Mine greening supervision method based on remote sensing image semantic segmentation

Abstract

The invention relates to the technical field of ecological environment data monitoring, and further relates to a mine greening supervision method based on remote sensing image semantic segmentation. The method comprises the following steps: step 1, based on original ecological indexes of a mine, establishing a six-column time sequence matrix for each mine landscape patch; 2, at each monitoring moment, constraint non-negative matrix factorization is completed on the time sequence matrix through minimum K-L divergence iteration, and the target is to disassemble the time sequence matrix into a product of a potential ecological function module weight matrix and a potential ecological function module index contribution matrix; and 3, carrying out weighted average on the resource-liability balance degree by using the area of each mine landscape patch so as to evaluate the ecological restoration effect of the mine. According to the invention, dynamic, fine and structured monitoring and evaluation of the mine ecological restoration process are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of ecological environment data monitoring, and specifically relates to a mine greening supervision method based on remote sensing image semantic segmentation. Background Art

[0002] With the continuous development of mineral resources, the problem of ecological damage in mining areas has become increasingly serious. Engineering activities such as open-pit mining, underground collapse, spoil dump accumulation, and road excavation have directly changed the surface morphology and soil structure, leading to multiple ecological degradation processes such as vegetation destruction, soil erosion, land degradation, and landscape fragmentation. To this end, since the end of the last century, my country has gradually established a mine ecological restoration system based on "mining and restoration" and "post-mining reclamation", which clearly stipulates that mining companies must fulfill their ecological restoration obligations during mining activities, and are supervised and inspected by the relevant government departments. However, at the practical level, the current mine ecological restoration work generally has the following key technical difficulties:

[0003] First, existing supervision and evaluation methods are often based on single indicators or limited factors, such as vegetation cover, soil organic matter, and heavy metal content, which fail to fully reflect the ecosystem response mechanisms to restoration effectiveness. For example, some supervision reports use improvements in NDVI as the basis for achieving restoration standards, neglecting the systematic evaluation of structural connectivity, geomorphic stability, and ecological risk liabilities, resulting in supervision data that deviates from actual ecological processes. Second, ecological restoration supervision generally employs qualitative grading or empirical weighting models, lacking a rigorous mathematical basis. Currently widely used methods such as the Analytic Hierarchy Process (AHP), comprehensive index method, and fuzzy comprehensive evaluation method rely on artificial weighting or subjective interpretation rules, resulting in poor model reproducibility and scale comparability. For example, the "Land Reclamation Grade Evaluation Standard" generates ecological grades through manual scoring and grading. While concise in form, this method struggles to accommodate the heterogeneity and dynamic evolution of ecological structures across land types. Third, temporal dynamics are insufficiently addressed. Most existing supervision methods only compare a certain point in time or two limited points in time, lack a process understanding based on time series data, and are unable to identify the implicit trends of "short-term improvement, long-term degradation" or "surface improvement, structural deterioration" in the process of ecological restoration. In particular, there is a serious lag in the risk identification of isolated patches caused by rapid vegetation growth but unrecovered soil, or decreased connectivity. Fourth, spatial structural information is not fully utilized. Current mainstream evaluation technologies focus on point-like indicator extraction and regional averaging, ignoring the core role of ecological pattern structure in maintaining ecological functions. For example, in mine slope control projects, although vegetation coverage increases, if landscape fragmentation increases and ecological network connectivity decreases, it may pose a greater threat to species migration and ecological stability. Existing evaluation models are difficult to reveal such spatial structural risks. Summary of the Invention

[0004] The main purpose of this invention is to provide a mine restoration supervision method based on semantic segmentation of remote sensing images. By constructing an ecological indicator matrix with bidirectional normalization of resources and liabilities, non-negative matrix decomposition is used to extract potential ecological functional modules, and the resource income and liability residual ratio at the patch level is calculated to form a quantifiable resource-liability balance. Then, combined with patch area, structural connectivity, and landscape fragmentation, a global supervision evaluation index is constructed to achieve dynamic, precise, and structured monitoring and evaluation of the mine ecological restoration process. This method does not require subjective weighting, has mathematical closure and data repeatability, and significantly improves the scientific nature, accuracy, and engineering practicality of supervision evaluation.

[0005] In order to solve the above problems, the technical solution of the present invention is achieved as follows:

[0006] A mine restoration supervision method based on remote sensing image semantic segmentation, the method comprising:

[0007] Step 1: Based on the original ecological indicators of the mine, including resource indicators and liability indicators, a six-column time series matrix is established for each mine landscape patch. The elements in the first three columns of the time series matrix are all resource indicators, forming the resource column, representing resource attributes, and the elements in the last three columns are all liability indicators, forming the liability column, representing liability attributes.

[0008] Step 2: At each monitoring moment, constrained non-negative matrix decomposition is performed on the time series matrix by iteratively minimizing the KL divergence. The goal is to decompose the time series matrix into the product of the potential ecological function module weight matrix and the potential ecological function module indicator contribution matrix, and to ensure that both the potential ecological function module weight matrix and the potential ecological function module indicator contribution matrix remain non-negative; for each potential ecological function module, its resource income and liability residual are calculated, and the ratio of resource income to liability residual is obtained as the resource-liability balance;

[0009] Step 3: Use the area of each mine landscape patch to perform a weighted average of the resource-liability balance as the main part of the restoration effect; at the same time, use the structural connectivity between each mine landscape patch as an amplification factor, and introduce the landscape fragmentation as a penalty term into the same expression to calculate the global supervision evaluation index to evaluate the effect of mine ecological restoration.

[0010] Furthermore, resource indicators include: soil organic matter density, vegetation cover and slope stability; the liability indicators include: comprehensive heavy metal pollution index, annual average potential soil and water loss and landscape fragmentation; in order to make the values of resource indicators always positive and the values of liability indicators always negative, and to eliminate dimensional differences, the full-time and space-time maximum value of each original ecological indicator is first obtained, and then each original ecological indicator is logarithmically compressed and normalized according to the maximum value.

[0011] Furthermore, in step 2, for each potential ecological function module, the process of calculating its resource benefits and liability residuals includes: adding the cumulative contributions of each potential ecological function module to the resource column to obtain the resource-side weight of the potential ecological function module; then adding the cumulative contributions of the same potential ecological function module to the liability column to obtain the liability-side weight; for each mining landscape patch, multiplying its weight on each potential ecological function module by the resource-side weight and summing them to obtain the resource benefits of the mining landscape patch; multiplying the indicator contribution on each potential ecological function module by the liability-side weight and summing them to obtain the liability residual. The ratio of resource benefits to liability residual is the resource-liability balance of the mining landscape patch. When it is greater than 1, it means that the restoration benefits have exceeded the liability residual; if it is less than 1, it means that the liability residual is still dominant and the governance efforts need to be strengthened.

[0012] Furthermore, at time t, each element M in the time series matrix M(t) of the mining landscape patch p p，χ (t) is:

[0013]

[0014] Where, sgn(x) = +1, when x∈{1, 2, 3}, represents the resource column; sgn(χ) = -1, when χ∈{4, 5, 6}, represents the liability column; x = 1, represents the soil organic matter density; χ = 2, represents the vegetation cover; χ = 3, represents the slope stability; χ = 4, represents the total heavy metal pollution index; χ = 5, represents the annual potential soil and water loss modulus; χ = 6, represents the landscape fragmentation; X p，χ (t) is the original ecological index of the mine landscape patch p in the time series matrix M(t) at time t, corresponding to the column χ; p′，k (t′) is the original ecological indicator of the corresponding column χ in the time series matrix M(t) of the mining landscape patch p′ at time t′.

[0015] Furthermore, let the number of mining landscape patches be P, then the number of potential ecological function modules is calculated by Perform floor operation to obtain the value.

[0016] Furthermore, assuming that the potential ecological function module weight matrix is P(t) and the potential ecological function module indicator contribution matrix is Q(t), the constrained non-negative matrix decomposition is performed on the time series matrix by minimizing the KL divergence iteration using the following formula:

[0017]

[0018] And the following equality exists:

[0019]

[0020] Among them, P p，k (t) represents the weight of mining landscape patch p on potential ecological function module k at time t; Q k，X (t) represents the indicator contribution of potential ecological function module k at time t.

