Intelligent detection method for precision of river topographic map based on multivariate image fusion

Abstract

The invention relates to the technical field of image processing, in particular to a river topographic map precision intelligent detection method based on multivariate image fusion. The method comprises the following steps: collecting multivariate image data of a river channel region to be detected, wherein the multivariate image data comprises an optical image, a synthetic aperture radar image and laser radar point cloud data; performing multivariate image preprocessing on the multivariate image data, and performing adaptive fusion on the preprocessed multivariate image data to obtain a fused image; extracting river topographic features in the fused image, and executing precision matching verification in combination with pre-acquired mapping topographic features in a to-be-detected river topographic map; and calculating a comprehensive precision grade of the precision matching verification result by using a pre-constructed error evaluation model so as to obtain a precision detection report of the topographic map of the river channel to be detected. According to the invention, automation of key feature extraction and precision verification can be realized, the detection efficiency is greatly improved, and the labor cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an intelligent method for detecting the accuracy of river topographic maps based on multi-image fusion. Background Technology

[0002] River topographic maps are crucial foundational data for hydrological monitoring, river management, and water conservancy project construction. Their accuracy directly impacts the scientific validity and reliability of decisions regarding river management and water security. Traditional methods for verifying the accuracy of river topographic maps primarily rely on manual on-site verification or comparison of single images, which suffers from low efficiency, high labor intensity, and significant susceptibility to subjective factors.

[0003] With the development of remote sensing technology, multi-source imagery data such as optical images, synthetic aperture radar (SAR) images, and lidar are widely used in the surveying and mapping field. Multi-source imagery data is complementary: optical images are rich in texture information, SAR images have all-weather operation capabilities, and lidar can provide high-precision elevation data. However, a mature multi-source image fusion technology for intelligent detection of river topographic map accuracy has not yet been established. Existing technologies often employ only a single image data source, resulting in limited information and weak anti-interference capabilities; some fusion algorithms have fixed weight allocations, making them unsuitable for complex scenarios with varying river topography; furthermore, the automation level of feature extraction and error assessment is low, making it difficult to meet the needs of large-scale, high-precision river topographic map detection. Summary of the Invention

[0004] Therefore, it is necessary for the present invention to provide an intelligent detection method for the accuracy of river topographic maps based on multi-image fusion, so as to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, a method for intelligent detection of river topographic map accuracy based on multi-source image fusion includes the following steps: Step S1: Collect multi-dimensional image data of the river area to be detected, including optical images, synthetic aperture radar images, and lidar point cloud data. Step S2: Perform multi-dimensional image preprocessing on the multi-dimensional image data, and perform adaptive fusion on the preprocessed multi-dimensional image data to obtain a fused image; Step S3: Extract the river channel topographic features from the fused image and perform accuracy matching verification by combining them with the map topographic features in the pre-acquired river channel topographic map to be detected; Step S4: Using the pre-built error assessment model, calculate the comprehensive accuracy level of the accuracy matching verification results to obtain the accuracy detection report of the river topographic map to be detected.

[0006] This application employs multi-source image fusion technology, integrating the complementary advantages of optical imagery, SAR imagery, and LiDAR point cloud data. This solves the problem of insufficient information from a single data source and improves the reliability of feature extraction. The improved weighted fusion algorithm achieves adaptive weight allocation, adapting to complex scenarios with different river topography, resulting in higher information richness and clarity of the fused images. By combining a deep learning semantic segmentation model with an intelligent matching algorithm, the extraction of key features and accuracy verification are automated, significantly improving detection efficiency and reducing labor costs. A multi-dimensional error evaluation model is constructed, which can comprehensively reflect the planar accuracy, elevation accuracy, and boundary accuracy of the topographic map. The detection results are accurate and reliable, meeting the needs of river surveying projects with different accuracy levels. Attached Figure Description

[0007] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the steps of the intelligent detection method for river topographic map accuracy based on multi-image fusion according to the present invention; Figure 2 This is a block diagram of the multi-image data acquisition task architecture in an embodiment of the present invention; Figure 3 This is a schematic diagram of the marking of permanent markers in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0008] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0009] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0010] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0011] To achieve the above objectives, please refer to Figures 1 to 3 This invention provides an intelligent detection method for the accuracy of river topographic maps based on multi-source image fusion, the method comprising the following steps: Step S1: Collect multi-dimensional image data of the river area to be detected, including optical images, synthetic aperture radar images, and lidar point cloud data. In one embodiment of the present invention, the acquisition of multi-dimensional image data of the river channel area to be detected specifically includes: optical images acquired through satellite remote sensing or aerial photography, with a resolution of not less than 0.5m, used to capture the texture details and surface cover information of the river channel topography; synthetic aperture radar (SAR) images acquired using a synthetic aperture radar system, which has all-weather and all-time imaging capabilities, used to acquire the structural features and topographic relief information of the river channel topography; and lidar point cloud data acquired through a lidar system, with a point cloud density of not less than 10 points / square meter, used to provide high-precision topographic elevation information.

[0012] Step S2: Perform multi-dimensional image preprocessing on the multi-dimensional image data, and perform adaptive fusion on the preprocessed multi-dimensional image data to obtain a fused image; In a further embodiment of the present invention, the acquired multi-dimensional image data is preprocessed as follows: permanent landmarks such as bridge piers, dam markers, and river boundary markers along the riverbank are selected as ground control points. Gauss-Kruger projection transformation is used to unify all image data to a geodetic coordinate system (such as CGCS2000) to eliminate errors caused by coordinate system differences. Atmospheric correction and radiometric calibration are performed on optical images to remove the influence of atmospheric scattering and uneven illumination. Radiometric calibration and speckle noise suppression are performed on SAR images. An adaptive median filtering algorithm is used to denoise both optical and SAR images, preserving terrain features while removing noise interference. A statistical filtering algorithm is used to remove discrete noise points from LiDAR point cloud data. Based on the SIFT feature matching algorithm, key feature points of each image are extracted, and mismatched points are eliminated using the RANSAC algorithm, achieving precise spatial alignment of optical, SAR, and LiDAR point cloud data. An improved weighted fusion algorithm is used to fuse the preprocessed multi-dimensional images to achieve information complementarity: the information entropy of optical, SAR, and LiDAR raster images is calculated separately. Standard deviation Contrast Information entropy reflects the richness of information in the image, standard deviation reflects the dispersion of gray values, and contrast reflects the clarity of terrain features. If there are no conflicts among the indicators, the entropy weighting method can be used to objectively assign weights to these three indicators, obtaining the weights of each image under different indicators. If there are conflicts, the weights are allocated and adjusted according to the indicator priority rules. Finally, the final fusion weight coefficients are obtained through linear weighting calculation. To ensure that the weight allocation adapts to the information characteristics of different images; according to the formula Perform pixel-level fusion, where To merge images Pixel value at that location, For the first Original-like images The pixel values ​​at each location are used to create a fused image that combines texture details, structural features, and high-precision elevation information.

