Abandoned mine spot monitoring system based on remote sensing image data

Abstract

The application discloses a waste mine map spot monitoring system based on remote sensing image data and relates to the technical field of remote sensing image recognition. A patch extraction module constructs a repair semantic domain based on multi-period remote sensing images, calculates pixel exposure enhancement to generate a new exposed carrier patch, a skeleton tracking module determines the convergence direction in combination with terrain data, performs minimum cost path search along the slope surface to generate a rain erosion gully original skeleton, a cause construction module searches the transverse cross-section ridge and valley position based on the original skeleton to generate a rain erosion gully cause explanation domain, and a semantic stripping module performs a spatial difference set Boolean operation with the new exposed carrier patch as the subtracted polygon and the cause explanation domain as the subtracted polygon to generate a damage recurrence candidate domain and output a map spot level change state. The application accurately strips the natural rain erosion disturbance of the soil repair area, solves the problem that the soil repair area is misjudged as damage recurrence, and greatly improves the accuracy of map spot state recognition.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image recognition technology, and more specifically to a monitoring system for abandoned mine patches based on remote sensing image data. Background Technology

[0002] Ecological restoration monitoring of abandoned mines requires continuous dynamic monitoring of the evolution of mine patches using multi-phase high-resolution remote sensing imagery and topographic data. In practice, monitoring mainly relies on direct comparison of images from different time periods. By analyzing changes such as reduced vegetation cover, expansion of exposed surfaces, and expansion of pits and waste rock piles, the system infers whether the target area is in a state of new damage, expansion of damage, or restoration. However, mine ecological restoration projects inevitably generate short-term surface disturbances during construction and maintenance. Typical procedures such as slope trimming, topsoil covering and revegetation, site leveling, and construction of drainage facilities will leave newly exposed patches on the images that are bright in color, rough in texture, and have irregular boundaries, making the underlying image environment extremely complex and highly dynamic.

[0003] In the context of complex ecological restoration, existing remote sensing change detection technologies have revealed serious limitations in their underlying principles. Conventional methods often employ pixel comparison models, heavily relying on bare area thresholds, spectral index differences, or direct image subtraction algorithms for judgment. Once newly exposed patches are detected, they tend to be identified as newly added mining damage. This judgment logic completely ignores the naturally generated topographic and hydrological erosion mechanisms in the soil-covered restoration area. In areas under restoration and maintenance, the structure of the new soil layer is generally loose, and vegetation has not formed a stable soil-fixing cover. Rainfall runoff is easily eroded along gravity slopes, evolving into elongated grooves, dendritic grooves, or linear erosion textures extending downslope within the exposed areas. The bright spots and irregular extension patterns left by natural rain erosion have a very high visual similarity to newly excavated land or waste rock dumping caused by illegal mining. Because existing algorithms lack the ability to track the geometric topology of spatial erosion skeletons and the ability to interpret the causes of geological formation, they cannot accurately separate natural rain erosion disturbances from the overall increase in exposed area. This easily leads to the misjudgment of normal soil flow as mining damage recurrence, causing serious false alarms and significantly increasing the cost of field verification. Therefore, the core technical problem that urgently needs to be solved in this field is: how to overcome the limitations of judging based on single area or spectral changes in complex disturbance backgrounds, track and accurately separate the areal repair disturbance areas that can be reasonably explained by the rain erosion gully erosion mechanism from the extracted newly exposed patches, and then separate and output the remaining independent exposed areas that cannot be traced back to natural causes as real candidates for damage recurrence, thus completely solving the problem of misjudgment of patch status caused by surface visual confusion. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a monitoring system for abandoned mine patches based on remote sensing image data, which solves the problem that the lack of a mechanism for explaining the causes of mine damage leads to the misjudgment of rain erosion repair disturbances as damage recurrence.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a monitoring system for abandoned mine patches based on remote sensing image data, comprising: The patch extraction module is used to acquire multi-period remote sensing images and background patches of abandoned mines. Based on the image vector registration error and pixel spatial resolution, a restoration semantic domain is constructed. Based on the normalized rank value, the amount of exposure enhancement of pixels in the background patches of abandoned mines is calculated to generate newly exposed carrier patches. The skeleton tracking module is used to combine terrain auxiliary data to determine the slope confluence direction of pixels in newly exposed carrier patches, calculate the local gully appearance value to extract tracking seed points, and perform minimum cost path search along the slope confluence direction based on comprehensive cost to generate the original skeleton of rain erosion gullies. The gene formation module is used to establish cross-sectional search ridge locations based on the original skeleton of rain erosion gullies, obtain the local interpretation width of rain erosion based on the ridge distance, generate a single-point local influence area, merge the single-point local influence areas to obtain the preliminary interpretation domain, and perform a spatial intersection Boolean operation between the preliminary interpretation domain and the newly exposed carrier patch to generate the rain erosion gully gene formation interpretation domain. The semantic stripping module is used to perform spatial difference Boolean operations with the geometric boundary of newly exposed carrier patches as the minuend polygon and the geometric boundary of the rain erosion gully origin interpretation domain as the subtrahend polygon. After connectivity extraction, it generates a candidate domain for damage recurrence and outputs the patch-level change status according to the spatial object combination type.

[0006] Preferably, the patch extraction module constructs a restoration semantic domain based on image vector registration error and pixel spatial resolution, including: Retrieve the historical status fields of the base map patches of abandoned mines; For abandoned mine base map patches that are under repair or have been repaired and maintained, the image vector registration error and pixel spatial resolution values ​​are uniformly converted into values ​​with the same spatial distance dimension and then superimposed to obtain the extended distance. Based on the extended distance, a buffer polygon is generated in the spatial coordinate system to enclose the boundary of the original abandoned mine patch. The image area covered by the buffer polygon is established as the restoration semantic domain.