[0021] Furthermore, when the constrained non-negative matrix decomposition is completed by iteratively minimizing the KL divergence of the time series matrix, P p，k The update formula of (t) is:

[0022]

[0023] Among them, P p，k，new (t) is the updated P p，k (t); P p，k′ (t) represents the weight of mining landscape patch p on potential ecological function module k′ at time t; Q k′，χ (t) represents the indicator contribution of the potential ecological function module k′ at time t; Q k，χ The update formula of (t) is:

[0024]

[0025] Among them, Q k，χ，new (t) is the updated Q k，χ (t).

[0026] Furthermore, at time t, the resource income R of the potential ecological function module k is k (t) is:

[0027] R k (t)=∑ χ∈{1，2，3} Q k，χ (t);

[0028] At time t, the residual liability L of the potential ecological function module k is k (t) is:

[0029] L k (t)=∑ χ∈{4，5，6} Q k，χ (t);

[0030] At time t, the resource-liability balance B of mining landscape patch p is p (t) is:

[0031]

[0032] Furthermore, the global supervision evaluation index E(t) is:

[0033]

[0034] Among them, A p is the area of the mining landscape patch p; is the structural connectivity of the mining landscape patch p, A p，q is the adjacency matrix, d p，q is the centroid distance of the mine landscape patch p, is the average distance of all adjacent pairs; is the global average connectivity of the mining landscape patch p; F FRAG，p (t) is the landscape fragmentation of mining landscape patch p at time t.

[0035] The mine restoration supervision method based on remote sensing image semantic segmentation of the present invention has the following beneficial effects:

[0036] This method constructs a unified resource and liability ecological indicator matrix in the technical system, regularizes the positive indicators and negative indicators respectively, and projects them onto the same number axis, making ecological data from different sources and with different physical meanings comparable, overcoming the problems of dimensional confusion and evaluation fragmentation in existing methods.

[0037] Secondly, by introducing the non-negative matrix decomposition technology in the information theory optimization framework, potential ecological function patterns are automatically extracted from large-scale ecological monitoring data, which can clearly depict the actual impact of various ecological driving factors on different ecological indicators during the restoration process, avoiding the subjective bias caused by artificial weighting and experience scoring.

[0038] Thirdly, in terms of functional interpretation, by performing sub-item calculations of resources and liabilities for each patch and forming a dynamic balance index, the ecological profit and loss status of each landscape unit can be quantitatively identified, thus achieving refined management and control of the governance unit.

[0039] Finally, the method comprehensively introduces area, structural connectivity, and landscape fragmentation into the global evaluation formula, so that the final supervision evaluation not only reflects local restoration results but also incorporates spatial network attributes and ecological risk information, forming a three-dimensional evaluation logic of space-structure-function coordination. This greatly improves the explanatory power and adaptability of the supervision results, and provides a strong data foundation and model support for dynamic supervision and scientific decision-making of ecological restoration projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of the method flow of a mine greening supervision method based on remote sensing image semantic segmentation provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0042] refer to Figure 1 : A mine restoration supervision method based on remote sensing image semantic segmentation, the method comprising:

[0043] Step 1: Based on the original ecological indicators of the mine, including resource indicators and liability indicators, a six-column time series matrix is established for each mine landscape patch. The elements in the first three columns of the time series matrix are all resource indicators, forming the resource column, representing resource attributes, and the elements in the last three columns are all liability indicators, forming the liability column, representing liability attributes.

[0044] In the mine restoration supervision method based on semantic segmentation of remote sensing images, the core idea of the first step is to convert the original ecological indicators that are mixed in time and space, diverse in types and different in dimensions into a dynamic resource-liability information carrier that can not only preserve the characteristics of the landscape pattern but also facilitate subsequent computer automatic deduction. This carrier is a time series matrix with landscape patches as rows, six types of representative indicators as columns, and monitoring moments as layers. During implementation, we first use low-altitude visible light and multispectral images from drones, satellite remote sensing data, and digital elevation models generated by lidar. Through object-oriented multi-scale segmentation algorithms and random forest classifiers, we divide the mountains disturbed by mining into landscape patches such as mining depressions, spoil dumps, reclaimed terraces, vegetation reconstruction areas, drainage ditches, and sedimentation pools, and assign a unique number to each patch. Subsequently, through coordinate unification and geometric correction, we align the images and digital terrain data of previous years to the same projection system. With the help of a patch tracking algorithm that combines area similarity with the shortest distance to the centroid, we ensure that spatial entities with the same number remain consistent in multi-temporal sequences. Even if patches merge, split, or disappear, the evolution events are recorded through time tags, thus laying a stable data skeleton.

[0045] After the patch framework was established, six landscape pattern indices were extracted to systematically reflect the progress of mine ecological restoration. Soil organic matter density, vegetation cover, and slope stability were defined as resource attributes, estimated using field soil sampling, near-infrared laboratory calibration, remote sensing inversion, linear blending of fine-pixel NDVI from drone imagery, and a high-resolution DEM load-shear-slip model. Heavy metal pollution index, potential soil erosion modulus, and landscape fragmentation were classified as liability attributes. Heavy metal pollution was interpolated from sampling points using a portable X-ray fluorescence spectrometer; soil erosion modulus was derived using multi-source inference using the RUSLE factor; and fragmentation was calculated using FRAGSTATS using the density and shape complexity of patches at the same scale. Before being written to the database, all raw values were radiometrically calibrated, cloud and fog removed, noise smoothed, and multi-temporal kriging interpolated for missing pixels to enhance spatiotemporal consistency and observation reliability. Because the physical dimensions of each indicator vary significantly and the need to ensure clear positive and negative meanings in subsequent algorithms is essential, a unified scaling process is implemented during the matrix construction phase. For each indicator, the global extreme value is retrieved across all patches and the entire historical monitoring period. Logarithmic compression is used to reduce dispersion, and the indicator is normalized to its maximum value. This ensures that resource indicators remain positive, while liability indicators are converted to negative numbers through sign flipping. This results in a symmetrical resource-liability distribution on the numerical axis and completely eliminates computational bias caused by unit differences. After normalization, the six columns of indicators are sequentially written into the matrix, with the first three columns always positive and the last three columns always negative, and no rows or columns are missing. A resource-liability matrix is generated for each monitoring moment. Multi-temporal matrices are stacked in time-labeled order to form a three-dimensional tensor, laying a unified, traceable, and weight-free mathematical foundation for subsequent non-negative matrix factorization and potential functional module analysis. When new remote sensing images or field monitoring data enter the system, the automated pipeline will reuse the same patch segmentation, indicator inversion and normalization rules to splice the new observations to the end of the tensor; if large-scale surface morphology changes occur during the monitoring period, the system can renumber the patch change events or create row records for the new patches, while the historical matrix remains archived as is, ensuring data integrity and continuity.

[0046] The semantic segmentation process for remote sensing imagery of mining landscape patches is essentially a cross-platform, multi-source, and multi-temporal remote sensing intelligent interpretation pipeline. Its goal is to interpret the complex mining surface into vector patch objects with clear ecological meaning and time-trackable tracking, leveraging spatial, spectral, textural, and three-dimensional terrain information. The entire process begins with data preparation: first, a three-layer remote sensing acquisition scheme consisting of satellite, drone, and lidar is designed based on the mining area's scope and terrain obstruction. The outermost layer uses high-resolution or WorldView imagery with resolutions better than two meters to provide a global baseline. The middle layer, using regular oblique drone aerial photography, acquires sub-meter-level visible-to-near-infrared stereo images to capture detailed changes in open pits, dump steps, and drainage ditches. The inner layer, using airborne full-waveform LiDAR, is scanned annually to generate centimeter-level DEMs and point cloud intensity information, addressing the challenges of orthorectification in areas with slope shadows and significant elevation differences. All raw images are first radiometrically corrected and aerosol-subtracted, followed by a secondary geometric correction based on a network of reference control points. This ensures that data accuracy across sensors and camera angles is consistent within a single pixel. The LiDAR-generated surface model is then co-orthorectified with the optical imagery to ensure a strict one-to-one correspondence between every pixel position at all times.