[0013] Step S3: Extract the river channel topographic features from the fused image and perform accuracy matching verification by combining them with the map topographic features in the pre-acquired river channel topographic map to be detected; In a further embodiment of the present invention, key topographic features are extracted based on the fused image and the topographic map of the river channel to be detected: An improved U-Net semantic segmentation model is used, with an attention mechanism added to focus on key areas of the river channel, extracting feature points such as river boundaries, slope toes, levee apexes, and river channel turning points, as well as contour information of areas with abrupt elevation changes (such as levee and beach edges); the LiDAR point cloud data is rasterized to obtain a digital elevation model (DEM), and elevation values ​​are extracted as feature parameters; the topographic map of the river channel to be detected is vectorized to extract corresponding information such as river boundaries, feature point coordinates, and elevation values, ensuring consistency with the feature types extracted from the fused image. The features extracted from the fused image are then matched and verified with the features of the topographic map in a multi-dimensional manner: the Iterative Closest Point (ICP) algorithm is used to match the coordinates of feature points in the fused image and the topographic map, calculating the planar coordinate deviation of the corresponding points. The Hausdorff distance was used to calculate the degree of agreement between the river channel boundaries extracted from the fused image and the topographic map boundaries; a smaller distance value indicates better boundary consistency. Elevation values ​​(from LiDARDEM) of the corresponding area in the fused image were extracted and compared with the elevation values ​​in the topographic map, and the absolute elevation error was calculated. With relative error ;in The river channel topography features, To map the topographic features.

[0014] Step S4: Using the pre-built error assessment model, calculate the comprehensive accuracy level of the accuracy matching verification results to obtain the accuracy detection report of the river topographic map to be detected.

[0015] In a further embodiment of the present invention, the construction of the error assessment model first uses the multi-dimensional error results obtained in the accuracy matching and verification stage as the basic input. The input includes the planar coordinate deviation of river channel topographic feature points, absolute elevation error, relative elevation error, and the degree of conformity of the river channel boundary contour. To ensure the reasonable contribution of various error indicators in the comprehensive assessment, the analytic hierarchy process (AHP) is used to determine the weights of each indicator. Specifically, a three-layer evaluation structure is constructed, where the target layer is the comprehensive accuracy level of the river channel topographic map, the criterion layer includes three evaluation criteria: planar accuracy, elevation accuracy, and morphological consistency, and the indicator layer corresponds to feature point matching accuracy, boundary conformity, absolute elevation error, and relative elevation error, respectively. By comparing the importance of the criterion layer and the indicator layer pairwise, a judgment matrix is ​​formed, and the consistency check is used to ensure the rationality of the matrix. When the consistency ratio is less than 0.1, the obtained weights are confirmed to be valid, thus obtaining the weight coefficients of each error indicator in the comprehensive assessment. After determining the weights, a fuzzy comprehensive evaluation method is introduced to perform level mapping processing on each error indicator. Specifically, based on the accuracy requirements for different mapping scales in the river surveying engineering specifications, membership functions are constructed for each error index, converting continuous error values ​​into membership degrees corresponding to three evaluation levels: "Excellent," "Qualified," and "Unqualified." For example, for the plane coordinate deviation index, when the error value is less than 0.5 times the allowable error threshold, its membership degree to the "Excellent" level is close to 1; when the error value is near the allowable error threshold, its membership degree to the "Qualified" level increases; and when the error value exceeds the allowable error threshold, its membership degree to the "Unqualified" level dominates. The absolute elevation error, relative elevation error, and boundary conformity all use membership functions with the same structure, but their threshold parameters are set according to the engineering accuracy requirements of the corresponding indexes. Furthermore, to adapt to the accuracy requirements of different river surveying tasks, the error assessment model introduces an adaptive adjustment mechanism for the accuracy threshold. Specifically, based on the scale information of the input river topographic map and the corresponding industry standard requirements, the allowable error range of each indicator is dynamically adjusted. For example, under a 1:500 scale, the horizontal error threshold is limited to no more than 0.1m, and the elevation error threshold is limited to no more than 0.05m. When the scale changes, the hierarchical node parameters in the membership function are adjusted simultaneously to ensure that the evaluation results are consistent with the actual engineering accuracy standards and avoid evaluation distortion caused by fixed thresholds. Finally, the membership results of each error indicator are weighted and summarized with the weight coefficients determined by the analytic hierarchy process (AHP) to form a comprehensive evaluation vector. The comprehensive accuracy level of the river topographic map to be tested is determined according to the principle of maximum membership degree or the weighted scoring rule. Based on this comprehensive accuracy level and the evaluation results of each individual indicator, an accuracy test report is generated. The report clearly lists the numerical results of each error indicator, the corresponding level judgment, and the comprehensive accuracy conclusion.

[0016] Figure 2This is a block diagram of the multi-image data acquisition task architecture in an embodiment of the present invention; the radar system / satellite remote sensing 201 is connected to the data processing terminal 203 through the network 202, which can be a wide area network or a local area network, or a combination of both.