[0007] Preferably, the patch extraction module calculates the pixel exposure enhancement based on the normalized rank value to generate newly exposed carrier patches, including: Visible and near-infrared band features of the current and previous remote sensing images are extracted from multiple remote sensing images to obtain pixel brightness values, bare soil tone characterization values ​​and vegetation reduction values. In the restoration semantic domain, the pixel brightness values, bare soil tone characterization values ​​and vegetation reduction values ​​are sorted in ascending order to obtain normalized rank values. The arithmetic mean of the normalized rank values ​​is set as the bare soil characterization ranking value. Based on the difference operation between the current period's naked character ranking value and the previous period's naked character ranking value, the set of positive value pixels is extracted to obtain the positive value difference. The positive value difference is sorted in ascending order to form a one-dimensional sequence and the difference between adjacent values ​​is calculated. The sequence breakpoint where the largest adjacent difference is located is found, and the high value group pixels located after the sequence breakpoint are marked as significantly naked enhancement pixels. Eight-neighbor connectivity extraction is performed on significantly exposed enhanced pixels. Connected domains with an area less than or equal to a preset area filtering lower limit are removed, and independent connected objects are retained and output as newly exposed carrier patches.

[0008] Preferably, the skeleton tracking module combines terrain-aided data to determine the slope convergence direction of pixels within newly exposed carrier patches, including: A three-level spatial scale progressive regression mechanism was established. Based on topographic auxiliary data, the direction of the fastest elevation drop of pixels in newly exposed carrier patches was calculated. The direction of the fastest elevation drop was normalized into a unit vector and set as the slope confluence direction. When the calculated elevation gradient value is less than the preset lower limit of the system gradient, the average direction vector of all effective slope aspect data in the neighboring area of ​​the pixel is extracted as the slope confluence direction. When the elevation descent direction cannot be obtained pixel by pixel within the newly exposed carrier patch, the highest and lowest elevation points on the boundary of the newly exposed carrier patch are extracted. The vector direction pointing from the highest elevation point to the lowest elevation point is set as the patch-level descent direction. The patch-level descent direction is then uniformly set as the slope convergence direction of all pixels within the newly exposed carrier patch.

[0009] Preferably, the skeleton tracking module calculates the local trench appearance value to extract tracking seed points, including: For pixels within newly exposed carrier patches, a cross-section is constructed by sampling sets along a preset direction. The brightness difference value is obtained by comparing the brightness of the pixel center with the brightness of the pixel at the distance between individual pixels extending to both sides of the cross-section. The largest brightness difference value is selected. The median absolute deviation value obtained based on the pixel brightness value is added to the preset numerical stability term to construct the normalized denominator. The local groove appearance value is obtained by dividing the maximum brightness difference value by the normalized denominator. Pixels whose local gully apparent value is greater than or equal to the local gully apparent value of other pixels in the eight-neighborhood and have effective downhill adjacent pixels along the slope confluence direction are selected and merged to generate a set of tracking seed points.

[0010] Preferably, the skeleton tracking module performs a minimum cost path search along the slope confluence direction based on the comprehensive cost, including: For the current skeleton point in the tracking seed point set, candidate pixels are selected from the pixels in the eight neighborhoods of the current skeleton point that are located within the boundary of the newly exposed carrier patch. The local groove cost is obtained by subtracting the local groove appearance value of candidate pixels from the maximum value of the local groove appearance value of all pixels in the newly exposed carrier patch. Calculate the inner product between the unit direction vector from the current skeleton point to the candidate cell and the unit vector of the slope confluence direction at the current skeleton point. Multiply the difference between the numerical value and the inner product by the preset slope confluence direction constraint weight to obtain the direction deviation cost. The local trench cost and the direction deviation cost are added together to obtain the comprehensive cost. The candidate pixel with the smallest comprehensive cost value is selected as the next tracking point. The process iterates to select the next tracking point until the next tracking point exceeds the boundary of the newly exposed carrier patch, forms a closed loop, or there are no downhill candidate pixels. The paths are then merged to generate the original skeleton of rain erosion gullies.

[0011] Preferably, the genetic construction module establishes cross-sectional search for gully ridge locations based on the original skeleton of rain-eroded gullies, including: Identify the topological bifurcation nodes within the original skeleton of rain erosion grooves, and generate simple groove segments by segmenting the original skeleton of rain erosion grooves based on the topological bifurcation nodes. For skeleton points on simple groove segments, the direction from the previous adjacent skeleton point to the next adjacent skeleton point is extracted as the local tangent, and the direction perpendicular to the local tangent is taken as the local normal. Starting from the skeleton point, the sampling coordinates are gradually determined along the local normal to both sides at a fixed step size of less than or equal to half the pixel spatial resolution. The brightness values ​​at the sampling coordinates are read based on the interpolation algorithm until the sampling coordinates reach the preset maximum search distance threshold or leave the boundary of the newly exposed carrier patch, and a cross-sectional brightness sequence is generated. Median smoothing is performed on the cross-sectional brightness sequence. The first local high point in the cross-sectional brightness sequence that satisfies the condition that the brightness value is not lower than the inner adjacent sampling point, not lower than the outer adjacent sampling point, and greater than the brightness value of the center of the skeleton point is set as the ridge position. If no local high point is found, the sampling point with the largest value in the cross-sectional brightness sequence is selected as the ridge position.

[0012] Preferably, the causal construction module generates a single-point local influence area based on the gully-ridge distance to obtain the local interpretation width of rain erosion, including: The straight-line distance between the left ridge position and the skeleton point is calculated as the left ridge distance, and the straight-line distance between the right ridge position and the skeleton point is calculated as the right ridge distance. The average of the left ridge distance and the right ridge distance is calculated to obtain the local interpretation width of rain erosion. Using the skeleton point as the center point, a spatial cross axis is constructed with local tangent and local normal to generate a planar polygon object. The width of the planar polygon object along the local normal is set to be equal to the local interpretation width of the rain erosion, and the total length of the planar polygon object along the local tangent is set to be equal to one pixel spatial resolution. The planar polygon object is then set as a single-point local influence area.

[0013] Preferably, the semantic stripping module performs a spatial difference Boolean operation using the geometric boundary of the newly exposed carrier patch as the minuend polygon and the geometric boundary of the rain erosion groove origin interpretation domain as the subtrahend polygon, and generates a damage recurrence candidate domain through connectivity extraction, including: Determine if the domain for interpreting the formation of rain erosion grooves is empty. If it is empty, directly set the newly exposed carrier patch as the preliminary remaining exposed area. If the domain for interpreting the formation of rain erosion gullies is not empty, perform a spatial difference Boolean operation with the geometric boundary of the newly exposed carrier patch as the minuend polygon and the geometric boundary of the domain for interpreting the formation of rain erosion gullies as the subtrahend polygon to obtain the preliminary remaining exposed area. Eight-neighbor connectivity extraction is performed on the remaining exposed areas. Connected regions with areas smaller than or equal to the preset area filtering lower limit are removed, and the remaining independent connected objects are set as candidate regions for damage recurrence.

