Mine re-greening supervision method based on remote sensing image semantic segmentation

Abstract

The present application relates to the technical field of ecological environment data monitoring, and more particularly to a mine re-greening supervision method based on remote sensing image semantic segmentation.The method comprises the following steps: step 1: based on the original ecological index of a mine, a six-column time sequence matrix is established for each mine landscape patch; step 2: at each monitoring time, the time sequence matrix is iteratively decomposed by constrained non-negative matrix factorization by minimizing the K-L divergence, and the objective is to decompose the time sequence matrix into the product of a latent ecological function module weight matrix and a latent ecological function module index contribution matrix; and step 3: the resource-liability balance degree is weighted and averaged using the area of each mine landscape patch to evaluate the effect of mine ecological restoration.The present application realizes dynamic, fine and structured monitoring and evaluation of the mine ecological restoration process.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of ecological environment data monitoring, and particularly relates to a mine re-greening supervision method based on remote sensing image semantic segmentation. BACKGROUND

[0002] With the continuous development of mineral resources, the problem of ecological destruction in mining areas is becoming increasingly serious. Engineering activities such as open stripping, underground subsidence, dump accumulation, and road excavation directly change the surface morphology and soil structure, leading to multiple ecological degradation processes such as vegetation destruction, soil erosion, land degradation, and landscape fragmentation. Therefore, since the end of the last century, China has gradually established a mine ecological restoration system, which is mainly based on "mining while repairing" and "post-mining reclamation". It is clearly stipulated that mining enterprises must fulfill their ecological restoration obligations during mining activities, and the government departments in charge will supervise and inspect. However, in practice, there are the following key technical problems in the current mine ecological restoration work:

[0003] First, the existing supervision and evaluation methods are mostly based on single indicators or limited factors, such as vegetation coverage, soil organic matter, and heavy metal content, which are difficult to fully reflect the ecological system response mechanism of restoration effectiveness. For example, some supervision reports use the improvement of NDVI as the basis for restoration compliance, ignoring the systematic evaluation of structural connectivity, landscape stability, and ecological risk debt, leading to the deviation of supervision data from the actual ecological process. Second, the ecological restoration supervision generally uses qualitative classification or experience weighting model, which lacks rigorous mathematical reasoning basis. The widely used methods such as analytic hierarchy process (AHP), comprehensive index method, and fuzzy comprehensive evaluation method rely on artificial weight setting or subjective interpretation rules, and have the problems of poor model repeatability and weak scale comparability. For example, in the "Land Reclamation Grade Evaluation Standard", the ecological grade is generated through manual scoring and grading, which is simple in form but difficult to adapt to the heterogeneity and dynamic evolution characteristics of ecological structure between different land types. Third, the time dynamics processing is not sufficient. Most existing supervision methods only compare a certain time point or limited two time points, lack of process understanding based on time series data, and cannot identify the hidden trends of "short-term improvement, long-term degradation" or "surface improvement, structural deterioration" in the ecological restoration process, especially in the identification of risks such as rapid vegetation growth but soil restoration, or connectivity decline leading to isolated patches. Fourth, the spatial structure information is not fully utilized. The current mainstream evaluation technology mainly focuses on point index extraction and regional average, ignoring the core role of ecological pattern structure in maintaining ecological function. For example, in the mine slope treatment project, although the vegetation coverage rate is improved, if the landscape fragmentation increases and the ecological network connectivity decreases, it may pose a greater threat to species migration and ecological stability, and the existing evaluation model cannot reveal this spatial structure risk. SUMMARY

[0004] The main purpose of the present application is to provide a mine re-greening supervision method based on remote sensing image semantic segmentation, by constructing a resource and liability bidirectional normalized ecological index matrix, using non-negative matrix decomposition to extract potential ecological function modules, and calculating the resource income and liability residual ratio of the patch level to form a quantifiable resource-liability balance degree; then combining the patch area, structural connectivity and landscape fragmentation, constructing a global supervision evaluation index, realizing the dynamic, fine and structured monitoring and evaluation of the mine ecological restoration process. This method does not need subjective weighting, has mathematical closure and data repeatability, and significantly improves the scientificity, accuracy and engineering practicability of supervision evaluation.

[0005] To solve the above problems, the technical scheme of the present application is as follows:

[0006] The mine re-greening supervision method based on remote sensing image semantic segmentation comprises the following steps:

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

[0008] Step 2: At each monitoring time, the time series matrix is iteratively decomposed by minimizing the K-L divergence to complete the constrained non-negative matrix decomposition, the objective is to decompose the time series matrix into the product of the potential ecological function module weight matrix and the potential ecological function module index contribution matrix, and to keep the potential ecological function module weight matrix and the potential ecological function module index contribution matrix non-negative; for each potential ecological function module, the resource income and liability residual are calculated to obtain the ratio of resource income and liability residual as the resource-liability balance degree;

[0009] Step 3: The resource-liability balance degree is weighted and averaged by the area of each mine landscape patch as the main part of the restoration effect; at the same time, the structural connectivity between each mine landscape patch is taken as an amplification factor, and the landscape fragmentation is taken as a penalty term to be introduced into the same expression, and the global supervision evaluation index is calculated to evaluate the effect of mine ecological restoration.

[0010] Further, the resource index includes soil organic matter density, vegetation coverage and slope stability; the liability index includes heavy metal comprehensive pollution index, annual potential soil and water loss and landscape fragmentation; in order to make the value of the resource index always positive and the value of the liability index always negative, and eliminate the dimensional difference, the maximum value of each original ecological index is obtained first, and then each original ecological index is logarithmically compressed and normalized according to the maximum value.

[0011] Further, in step 2, for each potential ecological function module, the process of calculating its resource benefits and liability residues 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 mine landscape patch, multiply the weight of each potential ecological function module on the patch by the resource-side weight and sum them up to obtain the resource benefits of the mine landscape patch; multiply the index contribution of each potential ecological function module by the liability-side weight and sum them up to obtain the liability residues, and the ratio of the resource benefits to the liability residues is the resource-liability balance degree of the mine landscape patch, which is greater than 1, indicating that the restoration benefits have exceeded the liability residues; if less than 1, it means that the liability residues are still dominant, and the governance efforts need to be increased;

[0012] Further, in time t, each element M p，χ (t) in the time sequence matrix M(t) of the mine landscape patch p is:

[0013]

[0014] Wherein, sgn(x) = +1, when x ∈ {1, 2, 3}, representing the resource column; sgn(x) = -1, when x ∈ {4, 5, 6}, representing the liability column; x = 1, representing that the column is soil organic matter density; x = 2, representing that the column is vegetation coverage; x = 3, representing that the column is slope stability; x = 4, representing that the column is total heavy metal pollution index; x = 5, representing that the column is annual average potential soil and water loss modulus; x = 6, representing that the column is landscape fragmentation; X p，χ (t) is the original ecological index of the χ column corresponding to the mine landscape patch p in the time sequence matrix M(t) at time t; X p′，k (t') is the original ecological index of the χ column corresponding to the mine landscape patch p' in the time sequence matrix M(t) at time t'.

[0015] Further, assuming that the number of mine landscape patches is P, the number of potential ecological function modules is obtained by rounding down the operation of .