[0017] Optionally, after obtaining the accuracy report of the topographic map of the river channel to be inspected, the report may also include: Based on the accuracy test report, the error distribution in the topographic map of the river channel to be tested is statistically analyzed; The error distribution is correlated with the corresponding locations in the river channel topographic features, and an error distribution map is constructed for visualization.

[0018] In one embodiment of the present invention, after obtaining an accuracy inspection report including planar coordinate deviation, absolute elevation error, relative elevation error, and boundary conformity, spatial statistical analysis is performed on the various error results in the report. Specifically, using the uniform raster resolution used in the construction of the fused image as the statistical unit, preferably with a raster size of 5m×5m, the number of feature points participating in the accuracy matching verification within each raster is counted. When the number of effective feature points within a raster is not less than 3, the average planar error value and the average elevation error value corresponding to that raster are calculated. When the number of effective feature points is insufficient, weighted compensation from adjacent rasteres is introduced, with the weight decreasing linearly with spatial distance from 0.6 to 0.8. An error statistical matrix covering the entire river channel area is formed in the above manner. The obtained error statistical matrix is ​​spatially correlated with the extracted river channel topographic features in the fused image to construct an error distribution map for visualization. Specifically, based on the coordinates of feature points and the contours of elevation abrupt change areas recorded in the aforementioned river channel topographic features, the error values ​​in the error statistics matrix are mapped to the corresponding river boundary lines, levee crest lines, and elevation abrupt change contours. A hierarchical rendering method is used to express the error values, with a first-level threshold of 0.05m for planar errors and 0.02m for elevation errors, each with at least four color ranges. When both planar and elevation errors exceed the limits in a given area, a composite error indicator is overlaid. The final generated error distribution map and accuracy test report share the same coordinate reference and can be directly used to identify areas of concentrated accuracy anomalies in the river channel topographic map.

[0019] It is worth noting that the obtained visualization results need to be sent back to the accuracy test report as part of that report.

[0020] Optionally, the multivariate image preprocessing in step S2 includes: Permanent landmarks are selected based on optical image data from multi-source imagery, and these permanent landmarks are used as ground control points. In one embodiment of the present invention, based on the acquired optical image data, structural target detection processing is performed on the riverbank area, preferably selecting artificial structures such as bridge piers, embankment control piles, and fixed boundary markers as candidate objects. By analyzing the geometric center offset of the candidate object in two consecutive images, when the offset distance is less than 0.2m and the change rate of edge morphology parameters is less than 5%, the structure is determined to have temporal stability; further, its geometric center coordinates are extracted as control point coordinates, and the control points are required to be spaced no more than 500m apart in the longitudinal direction of the river and to be distributed at least once on each side laterally.

[0021] Based on the spatial distribution and relative positional relationship between ground control points, the geometric transformation parameters of multi-dimensional image data are determined, thereby unifying the multi-dimensional image data to a preset geodetic coordinate system; In a further implementation, a geometric transformation relationship for the multi-source imagery is constructed based on the established set of ground control points. Specifically, an affine-quadratic polynomial combined transformation model is adopted, where the affine component corrects for translation, rotation, and scale differences, and the quadratic term compensates for local nonlinear distortions. The transformation parameters are calculated using the least squares method, and their validity is confirmed when the mean residual of the control points is less than 0.15m. A Gauss-Kruger projection transformation is then used to uniformly project the optical imagery, synthetic aperture radar imagery, and lidar point cloud data onto a preset geodetic coordinate system (such as CGCS2000). The coordinate reference used remains consistent with that used in subsequent accuracy detection stages, thereby eliminating spatial reference differences between different data sources.

[0022] Radiometric correction is performed on optical image data and synthetic aperture radar image data in multi-element image data unified to the geodetic coordinate system, and denoising processing is performed on radiometric correction results and lidar point cloud data unified to the geodetic coordinate system. In a further implementation, targeted quality correction processing is performed on the multi-source image data after coordinate unification. For optical images, radiometric calibration is performed based on imaging time and solar elevation angle parameters, and brightness unevenness is corrected using an atmospheric scattering model. For synthetic aperture radar images, radiometric normalization is performed, and an adaptive filtering method with a window size of 5×5 pixels is used to suppress speckle noise. For lidar point cloud data, outlier removal is performed based on neighborhood statistical distribution; points whose elevation values ​​deviate from the neighborhood mean by more than 2.5 standard deviations are removed. After processing, the data achieve comparability in terms of noise level and radiometric consistency.

[0023] It is worth noting that the atmospheric scattering model is constructed based on the radiative transfer mechanism of optical remote sensing imaging. Its theory originates from the classical atmospheric radiative transfer model, and parameters are simplified to suit the actual imaging conditions of the river region. Specifically, imaging parameters such as imaging time, solar altitude angle, and sensor observation angle are first retrieved from the metadata of the optical image. Combined with the visibility level at the time of image acquisition, the atmospheric assumption of single-scattering dominance is selected. Based on this, atmospheric scattering is decomposed into path radiation and surface reflection attenuation components. Path radiation is estimated by analyzing the minimum brightness value of deep water bodies or shadowed areas in the image, while surface reflection attenuation is corrected using the relationship between the solar altitude angle and the sensor's observation geometry. Finally, the estimated scattering compensation parameters are applied pixel-by-pixel to correct the radiance values ​​of the optical image, thus forming an atmospheric scattering correction model suitable for the multi-image preprocessing stage of this invention.

[0024] Feature matching is performed on the feature points of the denoised lidar point cloud data, optical image data and synthetic aperture radar image, and mismatched points in the feature matching results are removed to obtain preprocessed multi-dimensional image data.

[0025] In a further implementation, spatial feature points are extracted from the denoised multivariate image data, and cross-source matching is performed. Corner points and edge intersections are extracted from optical and radar images as two-dimensional features, and spatial inflection points with significant elevation change rates are extracted from lidar point clouds as three-dimensional features. These are then uniformly converted into two-dimensional projected coordinates for matching. After initial matching by establishing feature point description vectors, a random sampling consistency verification mechanism is introduced. When the reprojection error of a matching point is greater than 0.3m, it is determined to be a mismatched point and is removed. The final retained feature matching results serve as the constraint basis for the spatial consistency of the multivariate images, forming a preprocessed multivariate image dataset.

