boundary response map; according to the candidate boundary number and segment number, statistically analyzing the cross-boundary response pixel ratio, response direction, maximum continuous response length, and average response intensity, calculating the cross-boundary reconstruction leakage degree and reconstruction evidence integrity, and forming a segmented reconstruction leakage result.

[0012] S5 specifically includes: associating the stable candidate boundary set with the segment reconstruction leakage results; evaluating the segment quality based on the stability of identity, the degree of cross-boundary reconstruction leakage, and the completeness of two types of evidence, and selecting effective boundary segments; establishing connection relationships based on the endpoint distance, extension direction, boundary category, and adjacent objects on both sides of the effective boundary segments, and verifying the connection curves by crossing the main water body, swapping left and right banks, and repeated segmentation, to form the initial river and lake management boundary; performing constrained smoothing, ground coordinate correction, and topological verification on the initial river and lake management boundary; calculating the segment confidence based on the stability of identity, the degree of cross-boundary reconstruction leakage, the completeness of evidence, and the degree of connection continuity, and outputting the river and lake management boundary identification results.

[0013] The beneficial effects of this invention are as follows: This invention improves the accuracy of candidate boundary extraction for complex river and lake shorelines by performing geometric correction, spatial registration, and shoreline feature image recognition on aerial survey images to unify the ground coordinates, categories, and adjacent feature information of candidate boundaries. By constructing virtual flood, recession, and localized water accumulation images, the invention analyzes the changes in the position, length, category, and adjacent features of candidate boundaries under different water cover conditions, effectively eliminating temporary water body edges that change with water level.

[0014] This invention performs segmented identity stability evaluation on candidate boundaries, preserving boundary segments with stable position and visual identity under various water cover conditions, thus reducing the impact of single-temporal images on the river and lake management boundary identification results. By performing isolation reconstruction and open detection reconstruction on both sides of the candidate boundary respectively, the degree of cross-boundary reconstruction leakage is quantified, enabling the identification and elimination of positionally stable false boundaries such as internal road lines, slope texture lines, and internal vegetation lines.

[0015] This invention establishes boundary connections by combining endpoint distance, extension direction, boundary category, adjacent features on both sides, and connectivity with the main water body. This reduces topological errors such as crossing the main water body, connecting the left and right banks, and repeatedly segmenting the same feature. Through constrained smoothing, ground coordinate correction, and segmented confidence calculation, it improves the spatial continuity, coordinate accuracy, and traceability of identification results of river and lake management boundaries while maintaining the true turning points of flood walls, gates, etc. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent identification method for river and lake management boundaries based on aerial survey images according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example: Figure 1 As shown, this embodiment provides a method for intelligent identification of river and lake management boundaries based on aerial survey images, including the following steps: S1. Acquire aerial survey images, camera attitude data, ground positioning data, and ground elevation data of the target river and lake area. After correction, registration, and image recognition, obtain standard aerial survey images, coastal feature identification maps, and candidate boundary sets. S2. Based on the ground elevation data, the coastal feature identification map and the candidate boundary set, determine the allowable and prohibited change areas for virtual flooding, virtual receding and local water accumulation. Under the condition that the main outline of stable features remains unchanged, generate a group of generated images of water body coverage status, and match the candidate boundaries in the generated image group to form a cross-image boundary response sequence. S3. Calculate the position, length, category and changes of adjacent objects on both sides of the candidate boundary according to the cross-image boundary response sequence to form a water level response trajectory. Eliminate water level response type candidate boundaries and perform segmented identity stability evaluation on the remaining candidate boundaries to obtain a set of identity stable candidate boundaries. S4. Establish a dual-sided boundary analysis zone for the candidate boundary segments in the identity stable candidate boundary set, and perform isolation reconstruction and open detection reconstruction respectively. Calculate the degree of cross-boundary reconstruction leakage based on the reconstruction difference before and after opening the image features on the other side. S5. Select effective boundary segments based on the stability of identity and the degree of leakage in cross-boundary reconstruction. Establish connection relationships based on endpoint distance, extension direction, boundary category and adjacent objects on both sides. After topology verification, constrained smoothing and ground coordinate correction, output the confidence of river and lake management boundaries and sub-segments.

[0019] S1 specifically includes the following sub-steps: S110. Acquire the original aerial survey images, camera attitude data, ground positioning data, camera intrinsic parameters and ground elevation data of the target river and lake area, and form standard aerial survey images through geometric correction, spatial registration and radiometric unification.

[0020] The original aerial survey images were acquired by UAVs equipped with visible light cameras, with a forward overlap of 70%-85% and a lateral overlap of 60%-80%. Camera attitude data, including roll, pitch, and heading angles, were output by the flight control or inertial navigation system at the time of exposure. Ground positioning data was output by the airborne global satellite navigation system. Camera parameters were obtained from pre-aerial calibration, including focal length, principal point coordinates, and distortion coefficients. Ground elevation data came from existing digital elevation models, or from digital terrain models obtained by filtering out tree canopies, buildings, and bridges after generating digital surface models from the overlapping images through aerial triangulation.

[0021] Distortion is corrected based on camera intrinsic parameters, and then orthorectified using attitude, positioning, and ground elevation data, transforming the image to the same projection coordinate system. The median ground sampling distance of each image is used as the uniform pixel scale, in m / px, and bilinear interpolation is employed for resampling.