[0016] Further, assuming that the potential ecological function module weight matrix is P(t) and the potential ecological function module index contribution matrix is Q(t), the time sequence matrix is iteratively completed by minimizing the K-L divergence through the following formula:

[0017]

[0018] And there is the following equality relationship:

[0019]

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

[0021] Further, when the time series matrix is iteratively completed by minimizing the K-L divergence to constrain the non-negative matrix factorization, P p，k (t) is updated as follows:

[0022]

[0023] where P p，k，new (t) is the updated P p，k (t); P p，k′ (t) represents the weight of the mine landscape patch p on the potential ecological function module k' at time t; Q k′，χ (t) represents the index contribution of the potential ecological function module k' at time t; Q k，χ (t) is updated as follows:

[0024]

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

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

[0027] R k (t) =∑ χ∈{1，2，3} Q k，χ (t);

[0028] The liability residue L k (t) of the potential ecological function module k at time t is:

[0029] L k (t) =∑ χ∈{4，5，6} Q k，χ (t);

[0030] The resource-liability balance degree B p (t) of the mine landscape patch p at time t is:

[0031]

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

[0033]

[0034] wherein, A p is the area of the mine landscape patch p; is the structural connectivity of the mine 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 mine landscape patch p; F FRAG，p (t) is the landscape fragmentation of the mine landscape patch p at time t.

[0035] The mine re-greening supervision method based on remote sensing image semantic segmentation of the present application has the following beneficial effects:

[0036] The method constructs a unified resource and liability ecological index matrix on the technical system, uniformly projects the positive and negative indexes to the same number axis after normalizing them respectively, so that the ecological data of different sources and different physical meanings have comparability, and the problems of dimension confusion and evaluation fragmentation of the existing method are overcome.

[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 influence of various ecological driving factors on different ecological indexes in the restoration process, and avoid the subjective bias caused by artificial weighting and experience scoring.

[0038] Thirdly, in the function interpretation, the resource and liability of each patch are calculated, and a dynamic balance degree index is formed, so that each landscape unit can be quantitatively identified its ecological profit and loss state, and the fine management and control of the management unit are realized.

[0039] Finally, the area, structural connectivity and landscape fragmentation are comprehensively introduced into the global evaluation formula, so that the final supervision evaluation not only reflects the local restoration results, but also includes the spatial network properties and ecological risk information, forms a spatial-structural-functional three-dimensional coordinated evaluation logic, greatly improves the explanatory power and adaptability of the supervision results, and provides a strong data basis and model support for the dynamic supervision and scientific decision of the ecological restoration project. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The method flowchart of the mine re-greening supervision method based on remote sensing image semantic segmentation provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0041] In order to make the person skilled in the art better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor should belong to the protection scope of the present application.

[0042] Reference Figure 1 : A mine re-greening supervision method based on remote sensing image semantic segmentation, the method comprises:

[0043] Step 1: Based on the original ecological indicators of the mine, including resource indicators and liability indicators, a six-column time sequence matrix is established for each mine landscape patch; wherein the elements in the first three columns of the time sequence matrix are resource indicators, constituting the resource column, representing the resource attribute, and the elements in the last three columns are liability indicators, constituting the liability column, representing the liability attribute;

[0044] In the mine re-greening supervision method based on remote sensing image semantic segmentation, the core idea of the first step is to convert the original ecological indicators with spatio-temporal mixing, various types and different dimensions into a dynamic resource-liability information carrier that can not only preserve the characteristics of landscape pattern but also facilitate subsequent computer automatic deduction. This carrier is a time sequence matrix with landscape patches as rows, six representative indicators as columns, and monitoring time as layers. In implementation, first, use unmanned aerial low-altitude visible light and multispectral images, satellite remote sensing data and laser radar generated digital elevation model, through object-oriented multiscale segmentation algorithm and random forest classifier, divide the mined and excavated mountain into landscape patches such as mined-out depression, dump, reclamation terrace, vegetation reconstruction area, drainage ditch and sediment pool, and bind a unique number to each patch. Then, through coordinate unification and geometric correction, align the historical images and digital terrain data to the same projection system, and use the patch tracking algorithm combining area similarity and shortest distance of centroid to ensure the spatial entity of the same number in the multi-temporal sequence to be consistent. Even if the patch merges, splits or disappears, the evolution event is recorded through the time label, so as to lay a stable data skeleton.

[0045] After the patch framework was determined, six landscape pattern indices that could systematically reflect the ecological restoration process of the mine were extracted: soil organic matter density, vegetation cover, and slope stability were defined as resource attributes, which were estimated by field soil sampling, near-infrared laboratory calibration, remote sensing inversion, unmanned aerial vehicle image fine pixel NDVI linear mixing, and high-resolution DEM load-shear-sliding model, respectively; the comprehensive pollution index of heavy metals, potential soil erosion modulus, and landscape fragmentation were classified as liabilities, among which the heavy metal pollution was interpolated by portable X-ray fluorescence spectrometer sampling points, the soil erosion modulus was derived by multi-source RUSLE factors, and the fragmentation index was calculated by FRAGSTATS on the same scale patch boundary density and shape complexity. All index original values were first radiometrically calibrated, cloud and fog removed, noise points smoothed, and multi-temporal missing pixel spatio-temporal Kriging interpolated before being written into the database to improve the spatio-temporal coherence and observation reliability. Due to the large difference in physical dimensions of each index, it is necessary to ensure the clear positive and negative meaning in the subsequent algorithm, so a unified scale normalization process is implemented in the matrix construction stage: for each index, the global extreme value is searched in all patches and all historical monitoring periods, the logarithmic compression is used to weaken the dispersion, and then the maximum value is normalized, so that the resource class index remains positive, and the liability class index is converted to negative by sign flipping, so that the resource- liability symmetric distribution is presented on the number axis, and the calculation bias caused by unit difference is completely eliminated. After normalization, the six columns of indices are written into the matrix in turn, the first three columns are always positive, and the last three columns are always negative, and the matrix row and column do not contain missing values. Each monitoring time will generate a resource- liability matrix, and the multi-temporal matrix is stacked according to the time label order to form a three-dimensional tensor, which lays a unified, traceable, and objective weight-free mathematical foundation for subsequent non-negative matrix factorization and potential function module analysis. When new remote sensing images or field monitoring data enter the system, the automatic pipeline will reuse the same patch segmentation, index inversion, and normalization rules to splice the new observations to the end of the tensor; if there is a large-scale change in the surface form within the monitoring period, the system can renumber or create a row record for the new patch by determining the patch change event, while the historical matrix remains the same and is archived to ensure data integrity and continuity.

[0046] The semantic segmentation process of mine landscape patches from remote sensing images is essentially a cross-platform, multi-source, and multi-temporal intelligent analysis pipeline. Its goal is to interpret complex mine surface into vector patch objects with clear ecological semantics and time-traceable under the premise of making full use of spatial, spectral, texture, and three-dimensional terrain information. The whole process starts from data preparation: first, according to the scope of the mine and the terrain sheltering situation, design a three-layer remote sensing acquisition scheme of satellite-unmanned aerial vehicle-laser radar. The outermost layer uses high-resolution series or WorldView series images with a resolution better than two meters to provide a global base; the middle layer uses tilt photography unmanned aerial vehicle to obtain sub-meter visual-near-infrared stereo images regularly, to capture details such as open pit, dump steps, drainage ditches, etc.; the inner layer uses airborne full-waveform LiDAR to scan once a year to generate centimeter-level DEM and point cloud intensity information, to solve the problem of orthorectification in steep slope shadow and high-difference areas. All raw images are first radiometrically corrected and aerosol removed, and then geometrically refined according to the reference control point network to unify the data accuracy of different sensors and different shooting angles within one pixel. After that, the surface model generated by LiDAR is orthorectified with optical images to ensure that each pixel position has a strict one-to-one correspondence in all phases.