[0047] After image preprocessing, the object-oriented segmentation phase begins. To avoid over- or under-segmentation at a single scale, the system employs a multi-resolution hierarchical segmentation strategy: First, the closed-form analytical scale is automatically derived using the image's local variance and edge gradient. This formula simultaneously considers spectral homogeneity and shape compactness, resulting in a coarse segmentation. Each coarse segmentation area is then subdivided using superpixel aggregation to generate smaller patches with higher regularity. Pixel-level semantic segmentation is then performed using a deep convolutional neural network (such as DeepLabV3+), outputting probability maps for typical mining features—exposed rock in mining pits, waste rock dumps, dump steps, open drainage channels, reclaimed vegetation, natural forests and grasslands, exposed farmland, hardened building surfaces, and water sedimentation ponds. The system then superimposes this probability map with the object-oriented segmentation results and automatically fine-tunes the segmentation boundaries using the maximum probability principle and a dual threshold of boundary proximity, allowing patch boundaries to shrink or expand along the boundaries of the actual features and eliminating jagged edges caused by soft discrimination.

[0048] During the classification phase, the system extracts features for each candidate patch, including spectral mean, normalized difference vegetation index (NDVI), red edge index, gray-level co-occurrence matrix texture entropy, LiDAR reflectance intensity, average slope, surface roughness, shape index, aspect ratio, boundary complexity, and the mean squared variation (MSV) of the time series from the past three imagery periods, totaling over thirty dimensions. After three rounds of feature importance evaluation using LightGBM, redundant features with highly overlapping mutual information are removed, converging to a final input dimension of approximately fifteen. The classification model utilizes an ensemble learning framework consisting of a stack of random forests and gradient boosted decision trees. The first layer classifies the surface into five primary classes based on spectral and textural features. The second layer further subdivides the vegetation reclamation subclass into three stages: early herbaceous, mid-stage shrubs, and late-stage trees. LiDAR elevation features are expanded using tensor projection to explicitly account for differences in elevation gradients between shrubs and trees in the model. Training samples are manually annotated vectors from low-altitude ground truth drone flight strips. These samples are then overlaid with high-precision RTK transects for consistency, ensuring a strict limit of less than two percent for misclassified samples. The final classification accuracy is based on the kappa coefficient greater than 0.85 and the macro-average F1 score greater than 0.8. If the standards are not met, the system will automatically return to the feature selection stage to adjust the threshold or backtrack the segmentation scale.

[0049] To remove random noise and artifacts, the system performs a morphological closing operation on adjacent patches of the same type. A region growing algorithm is then used to aggregate patches using a dual threshold of spectral Euclidean distance and boundary curvature. This ensures that small, morphologically insignificant patches without ecological functions are incorporated into the main patch. Boundary simplification utilizes a combination of the Douglas–Peucker algorithm and least-squares smoothing to maintain the overall shape of the patch while reducing vertex redundancy. Patch vectors are then stored with additional multivariate fields: patch area, centroid coordinates, primary category, secondary category, texture composite index, mean vegetation index, mean slope, and fragmentation measure, automatically recorded for direct use by subsequent models.

[0050] Temporal alignment is key to ensuring the consistency of the row dimensions of the resource-liability matrix. The system first calculates the overlap area ratio matrix and centroid distance matrix for each patch set during each monitoring period with the previous period. Then, using a bipartite matching method, it finds the largest cross-mapping area. If a pair of patches has an overlap ratio greater than 0.6 or a centroid distance less than twice the minimum resolution and the same classification, they are considered the same evolutionary unit and inherit the same number. If a one-to-many or many-to-one match occurs, it is recorded as a split or merge event, and a uniform event suffix is added to the patch ID. New patches are serially numbered. New patches are added to the resource-liability matrix row and filled with zeros at historical moments, ensuring that the matrix dimension increases over time without missing columns. The patch temporal tracking process also uses LiDAR differential DEM to detect sudden elevation changes. If the classification results show that vegetation areas have changed to bare rock without a significant increase in elevation, it is considered a classification error and triggers manual review.

[0051] After completing the above process, the system generates a patch dataset covering all monitoring moments, forming a three-dimensional tensor index. Patch row numbers and monitoring moment column numbers are directly mapped to the resource-liability matrix data structure, eliminating the need for additional spatial index conversion for subsequent indicator normalization and matrix decomposition. The entire remote sensing image semantic segmentation chain is managed by the open-source remote sensing processing engine orchestrator. Each step is interconnected through containerized microservices, supporting GPU-accelerated image inference and CPU-parallel GIS operations. The process can complete fully automated processing of a single phase in a 100-square-kilometer mining area within four hours, significantly improving the data production efficiency of supervision projects and providing high-precision, temporally and spatially consistent basic landscape patch data for subsequent evaluation methods.

[0052] Step 2: At each monitoring moment, constrained non-negative matrix decomposition is performed on the time series matrix by iteratively minimizing the KL divergence. The goal is to decompose the time series matrix into the product of the potential ecological function module weight matrix and the potential ecological function module indicator contribution matrix, and to ensure that both the potential ecological function module weight matrix and the potential ecological function module indicator contribution matrix remain non-negative; for each potential ecological function module, its resource income and liability residual are calculated, and the ratio of resource income to liability residual is obtained as the resource-liability balance;

[0053] The algorithm first automatically determines the number of potential modules based on the matrix dimensions, setting it to the square root of the product of the number of patches and the number of indicators, rounded down, to avoid subjective parameter setting. Subsequently, by adding minimal positive values to the matrix entries and taking the absolute value of the liability column, the original data, which contains both positive and negative signs, is uniformly mapped to the non-negative domain to meet the decomposition algorithm's input non-negativity requirement. The initial sign of each entry is also recorded, ensuring that the decomposition results can still distinguish between resources and liabilities during post-processing. During the initialization phase, the "patch-module" weight matrix and the "module-indicator" contribution matrix are assigned values in a random but mean-constrained manner to ensure that both matrix elements are greater than zero and meet the overall scale constraint. The system then enters the iteration phase, using a multiplicative update rule to gradually adjust the two matrices so that their product gradually approaches the original matrix in terms of Kullback-Leibler divergence. In each iteration, the algorithm first calculates the reconstruction error based on the current estimate, then adjusts the matrix elements proportionally based on the error, and immediately performs a minimum threshold truncation after the adjustment to prevent underflow or the recurrence of zero entries.

[0054] To avoid local minima, the system sets up multiple sets of independent random initial values and runs them in parallel, simultaneously monitoring their convergence rates and residuals. The decomposition results with the smallest divergence and most stable convergence rate are retained. The search process terminates when the relative change in reconstruction error falls below a preset threshold or when the maximum number of iterations is reached. A complete trajectory of the error over iterations is output for later review by supervisory technicians. After decomposition, the algorithm first sums the "module-indicator" contribution matrix by resource and liability columns, deriving the cumulative positive contribution to resource attributes and the cumulative negative contribution to liability attributes of each potential module. This is then element-wise multiplied by the "patch-module" weight matrix and summed across the module dimension to calculate the resource benefit and residual liability values for each landscape patch. The ratio of resource benefit to residual liability for each patch is used to generate an intermediate measure, the resource-liability balance. A balance greater than one indicates that the patch's current resource accumulation outweighs its ecological liability, while a balance close to or less than one indicates that remediation measures still require improvement. At the same time, the system also performs random permutation tests and subsample cross-validation on the decomposition results. By repeatedly shuffling the rows and columns of the matrix or extracting non-overlapping subsets of monitoring periods, it verifies the stability and ecological interpretability of the potential modules, ensuring that the functional modules finally extracted are not accidental noise patterns, but rather drivers of the restoration process that can remain consistent across different time and spatial scales. The entire processing chain is completely data-driven in implementation, without relying on any empirical weights or manually set adjustment coefficients, and all algorithm hyperparameters are transparently recorded and open to the supervision interface, making it convenient for different project teams to directly reuse them in new mining areas or new restoration stages, ensuring the consistency of the evaluation system and the traceability of the results. More importantly, by embedding the resource-liability conversion mechanism into the decomposition framework, subsequent steps can directly call the patch-level balance when calculating the area-weighted average and connectivity amplification factor without the need for symbolic judgment, thereby significantly improving the efficiency and precision of supervision and evaluation, and providing a solid numerical foundation for real-time dynamic review of the mine ecological restoration process.

[0055] Step 3: Use the area of each mine landscape patch to perform a weighted average of the resource-liability balance as the main part of the restoration effect; at the same time, use the structural connectivity between each mine landscape patch as an amplification factor, and introduce the landscape fragmentation as a penalty term into the same expression to calculate the global supervision evaluation index to evaluate the effect of mine ecological restoration.