[0014] Preferably, the semantic stripping module outputs the patch-level change status based on the spatial object combination type, including: Generate patch-level change status based on the spatial object combination type within the same abandoned mine patch; When the newly exposed carrier patch corresponding to the abandoned mine patch is empty, the output shows no significant abnormal changes. When the newly exposed carrier patch corresponding to the abandoned mine patch is not empty and the corresponding candidate domain for damage recurrence is empty, output the status of repair disturbance pending verification. When there is a non-empty damage recurrence candidate domain in the abandoned mine patch and there is no rain erosion gully formation explanation domain in the abandoned mine patch, output the suspected damage recurrence status. When both the rain erosion groove formation interpretation domain and the non-empty damage recurrence candidate domain exist within the same abandoned mine patch, output a composite change pending verification status. Newly exposed carrier patches, the original skeleton of rain erosion grooves, the explanatory domain of rain erosion groove formation, the candidate domain of damage recurrence, and the patch-level change status are written into a pre-set spatial database.

[0015] Compared with existing technologies, it has the following advantages: This proposed system for monitoring abandoned mine patches based on remote sensing image data overcomes the limitations of conventional remote sensing change detection, which relies solely on the increase in exposed area or a single spectral evolution. It constructs a source-tracing and analytical mechanism that integrates surface spatial topology and grayscale profile morphology. For newly exposed patches inevitably generated during mine ecological restoration and management, this system introduces topographic data to establish surface water runoff paths. It then searches for and traces the starting points by combining apparent differences in brightness across local cross-sections, and performs a comprehensive cost path search along the water runoff direction to generate a fine gully framework. This mechanism eliminates the reliance of underlying monitoring models on simple pixel threshold comparisons, establishing a mathematical mapping between isolated changes in remote sensing pixel brightness and actual natural hydrological erosion trajectories on the surface. This provides highly reliable data support for accurately distinguishing between natural rain erosion and human-caused damage from illegal mining.

[0016] Based on the extracted skeleton tracking results, this scheme establishes an interpretation domain expansion model based on ridge span measurement. It clearly defines the spatial expansion boundaries of unilateral width and tangential length, and uses the initially extracted exposed geometric boundaries as the minuend polygons and the derived interpretation domain boundaries as the subtrahend polygons to perform rigorous spatial difference operations. This accurately isolates areal disturbance areas with reasonable natural hydrological causes from the original exposed contours. The independent connected objects retained after area lower limit filtering completely eliminate visual misleading components caused by topsoil erosion and peeling, transforming them into highly suspected real candidate targets of human-caused damage recurrence. The linked operation logic of spatial difference Boolean calculation and combination type state determination fills the gap in existing monitoring algorithms lacking a causal deconstruction level. From the underlying calculation approach, it cuts off the path of misreporting normal ecological disturbances as damage recurrence, significantly reducing the misjudgment rate and field verification costs in mine supervision. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system framework of the present invention.

[0018] Figure 2 This is a schematic diagram of the system execution flow of the present invention. Detailed Implementation

[0019] 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.

[0020] Please see Figures 1 to 2This application provides a monitoring system for abandoned mine patches based on remote sensing image data, including a patch extraction module, a skeleton tracking module, a gene formation module, and a semantic stripping module; It should be understood that the abandoned mine patch monitoring system based on remote sensing image data described in this embodiment can be deployed in a server cluster or workstation with high-performance computing capabilities. The underlying hardware of the system includes at least a processor and a memory communicatively connected to the processor. The memory stores computer program instructions for executing the patch monitoring processing logic described herein, and the processor calls these instructions and instantiates them into the patch extraction module, skeleton tracking module, causal construction module, and semantic stripping module described below. The system also includes a data interaction interface for receiving multi-period remote sensing images and abandoned mine background patch vector data from external sources, and outputting patch-level monitoring results to terminal devices. Based on the above hardware foundation, the processing of this system is achieved collaboratively by the following functional modules: The patch extraction module is used to acquire basic data input and construct the restoration semantic domain. The system uses its built-in raster data parsing engine to read the current and previous remote sensing images and obtain their corresponding spatial reference coordinate systems. The two images have the same or can be resampled to the same pixel spatial resolution. The system simultaneously reads in the abandoned mine base patch set, including each abandoned mine patch and its corresponding historical status field and image vector registration error. Based on the historical status field, the system determines whether each abandoned mine patch is under restoration or has been restored and is under maintenance. For abandoned mine patches that meet these conditions, the system expands outwards from their spatial polygon boundaries to construct the restoration semantic domain for the corresponding patch. For abandoned mine patches that do not meet these conditions, the subsequent analysis process is terminated.

[0021] Specifically, the outward extension distance of the restored semantic domain is set as the sum of the image vector registration error and the pixel spatial resolution, both of which are uniformly converted into spatial distance values ​​in meters and then superimposed. Based on the extension distance calculated above, the system generates a buffer polygon in the spatial coordinate system that encloses the boundary of the original abandoned mine patch. The image area covered by this buffer polygon is established as the restored semantic domain.

[0022] It should be noted that the abandoned mine baseline patch is a pre-defined spatial vector polygon object within the system, and the historical status field records the operational management stage of the patch. Pixel spatial resolution represents the smallest ground size that the image sensor can recognize, and image vector registration error (i.e., the registration error referred to in this invention) is the absolute value of the registration deviation between the current image and the vector patch in the planar coordinate system. The aforementioned distance extension mechanism superimposes the maximum spatial misalignment generated by system registration with the theoretical minimum imaging unit to form a buffer polygon. This mechanism encompasses the disturbance caused by repair construction at the patch edge and the deformation of temporary work passages, and simultaneously establishes the effective statistical population range for subsequent feature extraction and calculation through the numerical limitation of the aforementioned extended distance.