[0047] After image preprocessing, the object-oriented segmentation stage is entered. To avoid over-segmentation or under-segmentation, the system uses a multi-resolution hierarchical segmentation strategy: first, the local variance and edge gradient of the image are used to automatically derive the closed-form analytical scale, which considers both spectral homogeneity and shape compactness, to obtain coarse segmentation; then, superpixel aggregation is used inside each coarse segmentation area to generate small patches with higher regularity. Subsequently, a deep convolutional neural network (such as DeepLabV3+) is used to perform pixel-level semantic segmentation, outputting a probability map for typical mine features such as open pit bare rock, waste rock yard, dump steps, drainage channel, reclamation vegetation, natural forest and grass, bare farmland, building hardening surface, and water sedimentation tank; the system superimposes this probability map with the object-oriented segmentation result, and automatically fine-tunes the segmentation boundary through the maximum probability principle and the boundary approximation degree double threshold, so that the patch boundary shrinks or expands along the real feature boundary, eliminating the jagged edges caused by soft discrimination.

[0048] In the classification stage, the system extracts spectral mean, normalized difference vegetation index, red edge position index, texture entropy of gray level co-occurrence matrix, LiDAR intensity, average slope, surface roughness, shape index, aspect ratio, boundary complexity, and time series mean square deviation of the past three images for each candidate patch, totaling more than thirty dimensions. After three rounds of feature importance evaluation using LightGBM, redundant features with highly overlapping mutual information are removed, and the final input dimension converges to about fifteen dimensions. The classification model uses an ensemble learning framework composed of a random forest and a gradient boosting decision tree: the first layer of models classifies the ground surface into five main categories based on spectral and texture features, and the second layer further subdivides the vegetation reclamation subcategory into three stages: early herbaceous, mid-stage shrub, and late-stage forest. The LiDAR elevation feature is expanded using tensor projection to make the height gradient difference between shrubs and forests explicit in the model. The training samples come from manual vector labeling of unmanned aerial vehicle low-altitude ground truth strips, and are consistent with the point-to-area overlay of RTK high-precision sample lines. The misclassified sample ratio is strictly controlled within two percent. The final classification accuracy is measured by 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 automatically returns to the feature selection stage to adjust the threshold or backtrack the segmentation scale.

[0049] To remove random noise and shape pseudo-patches, the system performs morphological closing operation on adjacent patches of the same type, and then aggregates them according to the spectral Euclidean distance and boundary curvature double thresholds using the region growing algorithm, ensuring that small-scale and thin-shaped non-ecological function patches are absorbed into the main patch. The boundary simplification uses a combination of the Douglas-Peucker algorithm and least squares smoothing to maintain the overall outline of the patch shape while reducing vertex redundancy. The patch vector is then stored in the database, with additional multi-field information: automatically recording patch area, centroid coordinates, main category, secondary category, texture synthesis index, vegetation index average, slope average, and fragmentation measure, for direct calling by subsequent models.

[0050] Temporal alignment is crucial to ensure the consistency of the resource-liability matrix row dimension. The system first calculates the overlapping area ratio matrix and centroid distance matrix between each monitoring period patch set and the previous period, and then finds the maximum area cross mapping based on the two-point matching method. If a pair of patches meets the conditions of an overlap ratio greater than 0.6 or a centroid distance less than twice the minimum resolution and the same category, they are determined to be the same evolution unit and inherit the number. If there is one-to-many or many-to-one, it is recorded as a split or merge event, and an event suffix is added to the patch id. For completely new patches, the serial number is incremented, and the new patch will increase the row in the resource-liability matrix and fill zero at the historical time, ensuring that the matrix dimension increases with time without missing columns. The patch temporal tracking process also combines LiDAR differential DEM to detect height differences. If the classification result shows that the vegetation area becomes bare rock but the height difference does not increase significantly, it is determined to be a classification error and triggers manual review.

[0051] After the above process is completed, the system generates a patch dataset covering all monitoring time points, forming a three-dimensional tensor index; the patch row number and the monitoring time column number are directly mapped to the data structure of the resource-liability matrix, so that subsequent index normalization and matrix decomposition do not require additional space index conversion. The entire remote sensing image semantic segmentation link is managed by an open source remote sensing processing engine orchestrator, and the containers are interconnected between steps. The image inference is supported by GPU acceleration and GIS operation is parallel to CPU. The process can be completed within four hours for a one-time full-automatic processing of a one-hundred-square-kilometer mining area scene, greatly improving the data production efficiency of the supervision project and providing high-precision, spatiotemporally consistent landscape patch basic data for subsequent evaluation methods.

[0052] Step 2: At each monitoring time, the time series matrix is iteratively decomposed by minimizing the K-L divergence, aiming to decompose the time series matrix into the product of the latent ecological function module weight matrix and the latent ecological function module index contribution matrix, and to keep both the latent ecological function module weight matrix and the latent ecological function module index contribution matrix non-negative; for each latent ecological function module, calculate its resource income and liability residue, and obtain the ratio of resource income to liability residue as the resource-liability balance degree;

[0053] The algorithm first determines the number of latent modules automatically according to the dimensions of the matrix, setting it to the square root of the product of the number of patches and the number of indicators, rounded down, thereby avoiding manual subjective parameter setting. Subsequently, by adding a very small positive value to the matrix entries and taking the absolute value of the liability column, the original data containing positive and negative signs is mapped to the non-negative domain, meeting the requirement of the decomposition algorithm for non-negative input; at the same time, the original sign of each entry is recorded, so that the decomposition result can still distinguish between resources and liabilities in the post-processing stage. In the initialization stage, the "patch-module" weight matrix and the "module-index" contribution matrix are assigned values using a random but limited average value distribution, to ensure that the elements of the two matrices are greater than zero and satisfy the overall size constraint. Then enter the iteration link, the system uses the multiplication update rule to gradually adjust the two matrices, so that their product gradually approximates the original matrix in the sense of Kullback-Leibler divergence; in each iteration, the algorithm first calculates the reconstruction error based on the current estimated value, then adjusts the matrix elements in proportion to the error size, and immediately performs a minimum threshold truncation after adjustment to prevent numerical underflow or the appearance of zero entries again.

[0054] To avoid local minimum, the system sets multiple groups of independent random initial values and runs in parallel, synchronously detects the convergence speed and residual change of each group, and finally retains the decomposition result of the group with the smallest divergence and the most stable convergence rate. The entire search process takes the relative change of reconstruction error below the preset threshold or reaching the maximum iteration number as the stopping condition, and outputs the complete trajectory file of error change with iteration number for later review by supervisory technicians. After decomposition, the algorithm first sums up the "module-index" contribution matrix according to the resource column and liability column respectively to obtain the cumulative positive contribution value of each potential module to resource attributes and the cumulative negative contribution value to liability attributes, then element-wise multiplies with the "patch-module" weight matrix and sums up in the module dimension to calculate the resource income and liability residual value corresponding to each landscape patch. Using the ratio of resource income to liability residual of the same patch, the resource-liability balance degree, an intermediate quantity, can be generated. A balance degree greater than one indicates that the resource accumulation of the patch has exceeded the ecological liability at the current time, and a balance degree close to or less than one indicates that the restoration measures still need to be strengthened. At the same time, the system also performs random permutation test and sub-sample cross-validation on the decomposition results, by repeatedly shuffling the matrix rows and columns or extracting non-overlapping monitoring period subsets, to verify the stability and ecological interpretability of the potential modules, ensuring that the extracted functional modules are not accidental noise patterns, but consistent driving factors of the restoration process at different time and spatial scales. The entire processing link is completely data-driven in implementation, does not rely on any empirical weight or artificially set adjustment coefficient, and all algorithm hyperparameters are transparently recorded to the supervisory interface, making it easy for different project teams to directly reuse 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, the subsequent steps can directly call the patch-level balance degree when calculating the area-weighted average and connectivity amplification factor without the need for further sign judgment, significantly improving the efficiency and precision of the supervision evaluation, and providing a solid numerical basis for real-time dynamic review of the mine ecological restoration process.