[0056] First, the precise area of each patch must be extracted. This area is directly derived from raster or vector boundaries of the same resolution, obtained through raster accumulation or polygon measurement. Area is used to ensure that the evaluation results fully reflect the spatial distribution differences of mine ecological restoration projects: given the same ecological benefits, larger patches clearly contribute more to the overall landscape function. Therefore, the system integrates the resource-liability balance of all patches using an area-weighted approach to construct a benchmark value that reflects the majority of restoration results. After completing area weighting, the system further considers the quality of spatial connectivity between patches. In a fragmented landscape, even if individual patches compensate for liabilities, overall ecological function may still be difficult to restore due to insufficient connectivity. To this end, the system first calculates a distance matrix based on the coordinates of the patch centroids. Then, incorporating information on actual obstacles such as terrain impedance, accumulation height, or road obstructions, the system uses a shortest path or landscape resistance model to determine which patches are ecologically interconnected and constructs a neighbor relationship map. In this diagram, the system assigns a distance attenuation weight to each pair of adjacent patches, and then sums the edge weights of all directly or indirectly connected to the target patch to obtain the structural connectivity of the patch; when the connectivity is higher than the global average level, the system gives an amplification effect in the evaluation system, indicating that the patch is not only in good repair itself, but also provides a cross-patch channel for species migration and ecological circulation, which helps to accelerate the spread of resources to surrounding areas.

[0057] In addition to emphasizing positive connectivity, the system also needs to mitigate the negative risks associated with excessive fragmentation. Therefore, a landscape fragmentation index is applied to each patch. This index is derived from a comprehensive calculation of patch boundary complexity and shape index, with higher values indicating greater fragmentation. To prevent evaluation results from being dominated by local extremes, the system first performs an area-weighted average of fragmentation and then introduces a penalty term to offset the positive gains from the initial connectivity amplification. This ensures that beneficial networking patterns are encouraged between patches while also promptly revealing potential ecological risks caused by localized over-excavation or vegetation disruption. Throughout this process, area, connectivity, and fragmentation are all based on objective observational data, without any artificial adjustment coefficients or empirical thresholds. Connectivity calculations rely entirely on patch geometry and the resistance between them, while fragmentation is batch-derived using a unified landscape pattern analysis software, ensuring repeatability and comparability across time and space. After calculating the area-weighted balance, connectivity amplification factor, and fragmentation penalty for all patches, the system synthesizes them on a unified scale, first by positive accumulation and then by negative deduction, to generate a comprehensive supervision index encompassing spatial structure, functional benefits, and risk compensation. Because this index maintains a single numerical value, supervisors can directly read its absolute magnitude and the magnitude of change between adjacent periods during each monitoring period. A consistently high index indicates that overall resource accumulation is steadily covering ecological liabilities and is accompanied by an optimized landscape. Stagnant or declining indexes indicate the need for more frequent on-site inspections and targeted restoration measures in areas with weak connectivity or increased fragmentation. By integrating this evaluation logic, which considers both the resource-liability benefits within patches and the landscape network structure and fragmentation risk, the system comprehensively and meticulously reflects the dynamics of mine ecological restoration without the need for subjective weights or empirical factors, providing solid data support for management decisions and subsequent project adjustments.

[0058] Furthermore, resource indicators include: soil organic matter density, vegetation cover and slope stability; the liability indicators include: comprehensive heavy metal pollution index, annual average potential soil and water loss and landscape fragmentation; in order to make the values of resource indicators always positive and the values of liability indicators always negative, and to eliminate dimensional differences, the full-time and space-time maximum value of each original ecological indicator is first obtained, and then each original ecological indicator is logarithmically compressed and normalized according to the maximum value.

[0059] Soil organic matter density reflects the abundance of carbon and nitrogen sources in the overburden layer that are available for microbial metabolism and absorption by vegetation roots. It is a key indicator for measuring the degree of soil fertility recovery and carbon sequestration potential. In mining-disturbed areas, the original topsoil is often stripped or mixed, resulting in a significant attenuation of organic matter. Therefore, an increase in organic matter density directly represents the accumulation of ecological functions on the resource side. Vegetation cover reflects the proportion of the surface covered by green vegetation and can comprehensively describe the intensity of photosynthesis, evaporation rate, and surface temperature regulation effects. Increased cover means that plant communities are more continuous in space, which has a positive effect on suppressing dust, fixing carbon, and providing a habitat. Slope stability is used to evaluate the ability of the profile to resist sliding, collapse, and gully development. Its level depends on the integrity of the bedrock, the smoothness of the drainage system, and the reinforcement effect of the vegetation root system. High stability indicates that engineering support and ecological slope protection measures have effectively reduced gravity erosion, which is an important safety attribute on the resource side. The Heavy Metal Comprehensive Pollution Index (HMI) compares the concentrations of multiple elements, such as lead, cadmium, and arsenic, to environmental benchmark limits, revealing the toxic debt incurred by the leaching of waste rock and tailings from mining areas to soil and groundwater. A higher index indicates a higher potential ecological and food chain risk. The annual average potential soil erosion is derived from a coupled calculation of rainfall erosion, slope length and gradient, vegetation cover, and soil erodibility. Higher values indicate that exposed or loose areas are more susceptible to erosion by heavy rain, transporting sediment and nutrients downstream. This is a core indicator of surface erosion debt. Landscape fragmentation measures the spatial continuity of a landscape by counting the number of patches, their perimeter-area relationship, and their shape complexity. Higher values indicate that previously coherent habitats have been fragmented by mining sites, roads, and spoil dumps, hindering species migration and gene flow, thereby creating a structural debt. These six indicators cover the key processes of soil quality, vegetation status, terrain safety, chemical pollution, physical erosion, and spatial pattern. The increase of resource indicators or the decrease of liability indicators can objectively reflect the gradual recovery of ecological service functions and the gradual reduction of environmental pressure during mine ecological restoration. Therefore, they are selected as the column vector basis of the resource-liability matrix of this method.

[0060] The system first performs logarithmic compression on the raw indicators, leveraging the monotonically increasing nature of the logarithmic function to reduce the stretching effect of outliers on the overall distribution. This ensures that resource and liability data retain resolution in low-level regions and achieve moderate convergence in high-level regions. Normalization is then performed using the previously cached column-level maximum as the denominator, mapping each entry to a dimensionless range between zero and one. Because resources and liabilities naturally have opposite signs, the system immediately performs a sign flip on the three columns: the comprehensive heavy metal pollution index, the annual average potential soil erosion, and the landscape fragmentation. These values are multiplied by -1, shifting all values into the negative range. However, soil organic matter density, vegetation cover, and slope stability retain their positive signs. This ensures that resource indicators are consistently positive, liability indicators are consistently negative, and all six columns are on the same scale, providing clean data input for the subsequent resource-liability matrix factorization. The entire process involves no artificial empirical coefficients: the maximum value used in normalization is directly derived from objective observations, the base of logarithmic compression is the system's default constant, and sign flipping is a fixed operation, thus ensuring consistent processing links across different monitoring periods. When a new period of remote sensing imagery or measured data enters the system, if a new extreme value is found that exceeds the historical record, the normalization constant is immediately replaced and an incremental recalculation is triggered, ensuring that data for all time periods remains comparable based on the latest benchmark. At the same time, the system logs the extreme value replacement record, allowing supervisors to trace changes in indicator scales and their impact on the comprehensive evaluation value.

[0061] Furthermore, in step 2, for each potential ecological function module, the process of calculating its resource benefits and liability residual includes: adding the cumulative contribution of each potential ecological function module to the resource column to obtain the resource-side weight of the potential ecological function module; then adding the cumulative contribution of the same potential ecological function module to the liability column to obtain the liability-side weight; for each mining landscape patch, multiplying its weight on each potential ecological function module with the resource-side weight and summing them to obtain the resource benefits of the mining landscape patch; multiplying the indicator contribution on each potential ecological function module with the liability-side weight and summing them to obtain the liability residual. The ratio of resource benefits to liability residual is the resource-liability balance of the mining landscape patch.

[0062] When it is greater than 1, it means that the repair benefits have exceeded the residual liabilities; if it is less than 1, it means that the residual liabilities are still dominant and the governance efforts need to be strengthened.