[0023] Within the defined restoration semantic domain space, the patch extraction module calculates the exposure characterization ranking value of each pixel in the current and previous remote sensing images. The system extracts visible light and near-infrared band features from the two images, obtains pixel brightness values, bare soil tone characterization values, and vegetation reduction values ​​calculated based on the normalized vegetation index, and performs rank normalization calculation based on the restoration semantic domain space.

[0024] Specifically, the system sorts the three types of feature values ​​of all valid pixels within the repair semantic domain in ascending order. For any pixel, its normalized rank value on a single feature is equal to its sorting index divided by the total number of pixels in the repair semantic domain. If multiple valid pixels within the repair semantic domain have the same single feature value, the arithmetic mean of the sorting indices of these multiple valid pixels is taken as their sorting index. The system calculates the arithmetic mean of the normalized rank values ​​of the pixel on the three types of features in the current remote sensing image, sets this arithmetic mean as the current exposure characterization ranking value, and similarly calculates the exposure characterization ranking value of the previous pixel based on the corresponding features of the previous remote sensing image. After the above processing, the range of exposure characterization ranking values ​​for pixels in both periods is mapped to a numerical range of zero to one.

[0025] It should be noted that the pixel brightness value is obtained by calculating the arithmetic mean of the reflectance of the red, green, and blue light bands; the vegetation reduction value is calculated by subtracting the normalized vegetation index from a value; and the bare soil tone characterization value is obtained by calculating the ratio of the sum of the reflectance of the red and green light bands to the sum of the reflectance of the near-infrared and blue light bands. This rank normalization mechanism transforms the absolute radiance values ​​based on the time series into a relative ordering probability distribution within the same restoration semantic domain. Because the vegetation reduction value is inverted, the closer the value is to one, the higher the rank representing the pixel as exhibiting bare soil characteristics in the local space. This calculation method ensures that the output results of the relative bare soil order remain stable even under conditions of overall color grading or illumination deviation between two image periods, establishing a unified feature comparison benchmark.

[0026] Based on the above sorted distribution data, the patch extraction module further calculates the pixel exposure enhancement amount of each pixel within the abandoned mine patch, and generates a set of newly exposed carrier patches after filtering. The formula for calculating the pixel exposure enhancement amount is as follows: In the formula, It represents the amount of pixel exposure enhancement, used to characterize the difference in the relative feature changes of a pixel with spatial coordinate x between two monitoring periods. This represents the current period's naked characterization ranking value. This represents the previous period's bare surface characterization ranking value. The formula quantifies the changing trend of surface bareness characteristics through a difference operation; a difference greater than zero indicates that the coordinate point possesses enhanced surface bareness characteristics.

[0027] Specifically, the system extracts a set of positive pixels with greater than zero exposure enhancement from the repair semantic domain and arranges the positive differences in ascending order to form a one-dimensional sequence. The system traverses this one-dimensional sequence, calculates the numerical difference between any two adjacent values, and finds the sequence breakpoint where the largest adjacent difference is located. When multiple largest adjacent differences exist with equal values, the breakpoint located at the end of the sequence is selected as the sequence breakpoint. The system marks high-value pixels following this breakpoint as significantly exposed enhanced pixels. The system performs eight-neighbor connectivity extraction on these significantly exposed enhanced pixels and determines the area of ​​each connected region, removing connected regions with an area less than or equal to four times the square of the pixel spatial resolution. The independent connected objects retained within the abandoned mine patch are then output as newly exposed carrier patches (also collectively referred to as the set of newly exposed carrier patches in the macro data stream).

[0028] It should be noted that the maximum adjacent difference is used as a mathematical breakpoint to divide pixels based on the numerical distribution of the difference data. The execution rule for eight-neighbor connectivity extraction is to group adjacent pixels in the horizontal, vertical, and diagonal directions into the same connected primitive. The parameter four times the square of the pixel spatial resolution sets the lower limit benchmark for area filtering, used to eliminate small connected regions caused by sensor single-pixel noise or registration jaggedness. If the maximum adjacent difference of the detected one-dimensional sequence is less than the set system minimum threshold, the system determines that the current abandoned mine patch has no significant exposed enhancement and outputs an empty carrier patch set. The system minimum threshold can be preset as an empirical constant such as 0.05, or adaptively set by three times the mean square error of exposed enhancement of pixels in historical unchanging areas. The newly generated exposed carrier patch set output by the above processing carries the determined spatial coordinate boundaries and image features, and is seamlessly input into subsequent modules for tracking calculations as an effective data carrier and spatial constraint condition.

[0029] The skeleton tracking module is used to extract linear skeleton data exhibiting rain-eroded morphology within newly exposed carrier patches. The system receives the set of newly exposed carrier patches output by the patch extraction module and simultaneously loads red, green, and blue band data from the current remote sensing image, as well as topographic auxiliary data. After normalizing the reflectance values ​​of the red, green, and blue bands, the system calculates the arithmetic mean of the reflectance values ​​of the three bands as the pixel brightness value and generates a brightness image. Simultaneously, it combines the topographic auxiliary data to determine the slope convergence direction of each pixel within the newly exposed carrier patch.

[0030] Specifically, the system calculates the direction of the fastest elevation descent for pixels within newly exposed carrier patches based on terrain-aided data, normalizes this direction into a unit vector, and sets it as the slope confluence direction for that pixel. If the calculated elevation gradient value is less than the set lower limit of the system gradient, the system extracts the average direction vector of all valid slope aspect data in the neighboring area of ​​that pixel as the slope confluence direction for that pixel. If the elevation descent direction cannot be obtained pixel by pixel within the newly exposed carrier patch, the system extracts the highest and lowest elevation points on the boundary of the newly exposed carrier patch, sets the vector direction from the highest point to the lowest point as a unified patch-level descent direction, and uniformly sets the patch-level descent direction as the slope confluence direction for all pixels within the newly exposed carrier patch.