[0055] Step 3: Weighted average the resource-liability balance degree with the area of each mine landscape patch as the main part of the restoration effectiveness; at the same time, introduce the structural connectivity between landscape patches as an amplification factor and 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 exact area of each patch needs to be extracted, which is directly derived from the same resolution raster or vector boundary, and can be obtained by raster accumulation or polygon measurement. The introduction of area is to ensure that the evaluation results can fully reflect the spatial distribution differences of mine ecological restoration projects: under the condition of equal ecological benefits, the larger patch obviously contributes more to the overall landscape function, so the system integrates the resource-debt balance of all patches by area weighting, thereby constructing the benchmark value of the main part reflecting the restoration effectiveness. After completing the area weighting, the system further focuses on the spatial connection quality between patches, because in a fragmented landscape, even if individual local patches compensate for the debt, the overall ecological function may still be difficult to sustain recovery due to insufficient connectivity. To this end, the system first calculates a distance matrix according to the patch centroid coordinates, then combines the actual obstacle information such as terrain resistance, accumulation body height or road obstruction, and judges which patches can be interconnected in the ecological sense through the shortest path or landscape resistance model, and builds an adjacency graph. In this graph, the system assigns a distance decay weight to each pair of adjacent patches, and then sums up the edge weights of all patches 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 not only the patch restoration condition is good, but also it provides a cross-patch channel for species migration and ecological circulation, which helps to speed up the spread of resources to the surrounding areas.

[0057] In addition to emphasizing positive connectivity, the system also needs to suppress the negative risks brought by excessive fragmentation, so it calls the landscape fragmentation index for each patch, which is derived from the comprehensive calculation of patch boundary complexity and shape index, the higher the more serious the fragmentation. In order to avoid the evaluation results being dominated by local extreme values, the system first performs area-weighted average on the fragmentation, and then introduces a penalty term to offset the positive gain brought by the previous connectivity amplification, ensuring that both beneficial networked patterns between patches and potential ecological risks caused by local over-excavation or vegetation rupture are revealed in time. In this process, the introduction of area, connectivity and fragmentation is based on objective observation data, without any artificial adjustment coefficient or experience threshold; the calculation of connectivity completely depends on the geometric relationship between patches and the resistance between them, while the fragmentation is obtained in batches through a unified landscape pattern software, ensuring the repeatability and cross-time and space comparability of the index. After the area-weighted balance, connectivity amplification factor and fragmentation penalty term of all patches are obtained respectively, the system synthesizes them in a unified scale in the order of first positive accumulation and then negative deduction, generating a comprehensive monitoring index that covers the three meanings of spatial structure, functional benefits and risk compensation. Since this index still maintains a single numerical form, monitoring personnel can directly read its absolute size and adjacent period change amplitude at each monitoring period: when the index continues to rise, it means that the overall resource accumulation steadily covers ecological liabilities and is accompanied by pattern optimization; if the index is stagnant or declining, it suggests that site inspections should be intensified and targeted repair measures should be taken for areas with weak connectivity or increased fragmentation. Through such an evaluation logic that considers both patch internal resource-liability benefits and landscape network structure and fragmentation risks, the system can comprehensively and meticulously reflect the dynamics of mine ecological restoration without the need for subjective weights and empirical factors, providing solid data support for management decisions and subsequent engineering adjustments.

[0058] Further, the resource type index includes soil organic matter density, vegetation coverage and slope stability; the liability type index includes heavy metal comprehensive pollution index, annual average potential soil and water loss amount and landscape fragmentation; in order to make the value of the resource type index always positive and the value of the liability type index always negative, while eliminating the dimensional difference, the maximum value of each original ecological index is first obtained, and then each original ecological index is logarithmically compressed and normalized according to the maximum value.

[0059] Soil organic matter density reflects the abundance of carbon and nitrogen sources for microbial metabolism and vegetation root absorption in the soil layer, and is a key indicator of soil fertility recovery and carbon sink potential. In mining disturbed areas, the original topsoil is often stripped or mixed, leading to significant attenuation of organic matter. Therefore, the increase in organic matter density directly represents the accumulation of ecological functions at the resource end. Vegetation coverage reflects the proportion of the ground surface covered by green vegetation, and can comprehensively describe the intensity of photosynthesis, evapotranspiration rate, and surface temperature regulation effect. The improvement of coverage means that the plant community is more continuous in space, which has a positive effect on suppressing dust, fixing carbon, and providing habitat. Slope stability is used to assess the ability of the profile form 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 roots. High stability means that engineering support and ecological slope protection measures effectively weaken gravity erosion, which is an important safety attribute at the resource side. The comprehensive pollution index of heavy metals is the ratio of the concentrations of multiple elements such as lead, cadmium, and arsenic to the environmental benchmark limit value, which reveals the toxic debt caused by the seepage of waste rock and tailings in the mining area to the soil and groundwater. The larger the index, the higher the potential ecological and food chain risk. The annual potential soil and water loss is derived from the coupling operation of rainfall corrosion, slope length and gradient, vegetation coverage, and soil erodibility. The higher the value, the more likely it is that bare or loose areas will be eroded by heavy rain and transport sediment and nutrients downstream, which is a core indicator reflecting the debt of surface erosion. Landscape fragmentation measures the spatial continuity of the landscape by counting the number of patches, the length-area relationship, and the shape complexity. The larger the value, the more fragmented the original continuous habitat is cut by mining sites, roads, and dumping sites, hindering species migration and gene flow, thus constituting a structural debt. These six indicators cover key processes in soil quality, vegetation condition, topographic safety, chemical pollution, physical erosion, and spatial pattern. The increase in resource indicators or the decrease in debt indicators can objectively reflect the gradual recovery of ecological service functions and the gradual reduction of environmental pressure in the process of mine ecological restoration, so they are selected as the column vector base of the resource-debt matrix of this method.

[0060] The system first performs logarithmic compression on the original indicators, using the monotonic increasing property of the logarithmic function to reduce the stretching effect of outliers on the overall distribution, so that resource and liability data remain resolved in the low magnitude area and moderately convergent in the high magnitude area. Then, using the column-level maximum value cached earlier as the denominator, the system normalizes each entry to a dimensionless interval between zero and one. Since resources and liabilities are naturally opposite in sign, the system immediately performs a sign flip on the heavy metal pollution index, the annual average potential soil erosion, and the landscape fragmentation after completing the normalization, multiplying them by negative one to make all their values fall into the negative interval, while the soil organic matter density, vegetation cover, and slope stability remain positive. In this way, resource indicators are always positive, liability indicators are always negative, and all six columns are on the same scale, providing clean data input for subsequent matrix decomposition based on the resource-liability concept. The entire process does not involve any artificial empirical coefficients: the maximum value used for normalization is directly derived from objective observation, the base of logarithmic compression is a system default constant, and the sign flip is a fixed operation, ensuring consistent processing links between different monitoring periods. When new remote sensing images or measured data enter the system, if a new extreme value is found that exceeds the historical record, the normalization constant is immediately replaced and incremental recalculation is triggered, ensuring that all time period data remains comparable under the latest benchmark. At the same time, the system writes the extreme value replacement record to the log, allowing monitoring personnel to trace the changes in indicator scale and their impact on the comprehensive evaluation value.