[0022] Matching point pairs are established using the intersections of road or embankment turning points, corner points of riverside structures, and edges of stable vegetation. A random sampling consensus algorithm is used to eliminate mismatched points. The registration error is: in, This represents the root mean square error of registration, in pixels (px). Indicates the number of valid matching pairs; , Indicates the coordinates of the nth matching point; Indicates coordinate transformation relationships; This represents the Euclidean distance.

[0023] like If the difference is less than 1.5px, the registration result is retained; otherwise, the matching points are re-extracted. Histogram matching and brightness equalization are performed on the registered images, which are then stitched together to form a standard aerial survey image, and the correspondence between pixels and projected ground coordinates is saved.

[0024] For example, when water bodies occupy 70% of the overlapping area, 24 sets of stable feature matching points are selected. If the value is 0.82px, then the registration is successful.

[0025] S120. Perform image recognition on standard aerial survey images to create a coastal feature identification map. The image recognition model is trained using historical aerial survey images and labeled samples of the local area, and is labeled with eight categories: water area, bank slope area, dam area, road area, vegetation area, bare surface area, coastal structure area, and background area. The bare surface area is an area showing soil, gravel, or bare rock texture and not covered by other features; the coastal structure area includes bridges, gates, pumping stations, revetments, and adjacent buildings.

[0026] Standard aerial survey images are cropped into 1024px × 1024px image blocks, with adjacent blocks overlapping by 20%. The model outputs pixel class probabilities. The probabilities of the same class at overlapping locations are averaged, and the class corresponding to the highest average probability is used as the initial class. Pixels with a highest average probability lower than 0.60 are supplemented by classifying the most numerous class in their 8-neighborhood with a probability not lower than 0.60.

[0027] Connected component labeling is performed on adjacent pixels of the same type, recording the region number, region outline, mean color, standard deviation of color, main texture direction, and spatial adjacency relationship; where the main texture direction is determined by the direction of the largest eigenvector of the structure tensor formed by the gray-level gradients within the region. Two region outlines are considered spatially adjacent when the minimum distance is no greater than the ground distance corresponding to 1px. The region category confidence is the average probability of the pixel category belonging to that category within the region.

[0028] The water body area that is connected to the preset center point of the target river or lake, or has the largest connected area, is determined as the main water body area, and the remaining unconnected water body areas are determined as auxiliary water body areas. The coastal feature identification map records pixel coordinates, pixel category probability, region category, region category confidence, region number, region outline, color distribution, main texture direction, and spatial adjacency relationship.

[0029] For example, a grassy slope adjacent to water is initially identified as a vegetated area, but when it is adjacent to the main water body, the ground elevation rises in the direction away from the water body, and the outer side is connected to a road, it is corrected to a bank slope area.

[0030] S130. Based on the coastal feature identification map, extract and connect the edges of coastal features to form a candidate boundary set. The common contour of the water area and the non-water area is determined as the water edge, the outer contour of the bank slope area is determined as the bank slope edge, the side contour of the road area along the long axis is determined as the road edge, the common contour of vegetation and non-vegetation with a continuous length of not less than 5m is determined as the vegetation transition edge, and the outer contour of the coastal structure facing the main water area is determined as the structure edge.

[0031] For the embankment area, normal profiles are set every 0.5m along the main direction of the embankment. Gray-level and texture abrupt change points with a positional variation of no more than 2px in adjacent profiles are connected to form the embankment crest edge. A sequence of ground coordinate points is saved for each edge, and sampling bands with a width of 10px are set on both sides. The area category with the largest area proportion is used as the adjacent ground object; the side closer to the main water body area is defined as side 1, and the side farther from the main water body area is defined as side 2. A boundary category compatibility table is pre-established, in which the embankment crest edge and the outer edge of the road along the embankment belong to a compatible category, while the water body edge and the road edge do not belong to a compatible category.

[0032] Calculate the endpoint distance and orientation difference at the edges to be connected: in, Indicates edge End point and edge Ground distance between starting points; and They represent the corresponding endpoints; This represents the difference in the fitting direction within a 1m range of the two edge ends; It represents pi (π).

[0033] Connections are made when the endpoint distance is no greater than 0.6m, the direction difference is no greater than 15°, the boundary categories are the same or compatible, and the adjacent objects on both sides are arranged in the same way. A unique number is assigned to each candidate boundary, and the boundary point sequence, boundary category, boundary category confidence level, continuous length, extension direction, adjacent objects on both sides, color distribution, and main texture direction are recorded. The boundary category confidence level is the average pixel category probability related to the corresponding boundary category within the sampling band on both sides of the boundary.

[0034] For example, two sections of embankment top edges with a distance of 0.45m, a directional difference of 8°, and on opposite sides a bank slope area and a road area, are connected; internal lines on both sides that are road areas are not connected.

[0035] S2 specifically includes the following sub-steps: S210. Based on standard aerial survey images, ground elevation data, coastal feature identification maps, and candidate boundary sets, construct a counterfactual water body change constraint set. The expansion of the water body region towards the second side represents virtual flooding, the contraction of the current water body edge towards the first side represents virtual receding, and water covering in localized low-lying areas on the second side represents localized water accumulation.