[0031] It should be noted that the brightness image is a grayscale distribution matrix obtained based on the current visible light band conversion. The terrain auxiliary data is a spatial matrix containing elevation or slope information. The slope runoff direction indicates the surface water runoff path dominated by gravity when rainfall flows on the surface. The set lower limit of the system gradient is a preset slope limit, which can be set to a range of 2 to 5 degrees. The above processing logic establishes a three-level spatial scale progressive backtracking mechanism of pixel-by-pixel elevation gradient, local neighborhood average direction, and patch-level unified direction. This mechanism ensures that even in the absence of high-precision terrain data or in areas with flat terrain, the system can still output direction vector data with a clear gravity runoff direction, preventing subsequent tracking calculations from being interrupted due to missing direction data.

[0032] After determining the slope confluence direction, the skeleton tracking module calculates the local groove appearance value of each pixel in the newly exposed carrier patch, and selects and generates a set of tracking seed points from the local groove appearance value.

[0033] Specifically, based on a preset directional sampling set, the system constructs cross-sections along each direction in the directional sampling set for any pixel within a newly exposed carrier patch. It compares the brightness of the pixel's center with the brightness of pixels extending to the sides of the cross-section at individual pixel intervals to obtain brightness difference values ​​in each direction. The system selects the largest brightness difference value, adds the median absolute deviation value to a set numerical stability term to construct a normalized denominator, and divides the largest brightness difference value by the normalized denominator. The result is set as the local groove appearance value of that pixel. Within the newly exposed carrier patch, the system filters all pixels whose local groove appearance value is greater than or equal to the local groove appearance values ​​of other pixels in its eight-neighborhood, and which have effective downhill adjacent pixels along the slope confluence direction. Pixels meeting these conditions are merged and set as the tracking seed point set.

[0034] It should be noted that the direction sampling set is a discretized sequence of cross-sectional tangential angles, containing eight directions with equal intervals from 0 degrees to 157.5 degrees. The median absolute deviation is the median of the absolute values ​​of the differences between the brightness values ​​of each pixel within the newly exposed carrier patch and the median brightness of the patch, and is a standard statistical parameter. The numerical stability term is a preset minimum positive number used to ensure the continuity of feature response calculation. The local groove appearance value quantifies the mathematical probability that a pixel is located at the bottom of the groove center; a larger value indicates that the pixel has more brightly lit ridge protrusions on both sides. The above processing rule for generating tracking seed points based on local relative values ​​uses the extreme value attribute of a pixel in its neighborhood to establish the tracking starting point, eliminating the dependence on fixed grayscale constants based on human experience.

[0035] Based on the set of tracking seed points, the skeleton tracking module performs a minimum cost path search calculation along the slope confluence direction to generate and output the original skeleton of rain-eroded gullies.

[0036] Specifically, for any current skeleton point in the tracking seed point set, the system selects the next tracking point from candidate pixels within its eight neighborhoods that are located within the boundary of newly exposed carrier patches. The system obtains the local gully cost for each candidate pixel and calculates the inner product of the unit direction vector from the current skeleton point to the candidate pixel and the unit vector of the slope confluence direction at the current skeleton point. The system subtracts the difference of the inner product from the value and multiplies it by a set slope confluence direction constraint weight to obtain the direction deviation cost. The system adds the local gully cost and the direction deviation cost to obtain the comprehensive cost, and selects the candidate pixel with the smallest comprehensive cost as the next tracking point. The system iterates the above next tracking point acquisition logic until the next tracking point exceeds the boundary of the newly exposed carrier patch, forms a closed loop, or there are no downhill candidate pixels, at which point the search for the current path is terminated. The system merges multiple paths with shared pixels or end-to-end connections, outputting the original skeleton of rain-eroded gullies.

[0037] It should be noted that the local gully cost of a candidate pixel is obtained by subtracting the local gully apparent value of the candidate pixel from the maximum local gully apparent value of all pixels within the newly exposed carrier patch. The larger the local gully apparent value, the smaller the converted local gully cost. The slope runoff direction constraint weight is a dimensionless value that adjusts the ratio of apparent features to topographic priors, and its value can be set between 0 and 1, preferably 1. The vector inner product operation mechanism evaluates the spatial fit between the tracking direction and the slope runoff direction; the closer the inner product value is to 1, the lower the direction deviation cost. The above search mechanism transforms the causes of rain erosion into a computer-executed step of minimum-cost path planning, ensuring that the generated skeleton data strictly follows the gully texture direction formed by rainfall runoff erosion. If the merged skeleton data contains a branch structure, the system stores it as a directed graph data object with connectivity. If a valid path cannot be generated within a newly exposed carrier patch, the system assigns a check mark to the newly exposed carrier patch indicating that fine groove features cannot be stably identified and outputs it to ensure the integrity of the entire module's data flow and storage logic.

[0038] The generative construction module is used to generate regional objects with actual spatial coverage based on the original skeleton of rain-eroded grooves. The system receives the original skeleton of rain-eroded grooves and newly exposed carrier patches output by the skeleton tracking module, and reads the current brightness image and pixel spatial resolution. The system performs line segmentation on the original skeleton of rain-eroded grooves according to the bifurcation nodes to generate independent simple groove segments.

[0039] Specifically, the system determines whether there are bifurcation nodes within the original skeleton of the rain erosion gullies. If there are no bifurcation nodes, the system extracts the entire original skeleton of the rain erosion gullies into a single simple groove segment; if there are bifurcation nodes, the system uses the bifurcation nodes and the skeleton endpoints as the dividing line to split the original skeleton of the rain erosion gullies into multiple simple groove segments that do not contain bifurcation nodes.

[0040] It should be noted that a branch node is a topological intersection connecting three or more skeleton edges. The above splitting operation reduces the dimensionality of the tree or network structure to linear simple groove segments, eliminating the risk of computational chaos at intersections and providing an unambiguous geometric basis for subsequent point-by-point unified cross-sectional search analysis.

[0041] After generating simple groove segments, the generative construction module establishes spatial cross-sections point by point and searches for the groove and ridge boundary positions on the cross-sections. The system calculates the local normal based on the coordinates of adjacent skeleton points, reads the grayscale sequence of the current brightness image along the local normal, and determines the local brightness high points as groove and ridge positions by combining the set maximum search distance threshold.