[0061] Further, in step 2, for each potential ecological function module, the process of calculating its resource income and liability residue 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 mine landscape patch, multiply the weight of each potential ecological function module on the patch by the resource-side weight and sum them up to obtain the resource income of the mine landscape patch; multiply the indicator contribution of each potential ecological function module by the liability-side weight and sum them up to obtain the liability residue, and the ratio of resource income to liability residue is the resource-liability balance of the mine landscape patch,

[0062] When it is greater than 1, it means that the restoration income has exceeded the liability residue; if it is less than 1, it means that the liability residue is still dominant and needs to be increased.

[0063] In the remote sensing image-based semantic segmentation method for mine re-greening supervision, the core task of the second stage is to map the potential ecological function modules that have been decomposed back to the specific landscape patch level, so as to quantify the resource surplus and debt residue of each patch at the current time. The implementation process first needs to read two result matrices of the same dimension: one is called the module index contribution matrix, which records the interpretation intensity of each potential ecological function module on the six landscape pattern indexes; the other is called the patch module weight matrix, which records the participation weight of each landscape patch on each potential module. The system divides the module index contribution matrix into resource column and debt column according to the index attributes, then performs row-wise accumulation on the resource column to generate a resource-side weight for each potential module; similarly, performing the same row-wise accumulation on the debt column can obtain the paired debt-side weight. At this time, the potential module is assigned two clear and independent numerical labels, one representing the average contribution of the module in improving ecological resources, and the other representing the average contribution of the module in preserving ecological debt. Then the system enters the patch-level calculation cycle. For a single landscape patch, the program traverses all potential modules, multiplies the patch's weight on a module with the corresponding resource-side weight and accumulates it into the resource income accumulator; simultaneously, it multiplies the patch's index contribution on the module with the debt-side weight and accumulates it into the debt residue accumulator. At the end of the traversal, the value in the resource income accumulator represents the positive ecological benefit obtained by the patch at the current time through ecological restoration engineering, while the debt residue accumulator represents the negative ecological pressure that has not been offset by engineering measures. In order to provide directly interpretable dimension-unified indicators for supervisors, the system performs a ratio operation on the two to obtain the resource-debt balance degree; a ratio greater than one indicates that the resource-side income has substantially covered the debt-side residue, indicating that the patch is in a net restoration state; if the ratio is close to or less than one, it indicates that the debt still dominates and the treatment intensity should be increased. The entire cycle only needs to complete two multiplication and addition operations of the size of the potential module number and one division operation for each patch, and the calculation complexity grows linearly with the total number of patches. The system can complete the global update of large-scale mining areas in minutes and write the results to the supervision database in real time. In order to avoid numerical instability caused by extremely small denominators, the program automatically checks whether the debt residue accumulator is below the machine precision threshold before entering the ratio operation, and replaces it with the threshold if it is below the threshold, ensuring that the ratio is always definable. All input parameters come from objective observation data and inherent output of the decomposition algorithm, and 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 the balance degree of all patches is calculated, the system will compare the latest balance degree of each patch with the historical monitoring sequence, automatically draw the resource-liability trajectory over time, and analyze the trend of the trajectory slope. If the balance degree of an individual patch is found to be declining for consecutive periods or hovering around the critical value for a long time, the monitoring platform will push warning information to the ecological engineering management end, requiring the on-site team to conduct targeted investigation or reinforcement for the patch.

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

[0065]

[0066] wherein sgn(x) = +1 when x e {1, 2, 3}, representing the resource column; sgn(x) = -1 when x e {4, 5, 6}, representing the liability column; x = 1, representing that the column is soil organic matter density; x = 2, representing that the column is vegetation coverage; x = 3, representing that the column is slope stability; x = 4, representing that the column is total heavy metal pollution index; x = 5, representing that the column is annual average potential soil erosion modulus; x = 6, representing that the column is landscape fragmentation; x p，χ (t) is the original ecological index of the x column corresponding to the time series matrix M(t) of the mine landscape patch p at time t; x p′，χ (t') is the original ecological index of the x column corresponding to the time series matrix M(t) of the mine landscape patch p' at time t'.

[0067] Firstly, the resource column {χ = 1, 2, 3} is fixed as positive by sgn(χ), and the liability column {χ = 4, 5, 6} is fixed as negative, so that the resource indicators (i.e., soil organic matter density, vegetation coverage, slope stability) always increase in the positive direction on the number axis, while the liability indicators (i.e., total heavy metal pollution index, annual average potential water and soil loss modulus, landscape fragmentation) decrease in the negative direction on the number axis. This artificially introduced symbol opposition ensures that resources and liabilities can be directly separated by positive and negative positions during subsequent matrix decomposition, while avoiding logical confusion caused by symbol flipping of the same column data at different time periods. Secondly, the numerator uses the ratio of the measured value to the maximum value in the whole space-time, which means that any indicator will reach a value of one at its historical peak, and a relative amount less than one at other space-time points. In this way, heterogeneous indicators such as concentration, percentage, and stability coefficient, which have different dimensions, are all pulled back onto a relative scale line. The denominator selects the global extreme value rather than the local extreme value, ensuring that the new and old data have continuous comparability when cross-year monitoring. Thirdly, the logarithmic compression of ln(1 + ·) not only overcomes the stretching of extreme high values on the distribution, but also retains the resolution of small-scale differences for ecological assessment. The addition operation ensures that the logarithm is still defined when the measured value is zero, avoiding the numerical missing of bare rock patches or extremely low pollution patches due to zero values. In summary, this transformation makes resource indicators show gradual growth from zero to the positive upper limit during the restoration and promotion process, while liability indicators show gradual convergence from the negative upper limit to zero during the governance process. Since the maximum amplitude on both sides is limited by the same function, resource gain and liability reduction can be directly compared by absolute value in subsequent calculations, thereby achieving the non-dimensionalization of the core quantity of resource-liability balance degree. More importantly, the only scaling constant comes from objective observation and does not contain empirical factors, ensuring that the evaluation results of different mining areas, different restoration stages, and even different monitoring technical conditions can be aligned under the same theoretical framework. When new extreme values appear during monitoring, the system only needs to automatically update the constant and recalculate the corresponding column to extend the time series depth without damaging the logic of existing data. Because of this, matrix M(t) can achieve scale normalization and dynamic comparability of six-dimensional landscape pattern indices while maintaining the symbol distinction between resources and liabilities, laying a rigorous, stable, and objective data foundation for subsequent non-negative matrix factorization based on Kullback-Leibler divergence and accurate calculation of resource-liability balance degree.

[0068] Further, assuming the number of mine landscape patches is P, the number of potential ecological function modules is obtained by rounding down .