[0036] For each candidate boundary, a pixel-by-pixel search is performed along the normal direction pointing from the current water body edge to the second side. The distance before the first contact area with the dam area, road area, or riverside structure area with a confidence level of not less than 0.70 is determined as the usable distance for rising water. Along the first side, 40% of the local water body width is taken as the initial receding distance. If the water body edges on both sides intersect after contraction, the receding distance is gradually reduced until the water body edges on both sides no longer intersect, and the obtained distance is determined as the usable distance for receding water.

[0037] The width of a local water body is the distance between the intersection of a straight line passing through the midpoint of the reference candidate boundary and perpendicular to the main extension direction of the main water body, and the edges of the water bodies on both sides; for meandering river sections, the median width of three or more cross-sections is used. The available distances for both flooding and receding water are divided into 3-6 variation levels.

[0038] For pixels located on the second side, belonging to the bank slope area or exposed surface area, calculate the relative depression height: in, Indicates the relative depression height of pixel x; This indicates the ground elevation corresponding to that pixel; This represents the average elevation within a 5m radius neighborhood centered at pixel x. This refers to the ground neighborhood.

[0039] when A pixel is defined as a low-lying pixel if its elevation is not less than 0.15m and not less than twice the nominal error of the ground elevation data, and if there are no stable features obstructing the main water body area. Connecting adjacent low-lying pixels forms a low-lying area on the bank slope. The main outline of stable features is the outer outline of dam areas, road areas, and riverside structures areas with a region category confidence level of not less than 0.70 and an area of ​​not less than 4m².

[0040] Record candidate boundary numbers, water body change types and levels, coverage change distances, permitted change areas, prohibited change areas, main outlines of stable features, and bank slope depression area numbers to form a counterfactual water body change constraint set.

[0041] For example, when the edge of the water body is 5.2m away from the road area, the usable distance for flooding is 5m; when an 18m² depression is 0.28m lower than the surrounding area and there are no obstructions, it is identified as a low-lying area on the bank slope.

[0042] S220. Generate virtual flood images, virtual receding water images, and localized water accumulation images based on the counterfactual water body change constraint set. When generating the virtual flood image, establish a water body appearance sample library from image blocks within 10m of the bank section to be expanded, belonging to the water body area, and containing no saturated reflective pixels. Record the color mean, grayscale gradient distribution, and texture scale. Saturated reflective pixels are pixels where any color channel reaches more than 98% of the maximum grayscale value of that channel, and the standard deviation of grayscale in a 5px neighborhood is lower than the reflectivity threshold. Select the water body image block closest to the expansion location, adjust its color mean to the color mean of the neighboring water bodies, and then fill the allowable variation area.

[0043] When generating a virtual receding water image, image patches with the smallest elevation difference and a main texture direction difference of no more than 20° are selected from the adjacent second-side bank slope area or exposed surface area to fill the area covered by receding water. If no suitable image patch exists, a placeholder surface texture is generated based on the average color and main texture direction of adjacent exposed surfaces, and marked using a synthesized receding water pixel mask. The placeholder surface texture is only used to construct the water shrinkage scene and is not used as evidence of real land cover categories.

[0044] When generating images of localized water accumulation, the shortest connected path between the low-lying areas of the bank slope and the main water body is calculated without passing through prohibited change areas. If such a path exists, the water body extends along the path; otherwise, a closed water accumulation area is formed within the low-lying area. A transition zone of 1-5 pixels is set at the edge of the generated water body, and the width of the transition zone does not exceed 20% of the minimum distance between adjacent candidate boundaries. The strategy optimizes the traversal of water body change levels, transition zone width, and water accumulation coverage ratio.

[0045] For each set of parameters, a temporary image is generated. The same image recognition and edge extraction rules from S120 and S130 are applied to calculate the position, category, and differences between candidate boundaries and adjacent objects. The displacement of the stable object's main outline and the anomaly gradient of the generated region are also calculated. Parameter combinations with larger response differences and smaller outline displacement and anomaly gradients are selected. Temporary identification is only used for parameter selection; after parameter selection, a formally generated image set is formed. If the average displacement of the stable object's main outline exceeds 1px, the stable object category retention rate is less than 95%, or the changed pixels exceed the allowable change area, suboptimal parameters are used for regeneration.

[0046] For example, if the road outline shifts by 2.4px due to a 2m virtual rise in water level, the width of the transition zone or the distance of the change in water coverage should be reduced.

[0047] S230: Image recognition is performed on both the standard aerial survey images and the officially generated image set using the same model and parameters from S120. Edge extraction and connection rules from S130 are then used to generate the boundaries to be matched. For each candidate boundary, a boundary search zone is established centered on its ground coordinate point sequence. The width of the search zone is the sum of the corresponding water cover change distance and 1 meter.

[0048] Using the original candidate boundary as the reference boundary, the boundaries to be matched within the search band are resampled at intervals of 0.2m. The distance, direction, category, and consistency with adjacent objects are calculated to obtain the matching score. in, This represents the matching score between reference boundary a and the boundary b to be matched; This represents the average bidirectional nearest point distance; Indicates the search band width; Indicates the difference in direction; This indicates that the adjacent objects on both sides are consistent; This indicates consistency in boundary categories.

[0049] When the adjacent objects on both sides are identical, Take 1; take 0.5 if only one side is consistent; take 0 if neither side is consistent; when the categories are the same, The value is set to 1, 0.5 when the boundary category compatibility relationship is satisfied, and 0 otherwise. When the matching score is not lower than 0.75, it is determined as the corresponding boundary; when there is no qualified boundary, it is recorded as not detected; when it corresponds to multiple consecutive boundaries, it is recorded as split detection and merged point sequence.