[0042] Specifically, for any skeleton point on a simple groove segment, the system extracts the direction from the previous adjacent skeleton point to the next adjacent skeleton point as the local tangent. If the current skeleton point is an endpoint, the system extracts the direction of the line connecting that endpoint and its only adjacent skeleton point as the local tangent. The system uses the direction perpendicular to the local tangent as the local normal and establishes cross-sections on both sides of the skeleton point along this local normal. Starting from the skeleton point, to prevent pixel skipping and aliasing during oblique sampling of the discrete raster, the system gradually determines the sampling coordinates to both sides with a fixed step size less than or equal to half the pixel spatial resolution. It then reads the brightness values ​​at the corresponding sampling coordinates based on bilinear interpolation or nearest neighbor interpolation algorithms until the sampling coordinates reach the set maximum search distance threshold or leave the boundary of the newly exposed carrier patch, generating a cross-sectional brightness sequence. The system performs median smoothing on the cross-sectional brightness sequence with a length of three sampling points. In the smoothed cross-sectional brightness sequence, the system searches outwards from a position one pixel away from the skeleton point. The local high point that first satisfies the condition of having a brightness value no lower than its inner and outer adjacent sampling points, and a brightness value greater than the center brightness value of the skeleton point, is designated as the ridge position on that side. If no local high point satisfying the above conditions is found before reaching the maximum search distance threshold or leaving the boundary of the newly exposed carrier patch, the sampling point with the largest value in the corresponding side's cross-sectional brightness sequence is selected as the ridge position.

[0043] It should be noted that the local tangential direction reflects the instantaneous extension direction of the groove at the skeleton point, while the local normal is the cross-sectional direction perpendicular to the direction. A smoothing process with a length equal to the median of three sampling points is used to filter out noise caused by single-pixel brightness jumps due to sensor interference. The maximum search distance threshold sets the upper limit of the extreme value of the groove width on one side. This threshold can be empirically set based on the actual spatial width of typical rain-eroded grooves, for example, set to 10 times the pixel spatial resolution value, to prevent the search range from excessively spreading to irrelevant bright spots. The ridge position maps the bright raised structure at the edge of the groove formed by rainwater cutting on the ground surface. If the actual spatial length of a cross-section on a certain side is less than one pixel spacing, the system determines that there is no effective ridge on that side and skips the strip region generation logic for that skeleton point. The above processing establishes the groove boundary through dual constraints of the maximum brightness value and the center brightness, achieving adaptive measurement for different erosion widths while ensuring stable convergence of the algorithm.

[0044] Based on the left and right ridge locations obtained through the search, the causal construction module calculates the local interpretation width at each skeleton point and generates the causal interpretation domain for rain erosion gullies.

[0045] Specifically, the system calculates the straight-line distance between the left ridge position and the skeleton point as the left ridge distance, and calculates the straight-line distance between the right ridge position and the skeleton point as the right ridge distance. The system adds the left and right ridge distances and divides by 2 to obtain the local interpretation width of the rain erosion at that skeleton point. Then, for all valid skeleton points on the original skeleton of the rain erosion gullies, the system constructs a spatial cross axis with each skeleton point as the center point, using local tangents and local normals. This generates a polygonal object with a single-sided width along the local normal equal to the local interpretation width of the rain erosion, and a total length along the local tangent equal to one pixel spatial resolution, which is set as the single-point local influence area. The system spatially merges the single-point local influence areas corresponding to all skeleton points to obtain a preliminary interpretation domain. The preliminary interpretation domain is then spatially intersected with the newly exposed carrier patch using a Boolean operation to generate a rain erosion gully formation interpretation domain with strictly constrained boundaries. If no valid original skeleton of the rain erosion gullies is received within the newly exposed carrier patch, an empty rain erosion gully formation interpretation domain is output.

[0046] It should be noted that the distances between the left and right ridges are dynamically calculated spatial straight-line lengths, and their units are all uniformly set to meters. The single-point local influence area is a polygonal region that adaptively changes with the actual brightness distribution of the trough. The above-described method of averaging distances and generating strip-shaped regions expands the linear skeleton into a coverage area with planar topological features. By introducing spatial intersection Boolean operations, the interpretation range is strictly limited to the boundary of the newly exposed carrier patch, avoiding the generation of meaningless extended regions beyond the boundary of the exposed object, and providing a precise planar spatial basis for the subsequent difference stripping step.

[0047] The semantic stripping module is used to spatially separate newly exposed carrier patches based on the interpretation domain of rain erosion gullies, outputting monitoring results at the abandoned mine patch level. The system receives the interpretation domain of rain erosion gullies output by the generative construction module and simultaneously calls the set of newly exposed carrier patches and the set of abandoned mine background patches output by the patch extraction module. Based on the received spatial object data, the system performs spatial difference Boolean operations to generate candidate domains for damage recurrence, and combines this with preset state judgment logic to output the abandoned mine patch-level change status and verification markers.

[0048] Specifically, the system marks the received rain erosion groove formation interpretation domain as the rain erosion disturbance interpretation domain for repair. For each newly exposed carrier patch within the same abandoned mine map, the system determines whether its corresponding rain erosion groove formation interpretation domain is empty. If the corresponding rain erosion groove formation interpretation domain is empty, the system directly sets the newly exposed carrier patch as the preliminary remaining exposed area; if the corresponding rain erosion groove formation interpretation domain is not empty, the system performs a spatial difference Boolean operation using the geometric boundary of the newly exposed carrier patch as the minuend polygon and the geometric boundary of the corresponding rain erosion groove formation interpretation domain as the subtrahend polygon to obtain the preliminary remaining exposed area. The system performs eight-neighbor connectivity extraction on the preliminary remaining exposed area, determines the area of ​​each connected region, and removes connected regions with an area less than or equal to four times the square of the pixel spatial resolution. The system sets the independent connected objects retained after area removal as damage recurrence candidate domains. If no connected objects are retained after area removal, the system outputs an empty damage recurrence candidate domain.