[0026] Optionally, permanent markers may be selected, including: Detect the target area of ​​the structure in the optical image data, and screen the high-contrast target areas in the target area of ​​the structure as candidate target areas; In one embodiment of the present invention, edge enhancement processing is performed on optical image data to extract obvious straight edge and broken edge combination, and a target area of ​​the structure is formed by connected component analysis. On this basis, the difference between the pixel gray value in the target area of ​​the structure and the gray value of the surrounding background area is calculated. When the difference is greater than 0.25 and the region boundary has a continuous length of more than 8 pixels in at least two orthogonal directions, the region is determined to have significant contrast features and is identified as a high-contrast target area, which is used as a candidate target area for subsequent stability analysis.

[0027] Calculate the position offset and geometric features of each structure target in the candidate target area in continuous time phases; if the position offset of any structure target in different time phases is less than the preset stability threshold and the geometric features remain consistent, then the structure target is determined to be a candidate permanent marker. In a further embodiment, in optical images with an imaging interval of more than 30 days between at least two imaging phases, the geometric center coordinates of the same structure target are calculated, and its planar position offset is calculated accordingly. Simultaneously, geometric morphological characteristic parameters such as the area, aspect ratio, and principal orientation angle of the structure target are extracted. If the position offset of the structure in each time phase is less than 0.3 m, and the rate of change of the geometric morphological characteristic parameters does not exceed 5%, then the structure target is determined to have long-term stability in both spatial position and morphology, and is identified as a candidate permanent marker.

[0028] Identify the geometric center or unique spatial pointing feature point among the candidate permanent markers, and select the candidate permanent markers with the geometric center or unique spatial pointing feature point as permanent markers.

[0029] In a further implementation, key feature points for spatial registration are extracted from the identified candidate permanent markers. For structures with approximately regular contours and strong symmetry, the geometric center of their closed contours is calculated as spatial control points. For structures with irregular contours but obvious directional features, their principal axis endpoints or significant corner points are extracted as unique spatial directional feature points. To ensure the reliability of the control point positioning, the selected feature points must have a positioning uncertainty of less than one pixel in the image and a spatial distribution where the distance between adjacent feature points is no less than 200 m. Finally, candidate permanent markers that meet the above conditions are determined as permanent markers and used as ground control points for subsequent multivariate image geometric correction.

[0030] Figure 3 This is a schematic diagram of the marking of permanent markers in an embodiment of the present invention; such as Figure 3 As shown, the pink area represents the high-contrast target area detected in the optical image; the orange area is a schematic diagram of the candidate permanent marker locations selected within the high-contrast target area after stability and geometric feature consistency determination.

[0031] Of particular importance is the detection of structural target areas in optical image data, including: Edge feature extraction and connectivity analysis are performed on optical image data to screen candidate target regions that have a combination of straight edge or polygonal edge features and form a closed or semi-closed contour within a preset pixel scale. In one embodiment of the present invention, a multi-directional gradient operator is used to calculate the edge response of the image, extracting the gradient magnitudes in the horizontal, vertical, and diagonal directions respectively, and marking pixels with gradient magnitudes greater than 1.2 times the overall gradient mean of the image as edge points. Subsequently, connectivity analysis is performed on the edge points, using 8-neighborhood as the pixel connection rule to aggregate spatially continuous edge points into edge segments, and fitting straight lines or polygonal structures to these edge segments. The fitting results are then evaluated for contour closure; when an edge segment forms a closed or semi-closed contour within a scale of 30–300 pixels, the corresponding region is marked as a candidate target region.

[0032] If the candidate region maintains a consistent edge distribution within the range of adjacent pixels, the variation of geometric morphological parameters within a preset spatial scale is less than the morphological stability threshold, and the difference between the brightness value or color component of the candidate target region and the surrounding background region is greater than the preset contrast threshold, then the candidate target region is determined to be the target region of the structure.

[0033] In a further implementation, an edge direction histogram is plotted within the candidate region. If the main direction accounts for at least 60%, the edge direction distribution is considered consistent. Simultaneously, the area, perimeter, and aspect ratio of the candidate region are calculated, and their variation is calculated within adjacent pixel scales (3×3 pixel neighborhood or 5×5 pixel neighborhood). If the variation rate is less than 10%, the morphology is considered stable. Furthermore, the difference between the average brightness value or main color component of the candidate region and its outer 3-pixel background region is compared. If the difference is greater than 0.2, significant contrast features are confirmed. Candidate regions meeting the above conditions are determined as the target area for the structure.

[0034] Optionally, methods for filtering high-contrast target regions include: The grayscale information of the corresponding pixels in the target area of ​​the structure is analyzed to extract the grayscale value distribution of the target area of ​​the structure. In one embodiment of the present invention, a pixel grayscale histogram statistical method within a region is used to calculate the grayscale value of each pixel and its distribution probability in the entire candidate region, and the distribution is smoothed to reduce noise interference; at the same time, the grayscale average value, standard deviation and local gradient within the region are calculated to form the grayscale distribution characteristics of the region.

[0035] The local grayscale variation range of each target area of ​​a structure is calculated based on the grayscale value distribution. If the local grayscale variation range of any target area of ​​a structure is higher than the average grayscale variation range of the background area in the optical image data within a preset neighborhood, then the target area of ​​the structure is determined to be a candidate target area. In a further implementation, a sliding window method (5×5 pixels in neighborhood) is used to calculate the grayscale standard deviation of each local window within the candidate region, which is taken as the local grayscale variation amplitude. The mean and maximum values ​​of the grayscale standard deviations of all local windows within the entire candidate region are then calculated. Simultaneously, several background regions are selected within a 3-pixel radius outside the candidate region, and the average grayscale standard deviation of these background regions is calculated as the mean background grayscale variation. If the local grayscale variation amplitude is consistently higher than 1.2 times the mean background grayscale variation amplitude within its neighborhood, the region is considered to have significant texture differences and is thus identified as a candidate target region. The continuous boundaries and boundary lengths of candidate target regions in each direction are detected; if any candidate target region has continuous boundaries in at least two mutually perpendicular directions and the boundary length exceeds a preset minimum scale threshold, then the candidate target region is determined as a high-contrast target region.