[0050] If the score difference between two boundaries to be matched is less than 0.05, the best match is selected based on the color distribution on both sides and the main texture direction recorded in S130. For the area covered by the synthetic receding pixel mask, only the state of water body or non-water body is determined, and its specific land cover category is not used as evidence of subsequent changes.

[0051] Following the sequence of virtual receding water, original state, virtual rising water, and local water accumulation, the candidate boundary number, image number, change type and level, matching status, boundary point sequence, bidirectional average nearest point distance, continuous length, boundary category, adjacent objects on both sides, and matching score are recorded to form a cross-image boundary response sequence.

[0052] S3 specifically includes the following sub-steps: S310: Read the candidate boundary set formed by S130 and the cross-image boundary response sequence formed by S230. Use the candidate boundaries in the standard aerial survey image as reference candidate boundaries and the boundaries that match them in each officially generated image as response boundaries to calculate the water level response characteristics.

[0053] Water level response features are used to describe the changes in the position, length, detection status, adjacent objects, and boundary category of the same candidate boundary under different water body coverage conditions. The reference candidate boundary and the response boundary are resampled at 0.2m intervals. The nearest matching point is determined for the k-th sampling point on the reference candidate boundary, and the signed lateral position offset is calculated. in, This represents the signed lateral position offset of the k-th sampling point; Indicates the reference sampling point; Indicates the nearest matching point; Represents the unit normal vector pointing from the first side to the second side; symbol This represents the vector dot product. A positive value indicates a shift to the second side, and a negative value indicates a shift to the first side. This is based on all valid sampled points. The median serves as the lateral offset of the response boundary.

[0054] The rate of change of continuous length is calculated according to the following formula: in, Indicates the rate of change of continuous length; Indicates the continuous length of the response boundary; Indicates the continuous length of the reference candidate boundary.

[0055] When splitting is detected, The sum of the common coverage lengths of each split boundary and the reference candidate boundary is taken; if no boundary is detected, the sum is 0. When the main category of the adjacent object on the 1st or 2nd side changes relative to the reference candidate boundary, the change of the adjacent object is recorded respectively; when the response boundary category changes to another category that does not satisfy the boundary category compatibility relationship, the boundary category change is recorded.

[0056] Specific category changes in the area covered by the synthetic receding pixel mask are not included in the above changes. Normal detection or split detection with a matching score of not less than 0.75 is considered a valid detection; other cases are considered invalid matches. The water level response features consist of candidate boundary number, image number, change type and level, lateral position offset, continuous length change rate, adjacent objects and category change markers, matching score, and detection status.

[0057] For example, if the water body edge B05 moves 1.86m to the second side in a 2m flood image and the second side becomes a water body area, then record the positive offset and the change of adjacent objects; if the levee top edge B17 only shifts by 0.08m and the two sides remain unchanged, then record it as a low response.

[0058] S320. Arrange the water level response characteristics of the same reference candidate boundary in ascending order of water body coverage to form a water level response trajectory. The water level response trajectory includes each virtual receding level, the original water body state, and each virtual rising level, with local water accumulation scenarios as an additional branch. For each level of change, search for the nearest water body edge along the normal of the reference candidate boundary within the connected area of ​​the same main water body, using the lateral positional offset of that water body edge as a comparison benchmark.

[0059] When a candidate boundary moves in the same direction as the edge of a neighboring water body, the difference in offset is no greater than 0.5m and no greater than 20% of the larger offset, and this holds true in at least two consecutive change levels, the candidate boundary is determined to have moved synchronously with the edge of the water body. When the original location of the candidate boundary is identified as a water body area, and it is not detected or the continuous length retention rate is less than 30%, the candidate boundary is determined to have been covered by water. When a non-water body edge becomes a water body edge in two consecutive change levels, or when the second adjacent object becomes a water body area, it is determined to be a water-driven category replacement.

[0060] If any of the conditions are met, the corresponding candidate boundary is marked as a water level response candidate boundary. If the above conditions are not met, and the median absolute value of the lateral position offset in all valid generated images is not greater than the larger of the ground distances of 0.3m and 3px, the median continuous length retention rate is not less than 80%, the subject boundary category is maintained, and the number of images with changes in adjacent objects on both sides does not exceed 20% of the number of valid generated images, the candidate boundary is marked as an initial stable candidate boundary.

[0061] Valid generated images are officially generated images that pass the S220 validity check. If more than 50% of the officially generated images fail to obtain a valid response, and the candidate boundary is not covered by water, it is marked as a candidate boundary with insufficient response evidence.

[0062] For example, if the edge of a water body moves by 0.94m, 1.89m, and 2.82m in the 1m, 2m, and 3m flood images respectively, it should be marked as a water level response candidate boundary; the internal road line that does not move at all is temporarily listed as an initial stable candidate boundary, and will be excluded later through cross-boundary reconstruction leakage test.