[0063] In the mine restoration supervision method based on semantic segmentation of remote sensing imagery, the core task of the second phase is to map the decomposed potential ecological function modules back to the specific landscape patch level, thereby quantifying each patch's resource surplus and residual liability at the current moment. This implementation process first requires reading two result matrices of the same dimension: one, the module indicator contribution matrix, records the explanatory power of each potential ecological function module for six landscape pattern indicators; the other, the patch module weight matrix, records the weight of each landscape patch's participation in each potential module. The system then divides the module indicator contribution matrix into resource and liability columns based on indicator attributes. Then, within the module dimension, row-wise summation is performed on the resource column to generate a resource-side weight for each potential module. Similarly, the same row-wise summation is performed on the liability column to obtain its corresponding liability-side weight. At this point, the potential module is assigned two distinct and independent numerical labels: one representing the module's average contribution to enhancing ecological resources and the other representing its average contribution to preserving ecological liabilities. The system then enters a patch-level calculation loop. For each landscape patch, the program iterates through all potential modules, multiplying the patch's weight in a given module by the corresponding resource-side weight and accumulating the result in a resource benefit accumulator. Simultaneously, the patch's contribution to the module's indicator is multiplied by the liability-side weight and accumulated in a liability residual accumulator. At the end of the iteration, the value in the resource benefit accumulator represents the positive ecological benefits currently accruing to the patch through the restoration project, while the liability residual accumulator represents the negative ecological pressures that have not yet been offset by the project measures. To provide supervisors with a directly interpretable, unified metric, the system calculates a ratio between the two to determine the resource-liability balance. A ratio greater than one indicates that the resource-side benefits have substantially offset the liability-side residual, indicating a net restoration of the patch. A ratio close to or below one indicates that the liability is still dominant and that measures such as soil cover improvement, vegetation replanting, heavy metal passivation, or drainage management should be strengthened. The entire loop only needs to complete two multiplication and addition operations equal to the number of potential modules and one division operation for each patch. The computational complexity increases linearly with the total number of patches. The system can complete global updates of large-scale mining areas in minutes and write the results to the supervision database in real time. To avoid numerical instability caused by extremely small denominators, the program automatically checks whether the residual liability accumulator is below the machine precision threshold before entering the ratio calculation. If it is below the threshold, the threshold is used instead to ensure that the ratio is always definable. All input parameters come from objective observation data and the inherent output of the decomposition algorithm. No empirical weights or artificial adjustment factors are introduced in the calculation process, so the results remain comparable and consistent for different mining areas and different restoration stages.After completing the balance calculation for all patches, the system will conduct a longitudinal comparison of the latest balance of each patch with the historical monitoring sequence, automatically draw the resource-liability trajectory evolving over time, and perform a trend analysis on the slope of the trajectory; if it is found that the balance of individual patches has declined for several consecutive periods or has hovered near the critical value for a long time, the supervision platform will push early warning information to the ecological engineering management end, requiring the on-site team to conduct targeted investigations or reinforcements on the patch.

[0064] Furthermore, at time t, each element M in the time series matrix M(t) of the mining landscape patch p p，χ (t) is:

[0065]

[0066] Among them, sgn(χ) = +1, when χ∈{1, 2, 3}, represents the resource column; sgn(χ) = -1, when χ∈{4, 5, 6}, represents the liability column; χ = 1, represents the soil organic matter density; χ = 2, represents the vegetation cover; χ = 3, represents the slope stability; χ = 4, represents the total heavy metal pollution index; χ = 5, represents the annual potential soil and water loss modulus; X = 6, represents the landscape fragmentation; X p，χ (t) is the original ecological index of the mine landscape patch p in the time series matrix M(t) at time t, corresponding to the column χ; p′，χ (t′) is the original ecological indicator of the corresponding column χ in the time series matrix M(t) of the mining landscape patch p′ at time t′.

[0067] First, through sgn(χ), the resource column {χ = 1, 2, 3} is fixed to a positive sign, and the liability column {χ = 4, 5, 6} is fixed to a negative sign, so that the resource indicators (i.e., soil organic matter density, vegetation cover, and slope stability) always increase in the positive direction on the numerical axis, while the liability indicators (i.e., total heavy metal pollution index, annual average potential soil and water loss modulus, and landscape fragmentation) decrease in the negative direction on the numerical axis. This artificially introduced sign opposition ensures that resources and liabilities can be directly separated by positive and negative positions during subsequent matrix decomposition, while avoiding the logical confusion caused by sign reversal of the same column data in different time periods; secondly, the numerator uses the measured value and the full-time The ratio of null maxima means that any indicator achieves a ratio of one at its historical peak and a relative value less than one at other points in time and space. This brings heterogeneous indicators, such as concentrations, percentages, and stability coefficients, which originally had different dimensions, back onto a common scale. The denominator uses the global extreme value rather than the local extreme value to ensure continuous comparability between new and old data during cross-year monitoring. Furthermore, the logarithmic compression of ln(1+·) overcomes the stretching of the distribution caused by extremely high values while preserving the resolution of small-scale differences in ecological assessment. The addition of one ensures that the logarithm remains defined when the measured value is zero, avoiding missing values due to zero values in bare rock patches or patches with very low pollution. Overall, this transformation causes resource indicators to show a gradual increase from zero toward a positive upper limit during restoration, while liability indicators show a gradual convergence from a negative upper limit toward zero during remediation. Because the maximum amplitudes on both the positive and negative sides are bounded by the same function, resource gains and liability reductions can be directly compared in absolute terms in subsequent calculations, thus achieving a dimensionless interpretation of the core quantity of resource-liability balance. More importantly, the only scaling constant in the formula is Based on objective observations and lacking empirical factors, this ensures that evaluation results from different mining areas, restoration phases, and even different monitoring technologies are aligned within the same theoretical framework. When new extreme values appear during monitoring, the system automatically updates the constant and recalculates the corresponding columns, extending the time series depth without disrupting the existing data logic. Precisely because of this, the matrix M(t) maintains the sign distinction between resources and liabilities while achieving scale normalization and dynamic comparability of the six-dimensional landscape pattern index. This provides a rigorous, robust, and non-subjective data foundation for subsequent non-negative matrix factorization based on the Kullback–Leibler divergence and the precise calculation of the resource-liability balance.

[0068] Furthermore, let the number of mining landscape patches be P, then the number of potential ecological function modules is calculated by Perform floor operation to obtain the value.

[0069] In this method, the number of potential ecological function modules K is Its design idea is to keep the balance of hidden dimension and observation dimension in terms of information at approximately square root level: on the one hand, the number of columns of the time series matrix is fixed at 6 landscape pattern indices, so the product of the overall observation dimension and the number of patches P can be approximately regarded as the "total degrees of freedom" that need to be explained during decomposition; on the other hand, if P or 6 is directly used as the number of potential modules, not only will there be serious dimensional redundancy, but the non-negative matrix decomposition will also fall into the dilemma of sparse solution dilution or overfitting; multiplying the two and then taking the square root and rounding down is equivalent to geometrically compromising the "row information" and "column information" to the same order of magnitude, so that the average explanatory load borne by each potential module is neither too large nor too small. Specifically in the scenario of mine ecological restoration, this compromise strategy can ensure that the decomposition model has sufficient capacity to capture the differentiated ecological function patterns between patches, while avoiding the introduction of additional noise due to too many potential modules, which leads to unnecessary numerical diffusion in the subsequent calculation of resource benefits and residual liabilities; at the same time, rounding down ensures that the result is an integer, keeping the computational complexity and storage requirements within a controllable range, and as the monitoring network captures more new patches (that is, P grows), the number of potential modules will also expand synchronously at a power rate, so that the model capacity can automatically adapt to the needs of refined ecological patterns in long-term monitoring, ensuring the robustness and scalability of the evaluation system.

[0070] Furthermore, assuming that the potential ecological function module weight matrix is P(t) and the potential ecological function module indicator contribution matrix is Q(t), the constrained non-negative matrix decomposition is performed on the time series matrix by minimizing the KL divergence iteration using the following formula:

[0071]

[0072] And the following equality exists:

[0073]

[0074] Among them, P p，k (t) represents the weight of mining landscape patch p on potential ecological function module k at time t; Q k，χ (t) represents the indicator contribution of potential ecological function module k at time t.