[0036] In a further implementation, an edge tracking algorithm is used to detect the boundary continuity of the region in both the horizontal and vertical directions, and the boundary length is calculated. When the continuous boundary length of any candidate region in both the horizontal and vertical directions is greater than 8 pixels (corresponding to 4m in reality), combined with local grayscale contrast information, the candidate region is finally determined to be a high-contrast target region. Optionally, performing adaptive fusion in step S2 includes: The information entropy, standard deviation, and contrast of the preprocessed multivariate image data are calculated separately, and the information entropy, standard deviation, and contrast are reconciled according to the preset index priority rules to obtain the fusion weight. In one embodiment of the present invention, the information entropy H is expressed by the formula... Calculate the entropy and standard deviation of the probability distribution of the pixel grayscale histogram. The degree of dispersion of grayscale values ​​in the calculated area reflects the richness of texture; contrast C reflects the clarity of terrain features through local gradient magnitude or grayscale range. Index values ​​are calculated separately for each type of image within the same geographic region, and the corresponding matrices are saved.

[0037] The fusion weights are linearly weighted to obtain the fusion weight coefficients; In a further implementation, conflict coordination is performed on the calculated information entropy, standard deviation, and contrast indicators. First, the three indicators are normalized, mapping the values ​​to the range of 0 to 1. Then, the basic values ​​of the indicator weights are determined according to the preset indicator priority rules (information entropy 50%, standard deviation 30%, contrast 20%). For areas where there is an inverse distribution among the indicators and the difference coefficient is greater than 0.2, the weights are adjusted using local terrain complexity (calculated through LiDARDEM grid slope, normalized from 0 to 1). When the complexity is ≥0.5, the contrast weight is increased to 40%, the information entropy weight is reduced to 40%, and the standard deviation weight remains unchanged, ensuring a more reasonable allocation of indicators in conflict areas, and obtaining the final fused weight matrix.

[0038] In another implementation, if there is no inverse distribution or a difference coefficient less than 0.2 among the three types of indicators after normalization, it is considered that no conflicting indicators have occurred. In this case, the entropy weight method is used to calculate the weight of each indicator: the entropy contribution of each image information is calculated based on the normalized indicators, and the entropy weight coefficient is further calculated so that the indicators with richer information content receive higher weights. For example, the calculated weights are 0.48 for optical images, 0.32 for SAR images, and 0.20 for LiDAR.

[0039] According to the fusion weight coefficient, pixel fusion is performed on the preprocessed multi-dimensional image data to obtain the fused image.

[0040] In a further implementation, the formula is adopted. ,in For the first Image-like images pixel value, To correspond to the fusion weighting coefficients, during the fusion process, the LiDAR raster data can first be normalized in elevation to match its numerical range with that of optical and SAR images, thereby obtaining a fused image that combines texture details, structural features and high-precision elevation information.

[0041] Optionally, resolving conflicting performance indicators includes: Information entropy, standard deviation, and contrast are normalized separately to calculate the difference coefficient of the same index between different images; the index includes information entropy, standard deviation, and contrast. If the difference coefficient of any indicator between different images is greater than the preset difference threshold, and the indicator is inversely distributed between the images, then the indicator is determined to be a conflict indicator. In one embodiment of the present invention, each index is normalized to map the values ​​to the [0,1] interval, eliminating dimensional and amplitude differences between different images. By calculating the difference coefficient of the same index between different images, such as |H_optical−H_SAR|, |C_optical−C_LiDAR|, etc., when the difference coefficient of any index is ≥0.3 and shows a high-low inverse distribution (e.g., high information entropy in optical images and low information entropy in SAR images), it is determined that the index is in conflict.

[0042] The basic weights of each indicator are assigned according to the indicator priority rules, and the local terrain complexity of the image area corresponding to the image conflict indicator is calculated. The basic weights of each indicator in the corresponding image region of the conflict index are adjusted according to the local terrain complexity, and cross-validation is performed on the adjusted basic weights to obtain the fusion weights.

[0043] In a further implementation, conflict coordination is performed on conflict indicator regions. First, basic weights are assigned according to indicator priority rules: information entropy. 50% Contrast 30%, standard deviation A weighting factor matrix of 20% is formed; then, a sliding window of 3×3 is used to calculate the local terrain complexity K (based on LiDARDEM raster slope normalization 0~1). For areas where K≥0.5, the contrast is adjusted. Weight increased to 40%, information entropy Adjusted to 40%, standard deviation The weight allocation is maintained at 20%; for regions with K < 0.5, the original weight allocation is maintained. Finally, a cross-validation mechanism is introduced to iteratively optimize the weight allocation by comparing the extraction accuracy of river boundaries and elevation change areas before and after the fusion of conflict areas, ensuring that the clarity of terrain features after the fusion of conflict areas is ≥ 90%, and generating the final fusion weights.

[0044] It is worth noting that the basic weights of each indicator in the corresponding image area for adjusting the conflict index include: The basic weights of each indicator in the indicator priority rules are as follows: information entropy has a basic weight of 50%, standard deviation has a basic weight of 30%, and contrast has a basic weight of 20%. If the terrain complexity is ≥0.5, increase the base weight of contrast to 40% and adjust the base weight of information entropy to 40% accordingly. If the terrain complexity is less than 0.5, the basic weights of each indicator are maintained.