[0069] In this method, the number of potential ecological function modules K is obtained by rounding down The empirical formula is determined, and the design idea is to keep the hidden dimension and the observation dimension in the approximate square root level of information balance: on the one hand, the number of columns of the time series matrix is fixed as 6 landscape pattern indexes, so the product of the overall observation dimension and the number of patches P can be approximately regarded as the “total degree of freedom” that needs to be explained in decomposition; on the other hand, if P or 6 is directly taken as the number of potential modules, not only will there be serious dimension redundancy, but also the non-negative matrix decomposition will fall into the dilemma of sparse solution dilution or overfitting; after multiplying the two and then taking the square root and rounding down, it is equivalent to compromising the “row information amount” and “column information amount” to the same order of magnitude in the geometric sense, so that each potential module does not bear too much or too little explanation load. In the context of mine ecological restoration, this compromise strategy can ensure that the decomposition model has enough capacity to capture the differentiated ecological function patterns between patches, and avoid introducing additional noise due to too many potential modules, which leads to unnecessary numerical diffusion in subsequent resource benefit and liability residue calculation; at the same time, rounding down ensures that the result is an integer, keeping the calculation complexity and storage requirements within a controllable range, and as the monitoring network captures more new patches (i.e. P grows), the number of potential modules will also expand at a power speed, so that the model capacity automatically adapts to the demand for ecological pattern refinement in long-term monitoring, ensuring the robustness and scalability of the evaluation system.

[0070] Further, let the potential ecological function module weight matrix be P(t), and the potential ecological function module index contribution matrix be Q(t), then the time series matrix is iteratively completed by minimizing the K-L divergence as follows:

[0071]

[0072] And there is the following equation relationship:

[0073]

[0074] Where, P p，k (t) represents the weight of mine landscape patch p in potential ecological function module k at time t; Q k，χ (t) represents the index 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 processed by symbol normalization and logarithmic normalization into two completely non-negative and ecologically interpretable low-rank factors, so as to extract a set of implicit ecological function driving patterns from the massive patch-index observations, and quantify and track the dynamic contribution of these patterns to resource accumulation and liability decay. For this purpose, this method uses the multiplicative update NMF framework with Kullback-Leibler divergence as the loss function and non-negative constraint as the boundary condition, which is formally written as Furthermore, it is required that the approximate reconstruction equation is satisfied at any time t.

[0076] Unlike traditional NMF that uses Euclidean or Frobenius norms, KL divergence... With a natural relative entropy interpretation, it can quantify the difference in information distribution between two non-negative matrices. When the elements of M(t) exhibit a long-tailed distribution in the low-to-high value range, and the magnitudes of the resource and debt columns differ by tens of times, the KL metric is more effective than the squared error in penalizing the neglect of sparse peaks, thus maintaining high sensitivity even in the extreme case of "extremely low resources and extremely high debt" in the early stages of repair. To ensure that the number of potential functional modules is balanced with the observation dimension, and to avoid overfitting caused by the dimensionality curse, 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 ecological process dimension. 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 potential space; rounding down ensures that the number of factors is an integer and does not exceed the observation degrees of freedom.

[0077] In the specific solution, firstly... and Assign a random positive initial value proportional to the row mean to reduce gradient oscillations; then apply the classic Lee and Seung multiplication update rule iteratively in the sense of KL divergence: update P p，k When (t), M p，χ (t) / (P(t)Q(t)) p，χ As a gain coefficient and the existing Q k，χ Multiply (t) and then normalize to update Q. k，χ When (t), the same gain coefficient is applied to the current P. p，k (t) Multiplication and normalization; to avoid division by zero and to maintain numerical stability, a minimum machine value 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 decrease rate is less than 10⁻⁵ or the maximum number of iterations is 500, and uses ten sets of independent random initial values ​​for parallel solution, taking the set with the smallest KL residual and the smoothest convergence curve as the final result. The obtained P(t) row vector can be regarded as the probability density of patch affiliation in each functional module, and the sum of the elements is approximately the complete weight of the ecological function of the patch; the obtained Q(t) column vector describes the weighted contribution of each module to the six indicators, where the sum of the contributions to the resource column (χ = 1, 2, 3) is the resource-side weight of the module, and the sum of the contributions to the liability column (χ = 4, 5, 6) is the liability-side weight of the module. To map these potential quantities back to actionable engineering semantics, the system performs resource revenue and liability residual estimation at the patch dimension: for each patch p, its weight P on module k is calculated. p，kThe resource gain is obtained by multiplying the corresponding module's resource-side weight by (t) and summing over all k; the debt residue is obtained by multiplying the weight of the same patch on module k by the module's liability-side weight and summing the sum; the ratio of the two is the resource-liability balance. Since P(t) and Q(t) are all non-negative and the weights on both the resource and liability sides are generated by explicit summation in the row or column direction, the resource gain and debt residue are naturally on the same dimension and can be directly compared. This not only avoids subjective coefficients but also allows each step of the calculation to be traced back to the physical interpretation of the matrix elements: if a module is determined to be "vegetation regeneration dominant", it has a high contribution in the resource column and a low contribution in the liability column; if a module corresponds to "heavy metal enrichment", it has a high contribution in the liability column. After the iteration converges, the system simultaneously outputs the whole-period KL residual curve, the sparsity index of the two factor matrices, and the module-index contribution heatmap for the supervision experts to perform ecological semantic annotation; if some modules cannot find a reasonable ecological correspondence, the K-rerun decomposition can be increased, or new observation index columns can be introduced into the backend database to expand the dimensions. It is worth emphasizing that when a 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 a warm-start for a small number of iterations to complete the online update, which greatly reduces the refactoring overhead. If P surges due to the addition of new patches, the system will automatically recalculate. The old factor matrix is ​​expanded with zero padding to ensure that the model capacity and spatial granularity grow in sync.

[0078] Furthermore, when performing constrained nonnegative matrix decomposition on the time series matrix by iteratively minimizing the KL divergence, P p，k The update formula for (t) is:

[0079]

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

[0081]

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

[0083] The first formula updates P by assigning weights to the potential ecological function module dimension of the reconstructed residual of each mining landscape patch p. p，k(t), specifically by taking the ratio of the observed value M p，χ (t) to the current reconstruction value ∑ k′ P p，k′ (t)Q k′，χ (t) as the gain coefficient, so that those index entries which have not been sufficiently explained by the reconstruction are given a larger weight increase in the module k; then the contribution of the module to each index Q k，χ (t) is weighted by the weight P p，k (t) to obtain the new value P p，k，new (t), while a 10-12 is added to the denominator to prevent division by zero and to ensure non-negativity. The second equation updates the index contribution matrix Q k，χ (t) in the dual: it takes the weight P p，k (t) of all patches in module k as the weight, and averages the ratio of the observed value to the reconstructed value for each index, thus absorbing the unexplained residuals in the index dimension, and then multiplies it with the old value to obtain Q k，χ，new (t); here again a small constant is added to the denominator to ensure that the updated elements are always greater than zero.

[0084] When the two updates are alternated, the product of P(t) and Q(t) is continuously contracted 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, all subsequent iterations remain non-negative, so that the weights and contributions of the ecological function modules always correspond to the positive explainable amount in value. In the context of mine ecological restoration, this update mechanism means that each iteration will allocate more explanation ability to the function modules that can simultaneously reduce heavy metal pollution, reduce soil erosion, and improve vegetation and soil quality, and will automatically weaken the weights of modules that are insufficiently explained by the liabilities; when the iteration converges, P p，k (t) will focus on a small number of modules with strong resource contribution, high connectivity potential, or effective erosion control characteristics, while Q k，X (t) will present a typical positive restoration pattern with high resource columns and low liability columns, or a liability-dominated stress pattern. Since the relative error between the actual observation and the model prediction is essentially the error feedback of the algorithm to each data record, the algorithm can still converge stably in the case of fluctuations in remote monitoring image quality or local field measurement 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 total weight conservation inside the module and inside the patch, so that the updated matrix will not have a numerical explosion of infinite gain. Through this strict iterative logic, the invention not only mathematically ensures the monotonic decrease of KL divergence, but also economically uses the multiplication update to avoid the explicit step size adjustment of the negative gradient, and makes every weight adjustment in the ecological sense be interpreted as "reallocate the unexplained repair effect or residual pressure" to the most matched potential functional module, greatly improving the model's explanation transparency and the operability of regulatory decision-making.