[0063] S330. Perform segmented stability evaluation on the initial stable candidate boundaries to form a set of identity stable candidate boundaries. Divide the initial stable candidate boundaries into segments with a ground length of 5m. If the last segment is less than 5m, merge it into the previous segment. Calculate the location stability score, length retention score, category retention score, adjacent object retention score, and effective detection score, and calculate the degree of identity stability. in, Indicates the degree of stability of identity; The score represents the positional stability score, calculated by subtracting the median of the absolute value of the lateral positional offset from 1 and the upper limit of the allowed offset; if the result is less than 0, it is set to 0. This represents the median of the continuous length retention rate; it is set to 1 if it exceeds 1. This indicates the percentage of valid generated images that retain their category. This indicates the percentage of valid generated images where both adjacent objects on both sides are preserved. This indicates the percentage of images that were effectively detected.

[0064] The stability threshold is determined by manually verified stable edge and water level response edge samples. It is traversed within the range of 0-1 at 0.05 intervals, selecting the threshold with the largest sum of the two types of sample identification indicators; a threshold of 0.75 is used when validation samples are lacking. Segments with an identity stability level not lower than the stability threshold and an effective detection score not lower than 0.50 are retained. Evidence completeness is the ratio of the number of officially generated images that actually obtain effective water level response features to the planned number of generated images.

[0065] When the distance between adjacent reserved sections does not exceed the S130 connection threshold, the same candidate boundary number is used and different section numbers are assigned. The boundary point sequence, boundary category, adjacent objects on both sides, five scores, identity stability, evidence integrity, water level response trajectory and corresponding generated image number of each section are recorded to form an identity-stable candidate boundary set.

[0066] For example, if the stability of the first 30m segment of a 40m candidate boundary is 0.88, and the stability of the last 10m segment is only 0.42 due to the rise in water level and the boundary becoming the edge of the water body, then only the first 30m segment is retained.

[0067] S4 specifically includes the following sub-steps: S410: Read the standard aerial survey image generated by S110, the coastal feature identification map generated by S120, the candidate boundary set generated by S130, and the identity-stable candidate boundary set generated by S330, and establish a boundary analysis zone according to the candidate boundary number and segment number. The boundary analysis zone is a strip-shaped area centered on the identity-stable candidate boundary segment, covering along the length of the segment and extending to both sides.

[0068] For each 5m segment, the strip area is resampled into a rectangle with the cumulative length direction as the longitudinal direction and the normal direction from the first side to the second side as the transverse direction. The base width of each side is taken as the larger value between the ground distances of 3m and 30px, and must not exceed 50% of the distance from that side to the nearest other candidate boundary; if the available width is less than 10px, it is marked as insufficient analysis width.

[0069] A masking window is set every 1m along the longitudinal direction of the rectangular boundary analysis zone. The window is 1m long longitudinally and 30% of the available width on that side laterally, and must not cross the identity-stable candidate boundary. Unmasked images are retained between adjacent windows as visible context. The masking window is the region to be reconstructed to test whether the local content of the image on this side can be recovered from the features of the image on the other side.

[0070] The system simultaneously reads the color channels of standard aerial survey images and the region categories and region category confidence scores from the coastal feature identification map. Pixels with confidence scores below 0.60 are marked as low-confidence pixels, forming the first-side reconstruction input image and the second-side reconstruction input image. If there are fewer than three effective occlusion windows on each side, it is marked as insufficient reconstruction samples. Each pixel retains its rectangular coordinates, original pixel coordinates, projected ground coordinates, region number, and uniform pixel scale for use in writing back the reconstruction results.

[0071] When a stable candidate boundary intersects with another candidate boundary, the boundary analysis band is split at the intersection point to prevent the same occlusion window from spanning two candidate boundaries. The occlusion window's position, size, side, low-confidence pixel ratio, and coordinate correspondence are written into the reconstruction task record.

[0072] For example, when the available distances on both sides of section B17-03 are 4.8m and 2.6m respectively, the analysis widths are taken as 3m and 1.3m respectively.

[0073] S420. For each effective occlusion window, perform isolation reconstruction and open probe reconstruction respectively. A dual-branch masked image reconstruction model with a first-side coding branch, a second-side coding branch, and a shared decoding branch is adopted. The training data comes from image patches that do not overlap with the section to be detected in historical aerial survey images and current standard aerial survey images. During training, 20%-40% of pixels are randomly occluded and their content is restored; coastal land cover categories are used as auxiliary input channels after one-hot encoding. After training, the model parameters are fixed.

[0074] When reconstructing the occluded window on either side, the output of the opposite coding branch is multiplied by 0, retaining only the visible context of this side. For open probe reconstruction, the occluded window, model parameters, and local context remain unchanged; only the output coefficients of the opposite coding branch are adjusted from 0 to 1.

[0075] The reconstruction error is: in, This represents the reconstruction error of the occlusion window q; Represents the set of valid pixels within the window; Indicates the number of valid pixels; and These represent real pixels and reconstructed pixels, respectively. Represents the image gradient; It represents the sum of the absolute values ​​of each component.

[0076] The isolation reconstruction error and the open probe reconstruction error are obtained separately, and the cross-boundary dependency gain is calculated: in, Indicates cross-boundary dependency gain; and These represent the isolation reconstruction error and the open probe reconstruction error, respectively. This is used to prevent the denominator from being 0.