[0049] Subsequently, the system generates a patch-level change status based on the spatial object combination type within the same abandoned mine patch: if the set of newly exposed carrier patches corresponding to the abandoned mine patch is empty, the system outputs a status of no significant abnormal change; if the set of newly exposed carrier patches corresponding to the abandoned mine patch is not empty, and all its corresponding damage recurrence candidate domains are empty, the system outputs a status of repair disturbance pending verification; if there are non-empty damage recurrence candidate domains within the abandoned mine patch, and there are no rain erosion gully gene interpretation domains within the entire patch, the system outputs a suspected damage recurrence status; if both rain erosion gully gene interpretation domains and non-empty damage recurrence candidate domains exist within the same abandoned mine patch, the system outputs a status of composite change pending verification. The system writes the newly exposed carrier patches, rain erosion gully primary skeleton, rain erosion gully gene interpretation domains, damage recurrence candidate domains, patch-level change status, and verification markers extracted and generated in the aforementioned steps into a preset spatial database.

[0050] It should be understood that the rain erosion disturbance interpretation domain characterizes the spatial range within newly exposed carrier patches that can be reasonably explained by the rain erosion mechanism after soil remediation. Spatial difference Boolean operations, through polygon subtraction, isolate perturbation regions with reasonable causes from the original change contours. The parameter four times the square of the pixel spatial resolution sets the lower limit benchmark for filtering residual patches, used to filter out the tiny topological fragments inevitably generated at the edges of spatial vector subtraction. The above-mentioned patch-level change state determination logic establishes an independent and completely exhaustive full-coverage state branch output mechanism, accurately decomposing confusing image change phenomena into two types of spatial objects with clear business orientations. The processing method of jointly storing patch-level change states and spatial vector data establishes a data link between the underlying image recognition objects and the upper-level ecological restoration dynamic monitoring database, providing precise spatial anchors for subsequent human-computer interaction review and field verification of questionable patches.

[0051] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A monitoring system for abandoned mine patches based on remote sensing image data, characterized in that, include: The patch extraction module is used to acquire multi-period remote sensing images and background patches of abandoned mines. Based on the image vector registration error and pixel spatial resolution, a restoration semantic domain is constructed. Based on the normalized rank value, the amount of exposure enhancement of pixels in the background patches of abandoned mines is calculated to generate newly exposed carrier patches. The skeleton tracking module is used to combine terrain auxiliary data to determine the slope confluence direction of pixels in newly exposed carrier patches, calculate the local gully appearance value to extract tracking seed points, and perform minimum cost path search along the slope confluence direction based on comprehensive cost to generate the original skeleton of rain erosion gullies. The gene formation module is used to establish cross-sectional search ridge locations based on the original skeleton of rain erosion gullies, obtain the local interpretation width of rain erosion based on the ridge distance, generate a single-point local influence area, merge the single-point local influence areas to obtain the preliminary interpretation domain, and perform a spatial intersection Boolean operation between the preliminary interpretation domain and the newly exposed carrier patch to generate the rain erosion gully gene formation interpretation domain. The semantic stripping module is used to perform spatial difference Boolean operations with the geometric boundary of newly exposed carrier patches as the minuend polygon and the geometric boundary of the rain erosion gully origin interpretation domain as the subtrahend polygon. After connectivity extraction, it generates a candidate domain for damage recurrence and outputs the patch-level change status according to the spatial object combination type.

2. The abandoned mine patch monitoring system based on remote sensing image data according to claim 1, characterized in that, The patch extraction module constructs a restoration semantic domain based on image vector registration error and pixel spatial resolution, including: Retrieve the historical status fields of the base map patches of abandoned mines; For abandoned mine base map patches that are under repair or have been repaired and maintained, the image vector registration error and pixel spatial resolution values ​​are uniformly converted into values ​​with the same spatial distance dimension and then superimposed to obtain the extended distance. Based on the extended distance, a buffer polygon is generated in the spatial coordinate system to enclose the boundary of the original abandoned mine patch. The image area covered by the buffer polygon is established as the restoration semantic domain.

3. The abandoned mine patch monitoring system based on remote sensing image data according to claim 2, characterized in that, The patch extraction module calculates the enhancement amount of exposed pixels based on normalized rank values, generating newly exposed carrier patches, including: Visible and near-infrared band features of the current and previous remote sensing images are extracted from multiple remote sensing images to obtain pixel brightness values, bare soil tone characterization values ​​and vegetation reduction values. In the restoration semantic domain, the pixel brightness values, bare soil tone characterization values ​​and vegetation reduction values ​​are sorted in ascending order to obtain normalized rank values. The arithmetic mean of the normalized rank values ​​is set as the bare soil characterization ranking value. Based on the difference operation between the current period's naked character ranking value and the previous period's naked character ranking value, the set of positive value pixels is extracted to obtain the positive value difference. The positive value difference is sorted in ascending order to form a one-dimensional sequence and the difference between adjacent values ​​is calculated. The sequence breakpoint where the largest adjacent difference is located is found, and the high value group pixels located after the sequence breakpoint are marked as significantly naked enhancement pixels. Eight-neighbor connectivity extraction is performed on significantly exposed enhanced pixels. Connected domains with an area less than or equal to a preset area filtering lower limit are removed, and independent connected objects are retained and output as newly exposed carrier patches.

4. The abandoned mine patch monitoring system based on remote sensing image data according to claim 1, characterized in that, The skeleton tracking module, combined with terrain-aided data, determines the slope convergence direction of pixels within newly exposed carrier patches, including: A three-level spatial scale progressive regression mechanism was established. Based on topographic auxiliary data, the direction of the fastest elevation drop of pixels in newly exposed carrier patches was calculated. The direction of the fastest elevation drop was normalized into a unit vector and set as the slope confluence direction. When the calculated elevation gradient value is less than the preset lower limit of the system gradient, the average direction vector of all effective slope aspect data in the neighboring area of ​​the pixel is extracted as the slope confluence direction. When the elevation descent direction cannot be obtained pixel by pixel within the newly exposed carrier patch, the highest and lowest elevation points on the boundary of the newly exposed carrier patch are extracted. The vector direction pointing from the highest elevation point to the lowest elevation point is set as the patch-level descent direction. The patch-level descent direction is then uniformly set as the slope convergence direction of all pixels within the newly exposed carrier patch.