[0075] The core goal of the matrix decomposition stage is to decompose the mine ecological landscape pattern data M(t) that has been sign-regulated and log-normalized into two completely non-negative and ecologically easy-to-interpret low-rank factors, so as to extract a set of implicit ecological function driving patterns from the massive patch-indicator observations and quantitatively track the dynamic contribution of resource accumulation and debt decay through these patterns. To this end, this method adopts a multiplicative update NMF framework with Kullback–Leibler divergence as the loss function and non-negativity constraints as the boundary conditions, formally written as And it is required that at any time t the approximate reconstruction equation is satisfied

[0076] Unlike traditional NMF that uses the Euclidean norm or Frobenius norm, KL divergence It has a natural relative entropy interpretation and can quantify the difference in information distribution between two non-negative matrices. When the elements of M(t) show a long-tail distribution in the low-value-high value range and the resource column and the liability column differ by dozens of times, the KL metric can punish the neglect of sparse peaks better than the square error, thus maintaining high sensitivity in the extreme case of "extremely low resources and extremely high liabilities" in the early stage of repair. In order to ensure that the number of potential functional modules is balanced with the observation dimension and to avoid overfitting caused by the curse of dimensionality, this method sets K to The number of patches P represents the spatial granularity of the landscape, and the fixed number of columns 6 represents the dimension of the ecological process. The product of the two approximately describes the overall information content of the system, while the square root gives the equilibrium dimension when compressed into the latent space; rounding down ensures that the number of factors is an integer and does not exceed the observation degrees of freedom.

[0077] When solving specifically, first and Assign a random positive initial value proportional to the row mean to reduce gradient oscillation; then apply the Lee and Seung classic multiplication update rule to iterate in the sense of KL divergence: Update P p，k (t), M p，χ (t) / (P(t)Q(t)) p，χ As the gain factor with the existing Q k，χ (t) multiply and normalize, update Q k，χ (t) then the same gain coefficient is combined with the current P p，k (t) are multiplied and then normalized. To avoid division by zero and maintain numerical stability, a machine minimum of 10-12 is explicitly added to the denominator in each calculation, and entries below this threshold are truncated and reset after the iteration. The algorithm stops when the relative residual decreases less than 10-5 or when the maximum number of iterations is 500. Ten sets of independent random initial values are used for parallel solution, and the set with the smallest KL residual and the smoothest convergence curve is selected as the final result. The resulting row vector P(t) can be viewed as the probability density of the patch's attribution to each functional module, and the sum of its elements is approximately the full ecological weight of the patch. The resulting column vector Q(t) describes each module's weighted contribution to the six indicators. The sum of its contributions to the resource column (χ = 1, 2, 3) is the module's resource-side weight, and the sum of its contributions to the liability column (χ = 4, 5, 6) is the module's liability-side weight. In order to map these potential quantities back to operational engineering semantics, the system performs resource benefit and liability residual inference in the patch dimension: for each patch p, its weight P on module k is p，kThe resource benefit is obtained by multiplying the weight of the corresponding module's resource side and summing over all k. The weight of the same patch in module k is then multiplied by the weight of the module's liability side, and the sum is calculated to obtain the liability residual. The ratio of the two is the resource-liability balance. Because P(t) and Q(t) are both non-negative and the resource and liability weights are generated by explicit row or column summation, the resource benefit and liability residual are naturally of the same dimension and directly comparable. This not only avoids subjective coefficients but also allows each calculation step to be traced back to the physical interpretation of the matrix elements: if a module is judged to be "vegetation regeneration-dominated," its contribution is high in the resource column and low in the liability column; if a module corresponds to "heavy metal enrichment," its contribution is high in the liability column. After iterative convergence, the system simultaneously outputs the full-period KL residual curve, sparsity indices of the two factor matrices, and a module-indicator contribution heat map for ecological semantic annotation by supervision experts. If a reasonable ecological correspondence cannot be found for some modules, the dimensionality can be expanded by increasing the K-rerun decomposition or by introducing new observation indicator columns in the backend database. It is worth emphasizing that when the new monitoring time t+1 arrives, the system only needs to append the new matrix slice to the end of the tensor and use the previous period's P(t) and Q(t) as warm-start for a small number of iterations to complete the online update, which greatly reduces the re-decomposition overhead; if P surges due to the addition of new patches, the system will automatically recalculate The old factor matrix is expanded with zero filling to ensure that the model capacity and spatial granularity grow synchronously.

[0078] Furthermore, when the constrained non-negative matrix decomposition is completed by iteratively minimizing the KL divergence of the time series matrix, P p，k The update formula of (t) is:

[0079]

[0080] Among them, P p，k，new (t) is the updated P p，k (t); P p，k′ (t) represents the weight of mining landscape patch p on potential ecological function module k′ at time t; Q k′，χ (t) represents the indicator contribution of the potential ecological function module k′ at time t; Q k，χ The update formula of (t) is:

[0081]

[0082] Among them, Q k，x，new (t) is the updated Q k，χ (t).

[0083] The first formula updates P by assigning weights to the reconstruction residuals of each mining landscape patch p on the dimensions of the potential ecological function module. p，k(t), the specific approach is to take the observed value M of the patch for all indicators x p，χ (t) and the current reconstruction value ∑ k′ P p，k′ (t)Q k′，χ (t) is used as the gain coefficient, which makes the weight of the indicator items that have not been fully explained by reconstruction increase more for the module k; then the contribution of the module to each indicator Q k，χ (t) is the weighted average of the weights in the indicator dimension, and the adjustment amount in line with the KL divergence gradient direction is obtained, and then multiplied by the P of the previous iteration p，k (t) Get the new value P p，k，new (t), and add 10-12 to the denominator to prevent zero division and ensure non-negativity. The second formula gives the indicator contribution matrix Q when the dimension of the potential ecological function module remains unchanged. k，χ (t) Perform dual update: It updates the weights P of all patches on module k. p，k (t) is used as the weight, and the observation and reconstruction ratio of each indicator is weighted averaged, thereby absorbing the unexplained residual in the indicator dimension, and then multiplying it with the old value to obtain Q k，χ，new (t); Here, the same small constant is added to the denominator to ensure that the updated elements are stably greater than zero.

[0084] When these two updates are performed alternately, the product of P(t) and Q(t) continues to shrink in the direction of decreasing KL divergence within the artificial non-negative constraint, and the multiplication form ensures that if the initial value is non-negative, then all subsequent iterations remain non-negative, so that the weight and contribution of the ecological functional module always correspond to the positive interpretable amount in terms of numerical value. In the case of mine ecological restoration, this update mechanism means that each iteration will allocate more explanatory power to functional modules that can simultaneously reduce heavy metal pollution, reduce soil erosion, and improve vegetation and soil quality, and will automatically weaken the weights of those modules that do not adequately explain the liability indicators; when the iteration converges, P p，k (t) will focus on a few modules with strong resource contribution, high connectivity potential or effective erosion control characteristics, while Q k，X The column vector of (t) shows a typical positive repair pattern with high resource columns and low debt columns, or vice versa, a pressure pattern dominated by debt. Since the implicit ratio term in the update is essentially the relative error between the actual observation and the model prediction, the algorithm performs fine-grained error feedback on each data record, so it can still converge stably even when the remote monitoring image quality fluctuates or local field measurement points are missing; on the other hand, the weight normalization denominator ∑ X Q k，χ (t)+10-12 and

[0085] ∑ p P p，k(t)+10-12 ensures the conservation of total weight within modules and patches, preventing the updated matrix from exploding with infinite gain. Through this rigorous iterative logic, the present invention not only mathematically ensures a monotonically decreasing KL divergence, economically utilizing multiplication updates to avoid explicit step-size adjustments for negative gradients, but also allows each weight adjustment to be interpreted ecologically as "redistributing under-explained restoration effects or residual pressures" to the most suitable potential functional modules, significantly improving the model's explanatory transparency and the operability of supervisory decisions.