[0077] The pixel-wise absolute difference between the two reconstruction results is calculated, normalized to the standard deviation of the original pixel intensity of the window, and limited to 0-1 to obtain the cross-boundary response intensity. The response is recorded according to the direction of information inflow from the opposite side, and written back according to the projected ground coordinates to form a two-sided cross-boundary response map. If the proportion of low-confidence pixels exceeds 50%, the effective pixels are less than 100px, or any reconstruction is incomplete, the corresponding occluded window is marked as invalid reconstruction.

[0078] For example, when the isolation reconstruction error of the internal road line is 0.18 and the open detection reconstruction error is 0.06, the cross-boundary dependency gain is 0.67; when the two errors of the bank slope and the road boundary are 0.11 and 0.10 respectively, the cross-boundary dependency gain is only 0.09.

[0079] S430 uses the candidate boundary numbers and segment numbers from S330 to summarize the bilateral cross-boundary response maps and generate segmented leakage reconstruction results. The response threshold is determined by manually confirmed pseudo-boundary samples and real boundary samples, and is traversed in the range of 0-1 at intervals of 0.05. The value with the highest accuracy in distinguishing between the two types of samples is selected. When historical samples are missing, virtual dividing lines within the same land feature are used as high leakage samples, and the boundaries between water bodies and banks, and between banks and dike top roads are used as low leakage samples.

[0080] Pixels with cross-boundary response intensities not lower than the response threshold are defined as cross-boundary response pixels. The mean of bidirectional cross-boundary dependency gain, the proportion of cross-boundary response pixels, the maximum continuous response length, and the normalized average response intensity are statistically analyzed, and the degree of cross-boundary reconstruction leakage is calculated. in, This indicates the degree of cross-boundary reconstruction leakage in segment s; This represents the mean cross-boundary dependency gain of the effective occlusion window; Indicates the cross-boundary response pixel ratio; This represents the ratio of the maximum continuous response length to the segment length. This represents the normalized average response intensity, with each value ranging from 0 to 1.

[0081] The reconstruction evidence completeness is the ratio of the number of effective occlusion windows to the planned number of occlusion windows. When this value is below 0.50, the segment is marked as having insufficient reconstruction evidence and is not directly retained based on the low leakage result on the surface. The segment number, bidirectional cross-boundary dependency gain, response pixel ratio, maximum continuous response length, average response intensity, cross-boundary reconstruction leakage degree, reconstruction evidence completeness, and processing status are recorded to form the segment reconstruction leakage result for S510 to call.

[0082] For example, the cross-boundary dependency gain of the 5m road internal section is 0.63, the response pixel ratio is 0.71, and the maximum continuous response length is 4.2m, indicating a high degree of leakage; the corresponding values ​​for the embankment top section are 0.08, 0.12, and 0.6m, indicating that the image structures on both sides are relatively independent.

[0083] S5 specifically includes the following sub-steps: S510: Read the identity stability candidate boundary set formed by S330 and the partition segment reconstruction leakage result formed by S430. Using the candidate boundary number and the segment number as unique association keys, associate the identity stability degree, evidence integrity, cross-boundary reconstruction leakage degree and reconstruction evidence integrity to the same segment.

[0084] The stability condition is met when the stability of the identity and the completeness of its evidence reach their respective thresholds; the isolation condition is met when the degree of leakage in cross-boundary reconstruction does not exceed the leakage threshold and the completeness of the reconstructed evidence reaches the corresponding threshold. Each threshold is determined by manually confirmed samples from historical aerial survey images. Valid samples include the edge of the dike top, the outer edge of the road along the dike, and the outer edge of the flood control wall. Invalid samples include the edge of the water body, the inner line of the road, the slope texture line, and the inner line of the vegetation.

[0085] The search threshold combination was optimized using the following strategy: the identity stability threshold was iterated at intervals of 0.05 within the range of 0.60-0.90; the leakage threshold was iterated at intervals of 0.05 within the range of 0.20-0.50; and the two types of evidence thresholds were iterated at intervals of 0.10 within the range of 0.50-0.90. Weighted values ​​of the effective boundary retention rate, invalid boundary exclusion rate, and segment continuity rate were calculated using weights of 0.40, 0.35, and 0.25, respectively, and the combination with the highest weighted value was selected. The segment continuity rate is the ratio of the continuous retention length to the total length of manually confirmed boundaries.

[0086] When calibration samples are lacking, the identity stability threshold, identity stability evidence threshold, leakage threshold, and reconstructed evidence threshold are set to 0.75, 0.60, 0.35, and 0.50, respectively.

[0087] The section quality value is: in, Indicates the quality value of the section; Indicates the degree of stability of identity; Indicates the completeness of the evidence; Indicates the degree of leakage during cross-boundary reconstruction; This indicates the completeness of the reconstructed evidence.

[0088] Simultaneously meeting the stability condition and the isolation condition and A score of 0.70 or higher is considered a valid boundary segment; other segments are marked as low stability and removed, high leakage and removed, or have insufficient evidence of identity stability or insufficient evidence of reconstruction. The boundary point sequence, boundary category, adjacent objects on both sides, four evaluation values, segment quality value, start and end points, and extension directions at both ends of each valid segment are recorded to form a set of valid boundary segments.

[0089] For example, if the stability of the internal road line is 0.91, but the leakage rate of cross-boundary reconstruction is 0.72, it should be removed.

[0090] S520: Read the set of valid boundary segments, the coastal feature identification map, and the candidate boundary set. Using the valid boundary segments as connection nodes and the start and end points of the segments as connection endpoints, construct a boundary connection map. Only establish candidate connection edges for segments whose endpoint distance is no greater than the connection search distance and whose direction difference is no greater than 30°.