[0045] Optionally, step S3 involves extracting river channel topographic features from the fused image, including: The lidar point cloud data in the fused image is rasterized to obtain a digital elevation model of the river area to be detected, and the elevation values ​​of each area in the digital elevation model are extracted as elevation feature parameters. In one embodiment of the present invention, lidar point cloud data is projected onto a two-dimensional grid with a grid resolution of 1 meter × 1 meter. For each grid cell, the elevation values ​​of all point clouds within that cell are statistically analyzed, and the average value is taken as the grid elevation. Simultaneously, the elevation standard deviation is recorded for subsequent elevation anomaly removal and terrain complexity calculation. The resulting digital elevation model (DEM) forms an elevation matrix.

[0046] Based on the variation amplitude of elevation feature parameters of adjacent grid areas, identify elevation change areas and elevation distribution patterns in fused images, and determine the corresponding river structure areas based on the elevation distribution patterns. In a further implementation, elevation abrupt change regions are identified based on a digital elevation model. The elevation difference between each grid cell and its eight neighboring grid cells is calculated. ,like ≥0.3 meters and local elevation gradient (in terms of If the normalized (raster side length) value is ≥0.15, it is marked as an elevation change region. Elevation gradient analysis of continuous rasters can distinguish between gentle river channels (gradient <0.05) and steep river channels (gradient ≥0.05), thereby generating river channel structure regions.

[0047] Based on the corresponding pixel coordinates of the elevation change region and the river channel structure region, semantic segmentation is performed on the corresponding pixel regions in the fused image to extract the river channel structure feature points and the contour information of the elevation change region. In a further implementation, semantic segmentation is performed on the fused image based on the corresponding pixel coordinates of the elevation change regions and the river channel structure regions. An improved U-Net model is employed, using the optical, SAR, and DEM channels of the fused image as input, and introducing an attention gate mechanism to focus on key regions such as river boundaries, levee apexes, and river bends. The model outputs river channel structure feature points and contour masks of elevation change regions. Furthermore, morphological operations (opening and closing operations) are used to remove isolated pixels, resulting in continuous and clear river channel structure feature points and elevation change contours within the river channel boundary lines.

[0048] Of particular importance is the rasterization process performed on the LiDAR point cloud data in the fused imagery, which includes: Based on the preset geodetic coordinate system, the lidar point cloud data in the fused image is projected onto a plane to map the three-dimensional point cloud horizontal coordinates in the lidar point cloud data to a preset two-dimensional regular grid. In this embodiment, a two-dimensional regular grid is constructed based on the preset accuracy requirements for river channel mapping, with a grid resolution set to 1m×1m. Using geodetic coordinate system parameters, the three-dimensional coordinates of the lidar point cloud data are horizontally projected onto the two-dimensional regular grid, retaining only its planar coordinate components, so that each lidar point can be uniquely mapped to the corresponding grid cell.

[0049] Calculate the standard deviation of elevation values ​​of point cloud data in each grid cell of the mapped two-dimensional regular grid. If the number of point clouds in any grid cell is not less than the preset minimum number of points threshold and the standard deviation of elevation values ​​does not exceed the preset elevation dispersion threshold, then statistically calculate the elevation values ​​of all point clouds falling into the grid cell and take the arithmetic mean as the representative elevation value of the grid cell. In a further embodiment, the point cloud distribution within each grid cell of the mapped two-dimensional regular grid is detected one by one, and the standard deviation parameter of the point cloud elevation value within each grid cell is calculated. If the number of point clouds within the grid cell is not less than a preset minimum point count threshold (e.g., 3 points), and the standard deviation of the elevation value does not exceed a preset elevation dispersion threshold (e.g., 0.5 m), then the point cloud distribution within the grid cell is determined to be stable, and the arithmetic mean of all point cloud elevation values ​​within the grid cell is calculated to obtain a representative elevation value.

[0050] If the number of point clouds in any raster cell is less than the minimum number of points threshold, or the standard deviation of the elevation values ​​in the raster cell exceeds the elevation dispersion threshold, then the raster cell is marked as an invalid raster cell, and the elevation values ​​of the point clouds in the invalid raster cell are interpolated using adjacent raster cells; the arithmetic mean of all the elevation values ​​of the point clouds in the invalid raster cell after interpolation is taken as the representative elevation value of the invalid raster cell. In a further embodiment, if the number of point clouds in any grid cell is less than the minimum point count threshold, or the standard deviation of its elevation value exceeds the elevation dispersion threshold, the grid cell is determined to be an invalid grid cell. For an invalid grid cell, the grid cells that have been determined to be valid in its four-neighbor or eight-neighbor areas are selected, and the elevation value is interpolated and estimated using an inverse distance weighting method. The interpolation result is then used as the representative elevation value of the invalid grid cell.

[0051] The representative elevation values ​​of each grid cell are mapped to the mapped two-dimensional regular grid to construct a digital elevation model of the river area to be detected.

[0052] In a further embodiment, the elevation values ​​represented by each grid cell are remapped to a two-dimensional regular grid according to their spatial location, forming an elevation matrix that continuously covers the river channel area to be detected. This elevation matrix also preserves the grid position index relationship, thereby constructing a digital elevation model for subsequent elevation abrupt change analysis and river channel topographic feature extraction.

[0053] Notably, an improved U-Net model was constructed for extracting river channel structure and elevation abrupt change regions. This model employs a symmetrical encoder-decoder structure. The encoder consists of four convolutional blocks, each including two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, followed by a 2×2 max-pooling layer for downsampling. The decoder consists of four upsampling blocks, each using a 2×2 upsampling convolution combined with skip connections. The feature maps of the corresponding encoder layers are concatenated with the decoder feature maps to restore spatial resolution and detail. An attention mechanism is introduced into each decoder upsampling block. By spatially and channel-weighting the feature maps transmitted from the encoder, key feature regions such as river boundaries, levee apexes, and river bends are highlighted, while background interference is suppressed. The parameters of the attention gate include a 3×3 kernel size, the same number of output channels as the concatenated feature maps, and an attention weight map generated using the sigmoid function. The model input consists of three channels of fused imagery: optical imagery, SAR imagery, and DEM elevation data. The input size can be set to 512×512 pixels, with channels normalized to the range [0,1]. The loss function is a weighted sum of Dice Loss and cross-entropy loss (weight ratio 0.6:0.4) to balance boundary accuracy and class balance. The optimizer uses Adam, with an initial learning rate of 0.001, a batch size of 8, and 100 training epochs. The learning rate can be adaptively adjusted based on validation set metrics. The training dataset consists of multi-temporal, multi-type fused river images and manually annotated river boundary and elevation change masks. During training, data augmentation (random rotation ±10°, horizontal flipping, brightness adjustment ±20%) is used to improve the model's generalization ability. After training, the model outputs a river structure feature point mask and an elevation change contour mask, which can be integrated with rasterized elevation data to form a river topographic feature dataset.