[0086] Further, at time t, the resource benefit R k (t) of the potential ecological functional module k is:

[0087] R k (t) = ∑ χ∈{1，2，3} Q k，χ (t);

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

[0089] L k (t) = ∑ X∈{4,5,6} Q k，X (t);

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

[0091]

[0092] When the contribution value of X∈{1,2,3} is significantly greater than X∈{4,5,6}, it indicates that the ecological mechanism of module k is more inclined to positive repair function, such as efficient vegetation regeneration or landscape stabilization measures; on the contrary, if the liability side contribution is higher than the resource side, it indicates that the module captures more stress processes such as heavy metal enrichment, erosion intensification or landscape cutting. Next, the resource- liability balance degree The participation weight P p，k (t) of the patch p on all potential functional modules is taken as the projection coefficient, and the resource benefit R k (t) and the liability residue L k (t) of the module level are respectively mapped back to the patch scale: the numerator is the weighted superposition of the patch to all positive functions, and the denominator is the total ecological pressure borne by the patch; because the same weight vector acts on the two types of total amount, the balance degree value becomes a dimensionless ratio to measure the profit and loss. When B p (t) > 1, the resource benefit accumulated by the patch exceeds the remaining liability, indicating that the improvement of soil improvement, vegetation repair, slope reinforcement and other management measures has achieved a stage of net income in this position; if Bp (t)≤1, it indicates that the liability is still dominant, and it is necessary to continue to put in the passivation material, add the drainage ditch or optimize the vegetation configuration. The small constant 10-12 is added to the denominator to prevent the division by zero error when the negative liability tends to zero in the numerical value, and it will not have a perceptible impact on the interpretation logic, because in the extreme case of repairing the almost completely eliminated liability, the balance degree will be almost infinite, which just prompts the supervisor that the patch can enter the maintenance period rather than the key management period. In this way, P p，k (t), R k (t) and L k (t) together convert the high-dimensional complex ecological pattern information into a real-time index that can be interpreted for a single block landscape, not only mathematically inheriting the precise reconstruction of the observation distribution by non-negative matrix factorization, but also providing clear and executable zoning management basis for the supervision system in engineering: the upward inclination of the balance degree curve indicates that the repair measures are accelerating the effect, and the downward inclination indicates that the management intensity or method needs to be adjusted; the spatial distribution heat map of multiple patches can highlight the risk aggregation area at a glance, laying a reliable underlying data foundation for the subsequent area weighted and connectivity amplification steps, so as to ensure the closure and self-consistency of the entire mine ecological restoration supervision and evaluation chain in theory and practice.

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

[0094]

[0095] Wherein, A p is the area of the mine landscape patch p; is the structural connectivity of the mine 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 mine landscape patch p; F FRAG，p (t) is the landscape fragmentation degree of the mine landscape patch p at time t.

[0096] The area A p as an objective scale of geographical entity appears symmetrically in the numerator and denominator, which on the one hand provides a weighting basis for the resource-liability balance degree B p (t) and the structural connectivity amplification factor, so that the larger the area of the patch, the more significant the total contribution; on the other hand, by summing the areas of all patches and placing them in the denominator, the global normalization is realized, so that the value range of E(t) is not distorted due to scale differences when comparing across the mine area. Secondly, the patch-level balance degree B p(t) The resource benefit and liability residual are compressed into a dimensionless ratio, representing the local repair effectiveness. If the value is greater than one, it means net benefit, and if it is less than one, it means net liability. This index reflects the logic of "high yield large patch" priority to improve the total score under the area weight.

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

[0098] The adjacency matrix is combined with distance decay, which automatically considers patch area and spatial resistance in the connected edge weight, ensuring that the calculation result is positively correlated with the actual seed diffusion, animal migration path or hydrological channel length. Then, the global average connectivity is divided by the ratio, which maps the connection advantages or disadvantages of each patch to a relative quantity, forming a multiplicative amplification factor after " + 1" translation: when , the factor is 2, representing that the connectivity reaches the average and can obtain the baseline enhancement; if a patch is isolated, C p (t) Close to zero, the amplification factor tends to 1; if a patch is located in an ecological corridor node, C p (t) Much higher than the average, exponential growth is suppressed by sub-linear, which acknowledges its network contribution and avoids numerical imbalance beyond the resource / liability semantics; the denominator adds 10 -12 is only for numerical stability, much smaller than the connectivity real value, and does not affect the ecological interpretation.

[0099] The last part As a whole penalty term deducted from the total score, it does an area-weighted average of the square root of the overall landscape fragmentation in the mining area. The square root function makes the fragmentation risk and penalty marginal and decreasing: when the fragmentation is low, the new cut significantly reduces the index, urging supervision and timely intervention; when the fragmentation is already very high, the further damage of new fragmentation to the ecological network is decreasing due to the denominator, avoiding negative infinity of the index and maintaining comparability; area weighting also ensures that large high-fragmented patches contribute most to the penalty. Overall, the formula will make the local function index B pE(t) maps area and connectivity to the macro-level of the landscape, and then uses fragmentation risk for overall correction. Mathematically, it achieves a three-stage coupling of addition, multiplication, and subtraction: addition represents the normalization of area summation satisfying mass conservation; multiplication reflects the amplification of positive benefits by connectivity; and subtraction reflects the suppression of system stability by fragmentation. Numerically, since all sub-items are derived from measurable or calculable data and have no empirical coefficients, E(t) has a consistent scale over time and is portable when comparing different mining areas or different governance schemes within the same mining area. Its ecological implication is that when most area weights in a mining area are concentrated in patches with high balance, high connectivity, and low fragmentation, E(t) achieves a positive and large value, indicating that the current restoration strategy not only improves ecosystem services locally but also optimizes the network spatially. If a decline in connectivity advantage or an increase in fragmentation is monitored, even if the balance temporarily improves, the total score may decrease, thus suggesting that restoration projects should shift from single-point governance to structural measures such as corridor opening and gentle slope reclamation.

[0100] As the monitoring sequence lengthens, the slope of the exponential curve can be used to evaluate the overall restoration speed, and the fluctuation width can be used to evaluate the system stability. In the management platform, when the E(t) value is higher than the set warning threshold, the system automatically identifies it as a "safe zone," between the two thresholds it is an "observation zone," and below the lower threshold it triggers a "risk zone" alarm. Based on the contribution ratio of connectivity and fragmentation, targeted suggestions are given, such as replanting connecting corridors, reducing slope cutting, or implementing fragmented habitat reconnection projects. Since each data stream in the formula is extracted from remote sensing grids, GIS topology, or the output of the previous iteration using automated scripts, the entire evaluation chain can run within minutes after the monitoring data is synchronously uploaded and update the management panel in real time. This enables mine ecological restoration supervision to move from traditional static discrete sampling to dynamic continuous digital twins. The ecological balance, spatial pattern, and risk pressure can be accurately perceived through the E(t) value, providing decision-makers with a refined control basis based on big data.