[0091] The connection search distance is taken as the larger of 3 times the S130 connection threshold and 1.5m; when the endpoints are in areas obscured by tree canopies or riverside structures, and the boundary category and adjacent objects on both sides are consistent, the distance can be increased to 3m. A third-order Hermit curve is used to connect the endpoints. The starting and ending tangents are fitted by boundary points within 1m of the end of the segment, and the curve is restricted to a connection corridor centered on the endpoint connection line with a width equal to 50% of the endpoint distance.

[0092] Connection score: in, Indicates a section and section Connection score; Indicates the distance between endpoints; Indicates the search distance; Indicates the difference in direction; It is 30°; This indicates the consistency of the arrangement of adjacent objects. A value of 1 is used when both sides are consistent, 0.5 is used when only one side is consistent, and 0 is used otherwise. This indicates the continuity of the boundary category. It is 1 when the categories are the same, 0.5 when the boundary category compatibility relationship is satisfied, and 0 in other cases.

[0093] Topology verification is performed when the connection score is not lower than 0.75: if the continuous length of the connection curve passing through the main water body area exceeds 0.5m and is not within the bridge or culvert obstruction area, it is determined to cross the main water body; if the connected area of ​​the main water body adjacent to the first side changes after sampling along the normal of the connection curve, it is determined to be a swap of the left and right banks; if more than 80% of the sampling positions on both sides of the connection curve belong to the same connected area, it is determined to be a duplicate segmentation of the same feature. If any of these conditions are met, the candidate connection edge is deleted.

[0094] When multiple candidate connections exist for a single endpoint, the connection with the highest score that does not intersect with the selected curve is selected; multi-branch connections are only allowed at the branch locations of the main water body. The initial river and lake management boundary consists of the effective boundary segment, connection curve, connection score, topology verification results, and unconnected endpoints.

[0095] For example, a connection can be made when the endpoints of the dike crest section are 0.9m apart, the direction difference is 9°, and the adjacent objects are the same; the connection should be deleted when the connecting curve crosses the water body by 1.2m.

[0096] S530. Read the initial river and lake management boundaries, standard aerial survey images, coastal feature identification maps, and boundary category confidence levels. Implement constrained smoothing, ground coordinate correction, and topology verification. Constrained smoothing only applies to the intersection of effective boundary segments and connecting curves, and single-pixel jagged edges. Take 1m boundary points on both sides of the intersection as smoothing windows, and use cubic spline curves for fitting. Fix the window endpoints and boundary points with a boundary category confidence level of not less than 0.80. The normal offset of the smoothing points must not exceed the larger value between 1px corresponding to the ground distance and 0.15m.

[0097] When the outline or boundary category of a coastal structure undergoes a true inflection, the corresponding inflection point is retained. Subsequently, within a 3px range on both sides of the boundary point's normal direction, the edge position with the same boundary category, the same arrangement of adjacent objects on both sides, and the largest image gradient is searched, and the moving distance must not exceed 2px; if no qualified edge is found, the original coordinates are retained; the connecting curves within the occlusion gaps must not snap to road markings, tree canopy edges, or shadow edges.

[0098] After correction, the boundary is re-examined for crossing the main water body, self-intersection, intersection of left and right banks, changes in the main water body's connectivity area, and repeated back-and-forth. If any abnormalities are found, the corresponding smoothing or correction results are revoked. For complete sections without connections, the continuity score is 1; for connected sections, the average score of adjacent connections is used; if there are gaps to be reviewed, the score is no higher than 0.50.

[0099] The confidence level of the partition segment is: in, Indicates the confidence level of the partition segment; Indicates the degree of continuity of the connection; the meanings of the other symbols are consistent with S510.

[0100] Will Segments with a value not lower than 0.80, between 0.60 and 0.80, and below 0.60 are marked as high-confidence segments, medium-confidence segments, and segments requiring verification, respectively. The final output includes the river / lake management boundary number, segment number, sequence of projected ground coordinate points, boundary category, candidate boundary number, four evaluation values, connectivity score, topology verification status, connectivity continuity, segment confidence level, water level response trajectory index, set of generated image numbers, and source markers. Source markers include original valid boundary segments, automatically connected segments, smoothing correction segments, coordinate snapping correction segments, and gap segments requiring verification.

[0101] For example, if the maximum offset after smoothing is 0.08m and the arrangement of adjacent objects remains unchanged, the smoothing result can be retained; if there is a true 70° turn in the gate position, the inflection point should be retained.

[0102] All the above formulas are performed using dimensionless numerical calculations; the relevant formulas are based on empirical models that approximate the real situation, obtained through extensive data collection and software simulation fitting. The preset parameters and thresholds involved in the formulas can be conventionally set and adjusted by those skilled in the art according to the physical constraints of the actual application scenario.