[0054] The elevation feature parameters, river channel structural feature points, and contour information of elevation change areas are integrated according to their spatial location to obtain the river channel topographic features.

[0055] In a further implementation, the extracted elevation feature parameters, river channel structural feature points, and elevation change contour information are spatially integrated. All features are unified to the same geographic coordinate system (e.g., CGCS2000), and a three-dimensional spatial index structure (KD-Tree) is established. Feature points are used as indexes to quickly associate corresponding elevation values ​​and contour locations, achieving spatial association and data integration. This ultimately forms a river channel topographic feature dataset.

[0056] Optionally, the method for obtaining the cartographic topographic features in the topographic map of the river channel to be detected in step S3 includes: The topographic map of the river channel to be detected is vectorized, and the corresponding map topographic features are extracted from the vectorization results based on the feature type of the river channel topographic features.

[0057] In one embodiment of the present invention, the topographic map of the river channel to be detected is vectorized. The rasterized topographic map or scanned map is input into Geographic Information System (GIS) software. Edge detection and line extraction tools (such as the Canny operator combined with Hough transform) are used to convert the river channel boundaries, levee crest lines, and river channel turning point outlines into vector line features. The vector resolution is set to 1 meter to ensure smooth boundaries and spatial accuracy meeting the requirement of a 1:500 topographic map with a planar error ≤0.1 meters. Using the feature types of the river channel topographic features as a reference, corresponding cartographic topographic features are extracted from the vectorized results. The vector line features are classified by attribute, with the river centerline, levee crest lines, and beach boundaries marked as different feature types, and the node coordinates and line segment lengths of each feature are extracted. For levee or bridge locations, the vector point coordinates in the map are quickly matched using an additional spatial index (such as an R-tree index) for subsequent accuracy matching with the river channel feature points extracted from the fused image. The vectorized features are further processed to support elevation comparison analysis. The vector line features are spatially overlaid with the elevation labels or contour lines of the topographic map. An elevation value h_p is added to the node points of each vector line. At the same time, the average elevation, maximum elevation change ΔH, and slope of each river channel are calculated to form a complete map river channel feature dataset.

[0058] Optionally, performing precision matching verification in step S3 includes: Match the coordinates of river channel topographic features with the coordinates of river channel structural feature points in the map topographic features, and calculate the plane coordinate deviation of the corresponding feature points; In one embodiment of the present invention, the coordinates of river channel topographic features and river channel structural feature points in the map topographic features are matched. The Iterative Closest Point (ICP) algorithm is used to pair river channel structural feature points extracted from the fused image with river nodes in the vectorized topographic map. The initial matching threshold is set to 0.5 meters, and the iteration terminates when the root mean square deviation (RMSE) change is less than 0.01 meters. The planar coordinate deviation of each pair of corresponding points is calculated. , The average deviation and maximum deviation were calculated as accuracy indicators.

[0059] Calculate the degree of agreement between the river channel topographic features and the outline information of the corresponding elevation abrupt change areas in the topographic features of the map; In a further implementation, the degree of agreement between the river channel topographic features and the contour information of the corresponding elevation change areas in the map topographic features is calculated. The contour mask of the elevation change area extracted from the fused image is spatially overlapped with the contour of the map topographic map. The Hausdorff Distance (HD) is used to measure the maximum deviation between the contours. If HD ≤ 0.3 meters, the agreement is considered good. At the same time, the contour overlap rate (Intersection over Union, IoU) is calculated as an additional agreement index.

[0060] Extract the elevation values ​​of the corresponding areas from the river channel topographic features and the map topographic features, and calculate the absolute and relative elevation errors; In a further implementation, the elevation values ​​of corresponding areas in the river channel topographic features and the map topographic features are extracted, and the absolute elevation error is calculated. With relative error ,in To fuse the imagery corresponding to the raster DEM elevation values, To mark the elevation of the topographic map, extract the elevation error of each river feature point and outline node, and calculate the regional average elevation error and maximum error.

[0061] The deviation of plane coordinates, the degree of matching of contour information, and the absolute and relative errors of elevation are used as the results of accuracy matching verification.

[0062] In a further implementation, the plane coordinate deviation, contour matching degree, absolute elevation error, and relative error are integrated into the accuracy matching verification result. Each indicator can be weighted and summarized according to preset weights (such as coordinate deviation 40%, contour matching 30%, absolute elevation error 20%, and relative error 10%) to form a comprehensive accuracy value.

[0063] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application be incorporated into the invention.

[0064] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for intelligent detection of the accuracy of river topographic maps based on multi-source image fusion, characterized in that, Includes the following steps: Step S1: Collect multi-dimensional image data of the river area to be detected, including optical images, synthetic aperture radar images, and lidar point cloud data. Step S2: Perform multi-dimensional image preprocessing on the multi-dimensional image data, and perform adaptive fusion on the preprocessed multi-dimensional image data to obtain a fused image; Step S3: Extract the river channel topographic features from the fused image and perform accuracy matching verification by combining them with the map topographic features in the pre-acquired river channel topographic map to be detected; Step S4: Using the pre-built error assessment model, calculate the comprehensive accuracy level of the accuracy matching verification results to obtain the accuracy detection report of the river topographic map to be detected.