[0101] Suppose that only four landscape patches are divided in the mining area, numbered p = 1 to 4, and the corresponding measured original ecological indicators are as follows: soil organic matter density [X] p，1 (t1)] 8.00 g·kg respectively -1 , 4.00, 1.00, 6.00, Vegetation Cover [X] p，2 (t1)] respectively 70.0%, 35.0%, 15.0%, 55.0%, slope stability coefficient [X p，3 (t1)] were 0.85, 0.45, 0.30, and 0.65 respectively, and the total heavy metal pollution index [X] p，4 (t1)] are 2.00, 5.00, 10.0, and 3.00 respectively, representing 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) 0.30, 0.50, 0.80, 0.40. Since there is only one monitoring period, the spatio-temporal maximum of the six columns is simply the largest observed value in the column: maxX χ = {8.00, 70.0, 0.85, 10.0, 1500, 0.80}.

[0102] Substitute the normalization and sign adjustment formulas

[0103]

[0104] The four numerical vectors M 1：4 (t1) are obtained:

[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, rather than actually iterating, we give a feasible decomposition result directly (satisfying the non-negative constraints and the product row-column approximation M p，χ ), where the patch-module weight matrix is:

[0109]

[0110] The module-index contribution matrix is:

[0111]

[0112] The resource contribution of the three columns and the liability contribution of the three columns are summed by row to obtain the resource income R k (t1) and the liability residue L k (t1) of each module:

[0113]

[0114] The resource income and liability residue are then projected to the resource side and the liability side by the patch weights in each module to obtain the resource-liability balance:

[0115]

[0116] The calculation results are:

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

[0118] Area is calculated by remote sensing: {A1, A2, A3, A4} = {3.0, 2.0, 4.0, 1.0} hm 2 The patch centroid connectivity is calculated by using the adjacency matrix A p，q = 1 only for pairs (1, 2), (1, 4), (2, 3), (3, 4) are adjacent, and zero otherwise, and the distance matrix is taken as d 1，2 = 150 m, d 1，4 = 100 m, d 2，3 = 120 m, d 3，4 = 180 m, the average distance for all adjacent pairs Thus:

[0119]

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

[0121] The landscape fragmentation is directly calculated by the original measure {F FRAG，p (t1)} = {0.30, 0.50, 0.80, 0.40}. Now the global monitoring evaluation index can be calculated. First, the area-normalized positive molecule is calculated:

[0122] ∑ p A p = 10.0;

[0123]

[0124] Divided by the total area, we get 3.03. Then the fragmentation penalty term is calculated:

[0125]

[0126] Therefore, the final index is:

[0127] E(t1) = 3.03 - 0.74 ≈ 2.29.

[0128] Since E(t1) is obviously greater than zero and above the preset threshold value 1.0, it means that the overall global restoration benefit at monitoring time t1 has exceeded the ecological debt, and it is greatly strengthened by the connectivity gain; however, the local balance degree of patch 3 is only 0.40, and its fragmentation is the highest, which is still the core area of ecological pressure that has not been resolved, and measures such as heavy metal passivation, slope shielding, and corridor reconnection need to be prioritized in the next step of governance plan. The entire process from original index collection to E(t) output can be completed within a few minutes in the case of using script automation execution, and can be automatically updated incrementally with the arrival of the next period of remote sensing or ground monitoring data, realizing the real-time digital twin closed loop of mine ecological restoration monitoring.

[0129] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalent replacements; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1.A mine re-greening supervision method based on remote sensing image semantic segmentation, characterized in that, 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; wherein the elements in the first three columns of the time series matrix are resource indicators, constituting the resource column, representing the resource attribute, and the elements in the last three columns are liability indicators, constituting the liability column, representing the liability attribute; Step 2: at each monitoring time, the time series matrix is iteratively decomposed by minimizing the K-L divergence, the objective is to decompose the time series matrix into the product of the latent ecological function module weight matrix and the latent ecological function module indicator contribution matrix, and make the latent ecological function module weight matrix and the latent ecological function module indicator contribution matrix non-negative; for each latent ecological function module, the resource income and liability residue are calculated, and the ratio of the resource income and the liability residue is obtained as the resource- liability balance degree; Step 3: the area of each mine landscape patch is used to weight average the resource- liability balance degree as the main part of the restoration effect; at the same time, the structural connectivity between the mine landscape patches is taken as the amplification factor, and the landscape fragmentation is taken as the penalty term to be introduced into the same expression, and the global supervision evaluation index is calculated to evaluate the effect of mine ecological restoration; The resource indicators include: soil organic matter density, vegetation coverage 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 the resource indicators always positive and the values of the liability indicators always negative, and eliminate the dimensional difference, the maximum value of each original ecological indicator is obtained, and then each original ecological indicator is logarithmically compressed and normalized according to the maximum value; At time each element of the time series matrix of mine landscape patches is: ； wherein, , represents a resource column; , represents a liability column; , represents a column for soil organic matter density; , represents a column for vegetation coverage; , represents a column for slope stability; , represents a column for total heavy metal pollution index; , represents a column for annual average potential water and soil loss modulus; , represents a column for landscape fragmentation; is the original ecological index of the column corresponding to the mine landscape patch in the time series matrix at time ; is the original ecological index of the column corresponding to the mine landscape patch in the time series matrix at time ; is the original ecological index of the column corresponding to the mine landscape patch in the time series matrix at time ; is the original ecological index of the column corresponding to the mine landscape patch in the time series matrix at time ; is the original ecological index of the column corresponding to the mine landscape patch in the time series matrix at time ; is the original ecological index of the column corresponding to the mine landscape patch in the time series matrix at time ; The number of the mine landscape patches is The number of the potential ecological function modules is obtained by rounding down the operation of Let the potential ecological function module weight matrix be , and the potential ecological function module index contribution matrix be The time series matrix is iteratively completed by minimizing the K-L divergence through the following formula: ； And there is the following equation relationship: ； wherein, denotes the weight of the mine landscape patch at time on the potential ecological function module ; denotes the index contribution of the potential ecological function module at time ; At time the potential eco-functional module gains resource benefits : ； At time the potential eco-functional module has a debt residue of: ； In time At that time, mining landscape patches Resource-liability balance for: ； Global supervision evaluation index Is: ； wherein, is the area of the mine landscape patch ; is the structural connectivity of the mine landscape patch , is the adjacency matrix, is the centroid distance of the mine landscape patch , is the average distance of all adjacent pairs; is the global average connectivity of the mine landscape patch ; is the landscape fragmentation of the mine landscape patch at time . 2.The mine re-greening supervision method based on remote sensing image semantic segmentation according to claim 1, wherein, In step 2, for each latent ecological function module, the process of calculating the resource income and the liability residue includes: adding the cumulative contribution of each latent ecological function module to the resource column to obtain the resource side weight of the latent ecological function module; then adding the cumulative contribution of the same latent ecological function module to the liability column to obtain the liability side weight; for each mine landscape patch, the weight of the mine landscape patch on each latent ecological function module is multiplied by the resource side weight to obtain the resource income of the mine landscape patch; the index contribution of each latent ecological function module is multiplied by the liability side weight to obtain the liability residue, and the ratio of the resource income and the liability residue is the resource- liability balance degree of the mine landscape patch, when it is greater than 1, it means that the restoration income has exceeded the liability residue; if it is less than 1, it means that the liability residue is still dominant, and the treatment intensity needs to be increased. 3.The mine re-greening supervision method based on remote sensing image semantic segmentation of claim 1, wherein, When the time series matrix is iteratively completed by minimizing the K-L divergence to constraint non-negative matrix factorization, The update formula is: ； wherein, is the updated ; denotes the weight of the mine landscape patch on the potential ecological function module at time ; denotes the index contribution of the potential ecological function module at time ; The update formula of is: ； wherein is the updated .

Images