[0086] Furthermore, at time t, the resource income R of the potential ecological function module k is k (t) is:

[0087] R k (t)=∑ χ∈{1，2，3} Q k，χ (t);

[0088] At time t, the residual liability L of the potential ecological function module k is k (t) is:

[0089] L k (t)=∑ X∈{4,5,6} Q k，X (t);

[0090] At time t, the resource-liability balance B of mining landscape patch p is p (t) is:

[0091]

[0092] When the contribution value of X∈{1,2,3} is significantly greater than that of X∈{4,5,6}, it indicates that the ecological mechanism of module k is more inclined to positive restoration functions, such as efficient vegetation regeneration or landform stabilization measures; on the contrary, if the contribution of the liability side is higher than that of the resource side, it indicates that the module captures more stress processes such as heavy metal enrichment, erosion intensification or landscape fragmentation. Next, the resource-liability balance The participation weight P of patch p in all potential functional modules p，k (t) is used as the projection coefficient to convert the module-level resource benefits R k (t) and the residual liability L k (t) are mapped back to the patch scale: the numerator is the weighted superposition of all positive functions of the patch, and the denominator is the total ecological pressure borne by the patch; because the same weight vector acts on both types of totals, the balance value becomes a dimensionless ratio to measure profit and loss. p When (t)>1, the accumulated resource benefits of the patch exceed the remaining liabilities, which means that the management measures such as soil improvement, vegetation replanting, and slope reinforcement have achieved phased net benefits at this location; if Bp If (t)≤1, it means that the liability is still dominant and it is necessary to continue to put in passivation materials, add intercepting drainage ditches, or optimize vegetation configuration. Adding the extremely small constant 10-12 to the denominator can prevent the occurrence of zero division errors when the liability value approaches zero. At the same time, it will not have a perceptible impact on the interpretation logic. Because in the extreme case where the repair has almost completely eliminated the liability, the balance degree will be close to infinity, which just reminds the supervisor that the patch can enter the maintenance period rather than the key treatment period. In this way, P p，k (t), R k (t) and L k (t) They jointly transform high-dimensional and complex ecological pattern information into a real-time indicator that can be interpreted for a single landscape. This not only inherits the precise reconstruction of the observed distribution by non-negative matrix decomposition in mathematics, but also provides a clear and executable basis for zoning management for the supervision system in engineering: an upward-sloping balance curve indicates that restoration measures are taking effect at an accelerated pace, while a downward-sloping balance curve indicates that the management intensity or method needs to be adjusted; a heat map of the spatial distribution of multiple patches can clearly highlight risk concentration areas, laying a reliable underlying data foundation for subsequent area weighting and connectivity amplification steps, thereby ensuring that the entire mine ecological restoration supervision and evaluation chain is closed and self-consistent at the theoretical and practical levels.

[0093] Furthermore, the global supervision evaluation index E(t) is:

[0094]

[0095] Among them, A p is the area of the mining landscape patch p; is the structural connectivity of the mining landscape patch p, A p，q is the adjacency matrix, d p，q is the centroid distance of the mine landscape patch p, is the average distance of all adjacent pairs; is the global average connectivity of the mining landscape patch p; F FRAG，p (t) is the landscape fragmentation of mining landscape patch p at time t.

[0096] Area A p As an objective measure of geographical entities, it is placed symmetrically between the numerator and the denominator. On the one hand, its role is to provide a balance between resources and liabilities. p (t) and the structural connectivity amplification factor provide a weighted basis, so that patches with larger areas contribute more significantly to the total score; on the other hand, by summing the areas of all patches and placing them in the denominator to achieve global normalization, the value range of E(t) is not distorted by scale differences when comparing across mining areas. Secondly, the patch-level balance B p(t) itself has compressed resource benefits and residual liabilities into a dimensionless ratio, representing the effectiveness of local restoration. If the ratio is greater than one, it indicates net benefits; if it is less than one, it indicates net liabilities. This indicator, under the effect of area weight, reflects the logic of prioritizing the improvement of the total score of "large patches with high benefits".

[0097] Again, the connectivity correction term Reflects the quality of spatial network:

[0098] The adjacency matrix is combined with distance decay to automatically consider patch area and spatial resistance in the connected edge weight, ensuring that the calculation results are positively correlated with the actual seed spread, animal migration path or hydrological channel length; and then compared with the global average connectivity. By making a ratio, the connection advantage or disadvantage of each patch can be mapped to a relative quantity, thus forming a multiplicative amplification factor after the "+1" translation: When the factor is 2, it means that the baseline enhancement can be obtained when the connectivity reaches the average; if a plaque is isolated, C p (t) is close to zero, and the amplification factor approaches 1; if a patch is located at the node of the ecological corridor, C p (t) is much higher than the average, and the exponential growth is suppressed by sublinearity, which both recognizes its network contribution and avoids numerical imbalance beyond the resource / liability semantics; the denominator is increased by 10 -12 It is only for numerical stability, which is much smaller than the actual value of connectivity and does not affect the ecological interpretation.

[0099] The last part As an overall penalty item deducted from the total score, it takes the area-weighted average of the overall landscape fragmentation of the mining area and then squares it. The square root function makes the fragmentation risk and penalty show a marginal decreasing relationship: when the fragmentation is medium-low, the new cutting significantly lowers the index, urging the supervisor to intervene in time; when the fragmentation is already extremely high, the further damage to the ecological network caused by the new fragmentation decreases due to the denominator, avoiding the index from reaching negative infinity and maintaining comparability; area weighting also ensures that large and highly fragmented patches contribute the most to the penalty. Overall, the formula converts the local function index B into a pBy mapping area and connectivity to the landscape macro-level and then applying overall correction based on fragmentation risk, E(t) mathematically implements a three-step coupling: addition, which represents the normalized aggregation of areas to satisfy mass conservation, multiplication, and subtraction, which reflects the amplification of positive benefits by connectivity, and fragmentation, which suppresses system stability. Numerically, because all sub-items are derived from measurable or calculable data and lack empirical coefficients, E(t) maintains a consistent scale across time series and is portable when comparing different mining areas or different remediation plans within the same mining area. Its ecological implication is that when the majority of the mining area's area weights are concentrated in patches with high balance, high connectivity, and low fragmentation, E(t) achieves positive and large values, indicating that the current restoration strategy is both locally enhancing ecological services and spatially optimizing the network. If monitoring indicates a decline in connectivity or an increase in fragmentation, even if balance temporarily improves, the overall score may decline, suggesting that restoration projects should shift from single-point remediation to structural measures such as corridor development and gentle slope reclamation.

[0100] When the monitoring sequence is extended, the slope of the exponential curve can be used to evaluate the overall restoration speed, and the width of the fluctuation can be used to evaluate system stability. In the management platform, when the E(t) value exceeds the set warning threshold, the system automatically certifies it as a "safe zone," between the two thresholds as an "observation zone," and below the lower threshold, triggering a "risk zone" alarm. Targeted recommendations are then given based on the contribution ratio of connectivity to fragmentation, such as replanting connecting corridors, reducing slope cuts, or implementing fragmented habitat reconnection projects. Because each data stream in the formula is extracted from remote sensing rasters, GIS topology, or the output of the previous iteration using automated scripts, the entire evaluation chain can be run within minutes after the monitoring data is uploaded, and the management panel can be updated in real time. This enables mine ecological restoration supervision to move from traditional static discrete sampling to dynamic continuous digital twins. Through the single value of E(t), ecological gains and losses, spatial patterns, and risk pressures can be accurately perceived, providing decision makers with a basis for refined regulation based on big data.

[0101] Assume that the mining area is divided into four landscape patches, numbered p = 1 to 4, and the corresponding measured original ecological indicators are: soil organic matter density [X p，1 (t1)] 8.00 g·kg -1 ,4.00,1.00,6.00,vegetation cover[X p，2 (t1)] are 70.0%, 35.0%, 15.0%, 55.0%, respectively, and the slope stability coefficient [X p，3 (t1)] were 0.85, 0.45, 0.30, 0.65, and the total heavy metal pollution index [X p，4 (t1)] are 2.00, 5.00, 10.0, 3.00 respectively, and the annual average potential soil and water loss modulus [X p，5 (t1)] 500, 800, 1500, 600 t·km respectively -2 ·a-1 , landscape fragmentation [X p，6 (t1)] are 0.30, 0.50, 0.80, and 0.40 respectively. Because there is only one monitoring period, the maximum value of the six columns of indicators in all time and space is the maximum observation value in the column: maxX χ ={8.00, 70.0, 0.85, 10.0, 1500, 0.80}.