2. The intelligent detection method for river topographic map accuracy based on multi-image fusion according to claim 1, characterized in that, After obtaining the accuracy report of the topographic map of the river channel to be inspected, it also includes: Based on the accuracy test report, the error distribution in the topographic map of the river channel to be tested is statistically analyzed; The error distribution is correlated with the corresponding locations in the river channel topographic features, and an error distribution map is constructed for visualization.

3. The intelligent detection method for river topographic map accuracy based on multi-image fusion according to claim 1, characterized in that, Step S2, multivariate image preprocessing, includes: Permanent landmarks are selected based on optical image data from multi-source imagery, and these permanent landmarks are used as ground control points. Based on the spatial distribution and relative positional relationship between ground control points, the geometric transformation parameters of multi-dimensional image data are determined, thereby unifying the multi-dimensional image data to a preset geodetic coordinate system; Radiometric correction is performed on optical image data and synthetic aperture radar image data in multi-element image data unified to the geodetic coordinate system, and denoising processing is performed on radiometric correction results and lidar point cloud data unified to the geodetic coordinate system. Feature matching is performed on the feature points of the denoised lidar point cloud data, optical image data and synthetic aperture radar image, and mismatched points in the feature matching results are removed to obtain preprocessed multi-dimensional image data.

4. The intelligent detection method for river topographic map accuracy based on multi-image fusion according to claim 2, characterized in that, The selection of permanent markers includes: Detect the target area of ​​the structure in the optical image data, and screen the high-contrast target areas in the target area of ​​the structure as candidate target areas; Calculate the position offset and geometric features of each structure target in the candidate target area in continuous time phases; if the position offset of any structure target in different time phases is less than the preset stability threshold and the geometric features remain consistent, then the structure target is determined to be a candidate permanent marker. Identify the geometric center or unique spatial pointing feature point among the candidate permanent markers, and select the candidate permanent markers with the geometric center or unique spatial pointing feature point as permanent markers.

5. The intelligent detection method for river topographic map accuracy based on multi-image fusion according to claim 4, characterized in that, Methods for filtering high-contrast target regions include: The grayscale information of the corresponding pixels in the target area of ​​the structure is analyzed to extract the grayscale value distribution of the target area of ​​the structure. The local grayscale variation range of each target area of ​​a structure is calculated based on the grayscale value distribution. If the local grayscale variation range of any target area of ​​a structure is higher than the average grayscale variation range of the background area in the optical image data within a preset neighborhood, then the target area of ​​the structure is determined to be a candidate target area. The continuous boundaries and boundary lengths of candidate target regions in each direction are detected; if any candidate target region has continuous boundaries in at least two mutually perpendicular directions and the boundary length exceeds a preset minimum scale threshold, then the candidate target region is determined as a high-contrast target region.

6. The intelligent detection method for river topographic map accuracy based on multi-source image fusion according to claim 1, characterized in that, Step S2, which involves performing adaptive fusion, includes: The information entropy, standard deviation, and contrast of the preprocessed multivariate image data are calculated separately, and the information entropy, standard deviation, and contrast are reconciled according to the preset index priority rules to obtain the fusion weight. The fusion weights are linearly weighted to obtain the fusion weight coefficients; According to the fusion weight coefficient, pixel fusion is performed on the preprocessed multi-dimensional image data to obtain the fused image.

7. The intelligent detection method for river topographic map accuracy based on multi-image fusion according to claim 6, characterized in that, Coordination of conflicting performance indicators includes: Information entropy, standard deviation, and contrast are normalized separately to calculate the difference coefficient of the same index between different images; the index includes information entropy, standard deviation, and contrast. If the difference coefficient of any indicator between different images is greater than the preset difference threshold, and the indicator is inversely distributed between the images, then the indicator is determined to be a conflict indicator. The basic weights of each indicator are assigned according to the indicator priority rules, and the local terrain complexity of the image area corresponding to the image conflict indicator is calculated. The basic weights of each indicator in the corresponding image region of the conflict index are adjusted according to the local terrain complexity, and cross-validation is performed on the adjusted basic weights to obtain the fusion weights.

8. The intelligent detection method for river topographic map accuracy based on multi-image fusion according to claim 1, characterized in that, Step S3 extracts the river channel topographic features from the fused image, including: The lidar point cloud data in the fused image is rasterized to obtain a digital elevation model of the river area to be detected, and the elevation values ​​of each area in the digital elevation model are extracted as elevation feature parameters. Based on the variation amplitude of elevation feature parameters of adjacent grid areas, identify elevation change areas and elevation distribution patterns in fused images, and determine the corresponding river structure areas based on the elevation distribution patterns. Based on the corresponding pixel coordinates of the elevation change region and the river channel structure region, semantic segmentation is performed on the corresponding pixel regions in the fused image to extract the river channel structure feature points and the contour information of the elevation change region. The elevation feature parameters, river channel structural feature points, and contour information of elevation change areas are integrated according to their spatial location to obtain the river channel topographic features.

9. The intelligent detection method for river topographic map accuracy based on multi-image fusion according to claim 1, characterized in that, The method for obtaining the cartographic topographic features in the topographic map of the river channel to be detected in step S3 includes: The topographic map of the river channel to be detected is vectorized, and the corresponding map topographic features are extracted from the vectorization results based on the feature type of the river channel topographic features.

10. The intelligent detection method for river topographic map accuracy based on multi-source image fusion according to claim 1, characterized in that, Step S3, which involves performing precision matching verification, includes: Match the coordinates of river channel topographic features with the coordinates of river channel structural feature points in the map topographic features, and calculate the plane coordinate deviation of the corresponding feature points; Calculate the degree of agreement between the river channel topographic features and the outline information of the corresponding elevation abrupt change areas in the topographic features of the map; Extract the elevation values ​​of the corresponding areas from the river channel topographic features and the map topographic features, and calculate the absolute and relative elevation errors; The deviation of plane coordinates, the degree of matching of contour information, and the absolute and relative errors of elevation are used as the results of accuracy matching verification.

Images