[0103] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0104] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0105] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of river and lake management boundaries based on aerial survey images, characterized in that, Includes the following steps: S1. Acquire aerial survey images, camera attitude data, ground positioning data, and ground elevation data of the target river and lake area. After correction, registration, and image recognition, obtain standard aerial survey images, coastal feature identification maps, and candidate boundary sets. S2. Based on ground elevation data, coastal feature identification map and candidate boundary set, determine the allowable and prohibited change areas of virtual flooding, virtual receding and local water accumulation. Under the condition that the main outline of stable features remains unchanged, generate a group of generated images of water body coverage status, and match the candidate boundaries in the generated image group to form a cross-image boundary response sequence. S3. Calculate the position, length, category and changes of adjacent objects on both sides of the candidate boundary based on the cross-image boundary response sequence to form a water level response trajectory. Eliminate water level response type candidate boundaries and perform segmented identity stability evaluation on the remaining candidate boundaries to obtain a set of identity stable candidate boundaries. S4. Establish a two-sided boundary analysis zone for the candidate boundary segments in the identity-stable candidate boundary set, and perform isolation reconstruction and open detection reconstruction respectively. Calculate the degree of cross-boundary reconstruction leakage based on the reconstruction difference before and after opening the image features on the other side.

2. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 1, characterized in that, Also includes: S5. Select effective boundary segments based on the stability of identity and the degree of leakage in cross-boundary reconstruction. Establish connection relationships based on endpoint distance, extension direction, boundary category and adjacent objects on both sides. After topology verification, constrained smoothing and ground coordinate correction, output the confidence of river and lake management boundaries and sub-segments.

3. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 1, characterized in that, S1 specifically includes: The original aerial survey images, camera attitude data, ground positioning data, camera intrinsic parameters and ground elevation data are acquired, and after correction, registration, pixel scale unification and brightness equalization, standard aerial survey images are obtained. Image recognition models are used to identify water bodies, banks, dams, roads, vegetation, exposed surfaces, and coastal structures. The contours, color distributions, main texture directions, and spatial adjacency relationships of these areas are extracted to form a coastal feature identification map.

4. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 3, characterized in that, Also includes: Based on the coastal feature identification map, the edges of water bodies, bank slopes, dike tops, roads, vegetation transitions, and structures are extracted and connected according to the endpoint distance, extension direction, boundary category, and adjacent features on both sides to form a set of candidate boundaries.

5. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 1, characterized in that, S2 specifically includes: Based on ground elevation data, coastal feature identification maps, and candidate boundary sets, the change distances, allowable change areas, prohibited change areas, and stable feature outlines of virtual flooding, virtual receding water, and local water accumulation are determined, forming a counterfactual water body change constraint set. Based on the counterfactual water body change constraint set, extended water body, synthetic receding water and low-lying water accumulation images are generated. After edge transition and strategy optimization, a group of generated images that pass the verification of stable ground feature contour displacement and category retention rate is obtained. Image recognition and edge extraction are performed on the generated image group. Candidate boundaries are matched based on distance, direction, boundary category and adjacent objects on both sides. The matching status is recorded according to the water change type and level to form a cross-image boundary response sequence.

6. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 1, characterized in that, S3 specifically includes: Based on the cross-image boundary response sequence, the lateral position offset, continuous length change, boundary category change, and changes of adjacent objects on both sides of the candidate boundary are calculated to obtain the water level response characteristics. The water level response characteristics are arranged in the order of virtual receding water, original water body state, virtual rising water and local water accumulation to form a water level response trajectory. Based on the synchronous movement of the candidate boundary and the edge of the water body, the coverage by the water body and the category replacement, the water level response candidate boundary and the initial stable candidate boundary are distinguished.

7. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 6, characterized in that, Also includes: The initial stable candidate boundary is divided into segments. The stability of the identity is calculated based on the location, length, category, retention of adjacent objects on both sides, and effective detection. The set of stable candidate boundaries is then selected.

8. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 1, characterized in that, S4 specifically includes: A two-sided boundary analysis band is established for the candidate boundary segments in the identity-stable candidate boundary set, and occlusion windows are set on both sides of the boundary without crossing the boundary to form a two-sided reconstructed input image; A dual-branch mask image reconstruction model is adopted. For each occluded window, isolation reconstruction by cutting off the image features on the opposite side and open detection reconstruction by opening the image features on the opposite side are performed respectively. The cross-boundary dependency gain is calculated based on the error difference between the two reconstruction results, and a dual-boundary response map is generated. Based on the candidate boundary number and segment number, the cross-boundary response pixel ratio, response direction, maximum continuous response length and average response intensity are statistically analyzed. The cross-boundary reconstruction leakage degree and reconstruction evidence integrity are calculated to form the segment reconstruction leakage results.

9. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 2, characterized in that, S5 specifically includes: The set of candidate boundaries for stable identities is correlated with the leakage results of segment reconstruction. The segment quality is evaluated based on the degree of identity stability, the degree of leakage in cross-boundary reconstruction, and the completeness of the two types of evidence, and effective boundary segments are selected. The connection relationship is established based on the endpoint distance, extension direction, boundary category and adjacent objects on both sides of the effective boundary section. The connection curve is then checked by crossing the main water body, exchanging left and right banks and repeated segmentation to form the initial river and lake management boundary.

10. The intelligent identification method for river and lake management boundaries based on aerial survey images according to claim 9, characterized in that, Also includes: The initial river and lake management boundaries are subjected to constrained smoothing, ground coordinate correction, and topological verification. Based on the stability of identity, the degree of leakage in cross-boundary reconstruction, the completeness of evidence, and the degree of connection continuity, the confidence of each segment is calculated, and the river and lake management boundary identification results are output.

Images