[0102] Substitute into the normalization and symbol regularization formula

[0103]

[0104] We can get four numerical vectors M 1：4 (t1):

[0105]

[0106] The number of rows P = 4, and the number of columns is fixed at 6, so the number of potential ecological function modules is:

[0107]

[0108] For demonstration purposes, we will directly give a feasible decomposition result (satisfying the non-negative constraint and the product matrix approximating M) without actual iteration. p，χ ), where the patch-module weight matrix is:

[0109]

[0110] Module-Indicator Contribution Matrix:

[0111]

[0112] The resource contribution of the three columns and the liability contribution of the three columns are added row by row to get the resource income R of each module. k (t1) and the residual liability L k (t1):

[0113]

[0114] Then, the weights of the patches on each module are projected onto the resource side and the liability side to obtain the resource income, liability residue, and form the resource-liability balance:

[0115]

[0116] Calculation results:

[0117] B1=2.75, B2=1.36, B3=0.40, B4=2.27.

[0118] The area calculated using remote sensing data is {A1, A2, A3, A4} = {3.0, 2.0, 4.0, 1.0} hm 2 The connectivity of the patch centroid is calculated using the adjacency matrix A. p，q =1 only in (1, 2), (1, 4), (2, 3), (3, 4) are adjacent in pairs, and the rest are zero. The distance matrix takes d 1，2 =150m,d 1，4 =100m,d 2，3 =120m,d 3，4 =180m, the average distance between all adjacent pairs then:

[0119]

[0120] Values: C1 = 0.82, C2 = 0.75, C3 = 0.69, C4 = 0.75,

[0121] Landscape fragmentation is directly measured using the original measure {F FRAG，p (t1)}={0.30, 0.50, 0.80, 0.40}. Now we can substitute the global supervision evaluation index; first calculate the area normalized positive numerator:

[0122] ∑ p A p =10.0;

[0123]

[0124] Divide by the total area to get 3.03. Then calculate the fragmentation penalty:

[0125]

[0126] So the final index is:

[0127] E(t1)=3.03-0.74≈2.29.

[0128] Because E(t1) is significantly greater than zero and above the preset threshold of 1.0, this indicates that the overall restoration benefits at monitoring time t1 have already outweighed the ecological liabilities and have been significantly enhanced by connectivity gains. However, Patch 3 has a local balance of only 0.40 and the highest degree of fragmentation, remaining a core area of unrelieved ecological pressure. Prioritizing heavy metal passivation, slope shading, and corridor reconnection measures in the next phase of governance plans is crucial. The entire process, from raw indicator collection to E(t) output, can be completed within minutes using scripted automation. It can also be automatically updated with the arrival of the next round of remote sensing or ground monitoring data, achieving a real-time digital twin closed-loop for mine ecological restoration supervision.

[0129] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A mine restoration supervision method based on remote sensing image semantic segmentation, characterized by: The method comprises: Step 1: Based on the original ecological indicators of the mine, including resource indicators and liability indicators, a six-column time series matrix is established for each mine landscape patch. The elements in the first three columns of the time series matrix are all resource indicators, forming the resource column, representing resource attributes, and the elements in the last three columns are all liability indicators, forming the liability column, representing liability attributes. Step 2: At each monitoring moment, constrained non-negative matrix decomposition is performed on the time series matrix by iteratively minimizing the KL divergence. The goal is to decompose the time series matrix into the product of the potential ecological function module weight matrix and the potential ecological function module indicator contribution matrix, and to ensure that both the potential ecological function module weight matrix and the potential ecological function module indicator contribution matrix remain non-negative; for each potential ecological function module, its resource income and liability residual are calculated, and the ratio of resource income to liability residual is obtained as the resource-liability balance; Step 3: Use the area of each mine landscape patch to perform a weighted average of the resource-liability balance as the main part of the restoration effect; at the same time, use the structural connectivity between each mine landscape patch as an amplification factor, and introduce the landscape fragmentation as a penalty term into the same expression to calculate the global supervision evaluation index to evaluate the effect of mine ecological restoration.

2. The mine restoration supervision method based on remote sensing image semantic segmentation according to claim 1, characterized in that: Resource indicators include: soil organic matter density, vegetation cover and slope stability; the liability indicators include: comprehensive heavy metal pollution index, annual average potential soil and water loss and landscape fragmentation; in order to make the values of resource indicators always positive and the values of liability indicators always negative, and to eliminate dimensional differences, the full temporal and spatial maximum value of each original ecological indicator is first obtained, and then each original ecological indicator is logarithmically compressed and normalized according to the maximum value.

3. The mine restoration supervision method based on remote sensing image semantic segmentation according to claim 2, characterized in that: In step 2, for each potential ecological function module, the process of calculating its resource benefits and liability residuals includes: adding the cumulative contributions of each potential ecological function module to the resource column to obtain the resource-side weight of the potential ecological function module; then adding the cumulative contributions of the same potential ecological function module to the liability column to obtain the liability-side weight; for each mining landscape patch, multiplying its weight on each potential ecological function module with the resource-side weight and summing them to obtain the resource benefits of the mining landscape patch; multiplying the indicator contribution on each potential ecological function module with the liability-side weight and summing them to obtain the liability residual. The ratio of resource benefits to liability residual is the resource-liability balance of the mining landscape patch. When it is greater than 1, it means that the restoration benefits have exceeded the liability residual; if it is less than 1, it means that the liability residual is still dominant and the governance efforts need to be increased.

4. The mine restoration supervision method based on remote sensing image semantic segmentation according to claim 1, characterized in that: At time t, each element M in the time series matrix M(t) of the mining landscape patch p p,χ (t) is: Among them, sgn(χ)=+1, when χ∈{1,2,3}, represents the resource column; sgn(χ)=-1, when χ∈{4,5,6}, represents the liability column; χ=1, represents the soil organic matter density; χ=2, represents the vegetation cover; χ=3, represents the slope stability; χ=4, represents the total heavy metal pollution index; χ=5, represents the annual potential soil and water loss modulus; χ=6, represents the landscape fragmentation; X p,χ (t) is the original ecological index of the mine landscape patch p in the time series matrix M(t) at time t, corresponding to the column χ; p′,χ (t′) is the original ecological indicator of the corresponding column χ in the time series matrix M(t) of the mining landscape patch p′ at time t′.

5. The mine restoration supervision method based on remote sensing image semantic segmentation according to claim 4, characterized in that: Assuming the number of mining landscape patches is P, the number of potential ecological function modules is calculated by Perform floor operation to obtain the value.

6. The mine restoration supervision method based on remote sensing image semantic segmentation according to claim 5, characterized in that: Assuming that the potential ecological function module weight matrix is P(t) and the potential ecological function module indicator contribution matrix is Q(t), the constrained non-negative matrix decomposition is performed on the time series matrix by minimizing the KL divergence iteration using the following formula: And the following equality exists: Among them, P p,k (t) represents the weight of mining landscape patch p on potential ecological function module k at time t; Q k,χ (t) represents the indicator contribution of potential ecological function module k at time t.

7. The mine restoration supervision method based on remote sensing image semantic segmentation according to claim 6, characterized in that: When the constrained non-negative matrix decomposition is completed by iteratively minimizing the KL divergence of the time series matrix, P p,k The update formula of (t) is: Among them, P p,k,new (t) is the updated P p,k (t); P p,k′ (t) represents the weight of mining landscape patch p on potential ecological function module k′ at time t; Q k′,χ (t) represents the indicator contribution of the potential ecological function module k′ at time t; Q k,χ The update formula of (t) is: Among them, Q k,χ,new (t) is the updated Q k,χ (t).

8. The mine restoration supervision method based on remote sensing image semantic segmentation according to claim 7, characterized in that: At time t, the resource income R of the potential ecological function module k is k (t) is: R k (t)=∑ χ∈{1,2,3} Q k,χ (t); At time t, the residual liability L of the potential ecological function module k is k (t) is: L k (t)=∑ χ∈{4,5,6} Q k,χ (t); At time t, the resource-liability balance B of mining landscape patch p is p (t) is:

9. The mine restoration supervision method based on remote sensing image semantic segmentation according to claim 8, characterized in that: The overall supervision evaluation index E(t) is: Among them, A p is the area of the mining landscape patch p; is the structural connectivity of the mining landscape patch p, A q,q is the adjacency matrix, d p,q is the centroid distance of the mine landscape patch p, is the average distance of all adjacent pairs; is the global average connectivity of the mining landscape patch p; F FRAG,p (t) is the landscape fragmentation of mining landscape patch p at time t.

Images