5. The abandoned mine patch monitoring system based on remote sensing image data according to claim 4, characterized in that, The skeleton tracking module calculates the local trench appearance value and extracts tracking seed points, including: For pixels within newly exposed carrier patches, a cross-section is constructed by sampling sets along a preset direction. The brightness difference value is obtained by comparing the brightness of the pixel center with the brightness of the pixel at the distance between individual pixels extending to both sides of the cross-section. The largest brightness difference value is selected. The median absolute deviation value obtained based on the pixel brightness value is added to the preset numerical stability term to construct the normalized denominator. The local groove appearance value is obtained by dividing the maximum brightness difference value by the normalized denominator. Pixels whose local gully apparent value is greater than or equal to the local gully apparent value of other pixels in the eight-neighborhood and have effective downhill adjacent pixels along the slope confluence direction are selected and merged to generate a set of tracking seed points.

6. The abandoned mine patch monitoring system based on remote sensing image data according to claim 5, characterized in that, The skeleton tracking module performs a minimum cost path search along the slope confluence direction based on the overall cost, including: For the current skeleton point in the tracking seed point set, candidate pixels are selected from the pixels in the eight neighborhoods of the current skeleton point that are located within the boundary of the newly exposed carrier patch. The local groove cost is obtained by subtracting the local groove appearance value of candidate pixels from the maximum value of the local groove appearance value of all pixels in the newly exposed carrier patch. Calculate the inner product between the unit direction vector from the current skeleton point to the candidate cell and the unit vector of the slope confluence direction at the current skeleton point. Multiply the difference between the numerical value and the inner product by the preset slope confluence direction constraint weight to obtain the direction deviation cost. The local trench cost and the direction deviation cost are added together to obtain the comprehensive cost. The candidate pixel with the smallest comprehensive cost value is selected as the next tracking point. The process iterates to select the next tracking point until the next tracking point exceeds the boundary of the newly exposed carrier patch, forms a closed loop, or there are no downhill candidate pixels. The paths are then merged to generate the original skeleton of rain erosion gullies.

7. The abandoned mine patch monitoring system based on remote sensing image data according to claim 1, characterized in that, The generative construction module establishes cross-sectional searches for gully ridge locations based on the original framework of rain-eroded gullies, including: Identify the topological bifurcation nodes within the original skeleton of rain erosion grooves, and generate simple groove segments by segmenting the original skeleton of rain erosion grooves based on the topological bifurcation nodes. For skeleton points on simple groove segments, the direction from the previous adjacent skeleton point to the next adjacent skeleton point is extracted as the local tangent, and the direction perpendicular to the local tangent is taken as the local normal. Starting from the skeleton point, the sampling coordinates are gradually determined along the local normal to both sides at a fixed step size of less than or equal to half the pixel spatial resolution. The brightness values ​​at the sampling coordinates are read based on the interpolation algorithm until the sampling coordinates reach the preset maximum search distance threshold or leave the boundary of the newly exposed carrier patch, and a cross-sectional brightness sequence is generated. Median smoothing is performed on the cross-sectional brightness sequence. The first local high point in the cross-sectional brightness sequence that satisfies the condition that the brightness value is not lower than the inner adjacent sampling point, not lower than the outer adjacent sampling point, and greater than the brightness value of the center of the skeleton point is set as the ridge position. If no local high point is found, the sampling point with the largest value in the cross-sectional brightness sequence is selected as the ridge position.

8. The abandoned mine patch monitoring system based on remote sensing image data according to claim 7, characterized in that, The causal construction module generates a single-point local influence area based on the gully-ridge distance to obtain the local interpretation width of rain erosion, including: The straight-line distance between the left ridge position and the skeleton point is calculated as the left ridge distance, and the straight-line distance between the right ridge position and the skeleton point is calculated as the right ridge distance. The average of the left ridge distance and the right ridge distance is calculated to obtain the local interpretation width of rain erosion. Using the skeleton point as the center point, a spatial cross axis is constructed with local tangent and local normal to generate a planar polygon object. The width of the planar polygon object along the local normal is set to be equal to the local interpretation width of the rain erosion, and the total length of the planar polygon object along the local tangent is set to be equal to one pixel spatial resolution. The planar polygon object is then set as a single-point local influence area.

9. The abandoned mine patch monitoring system based on remote sensing image data according to claim 1, characterized in that, The semantic stripping module performs a spatial difference Boolean operation using the geometric boundaries of newly exposed carrier patches as the minuend polygon and the geometric boundaries of the rain erosion groove etiology interpretation domain as the subtrahend polygon. Through connectivity extraction, it generates candidate domains for damage recurrence, including: Determine if the domain for interpreting the formation of rain erosion grooves is empty. If it is empty, directly set the newly exposed carrier patch as the preliminary remaining exposed area. If the domain for interpreting the formation of rain erosion gullies is not empty, perform a spatial difference Boolean operation with the geometric boundary of the newly exposed carrier patch as the minuend polygon and the geometric boundary of the domain for interpreting the formation of rain erosion gullies as the subtrahend polygon to obtain the preliminary remaining exposed area. Eight-neighbor connectivity extraction is performed on the remaining exposed areas. Connected regions with areas smaller than or equal to the preset area filtering lower limit are removed, and the remaining independent connected objects are set as candidate regions for damage recurrence.

10. The abandoned mine patch monitoring system based on remote sensing image data according to claim 9, characterized in that, The semantic stripping module outputs the patch-level change status based on the spatial object combination type, including: Generate patch-level change status based on the spatial object combination type within the same abandoned mine patch; When the newly exposed carrier patch corresponding to the abandoned mine patch is empty, the output shows no significant abnormal changes. When the newly exposed carrier patch corresponding to the abandoned mine patch is not empty and the corresponding candidate domain for damage recurrence is empty, output the status of repair disturbance pending verification. When there is a non-empty damage recurrence candidate domain in the abandoned mine patch and there is no rain erosion gully formation explanation domain in the abandoned mine patch, output the suspected damage recurrence status. When both the rain erosion groove formation interpretation domain and the non-empty damage recurrence candidate domain exist within the same abandoned mine patch, output a composite change pending verification status. Newly exposed carrier patches, the original skeleton of rain erosion grooves, the explanatory domain of rain erosion groove formation, the candidate domain of damage recurrence, and the patch-level change status are written into a pre-set spatial database.

Images