A multi-evidence fusion type remote sensing heavy metal pollution risk index construction method

Abstract

This invention proposes a multi-evidence fusion-based remote sensing method for constructing a heavy metal pollution risk index, belonging to the field of surveying and mapping engineering. It addresses the problem that existing models lack training samples, making it difficult to obtain reliable spatial risk results and hindering low-cost identification of investigation areas. The method includes: acquiring the heavy metal content and coordinates of soil sampling points in mining areas; calculating the Nemerow Comprehensive Pollution Index (PN) based on the content and background values; classifying risk sampling points using a PN threshold; constructing a multi-evidence chain risk feature set based on the pollution generation, migration, accumulation, and response processes; when sampling data is available, training the model using the feature set as input and the PN risk state as the target, outputting the warning risk probability as the sample-available index (RWI); when sampling data is unavailable, unifying and normalizing the risk direction of the feature set, and performing intra-layer and inter-layer weighted fusion to obtain the sample-free index (RWIn); finally, outputting the results according to the index classification and zoning. This method is applicable to the field of soil environmental monitoring.

Description

Technical Field

[0001] This invention belongs to the field of surveying and mapping engineering, and in particular relates to a method for constructing a multi-evidence fusion remote sensing heavy metal pollution risk index. Background Technology

[0002] In the fields of soil environmental monitoring and mining area ecological restoration, the identification and assessment of heavy metal pollution risks are based on ground sampling and physicochemical analysis results. Current technologies typically employ single-factor pollution index methods, the Nemerow composite pollution index method, or the potential ecological risk index method to assess soil quality in mining areas. These methods are based on laboratory-measured data of heavy metal contents such as chromium, nickel, copper, zinc, cadmium, lead, arsenic, and mercury, comparing them to national soil environmental quality standards or regional background values ​​to calculate the degree of pollution. For example, the Nemerow composite pollution index is widely used in site investigations because it can account for both extreme values ​​of single factors and average pollution levels, while the potential ecological risk index further incorporates a toxicity coefficient to quantify ecological hazards. However, these traditional assessment methods are essentially based on static determinations of discrete sampling points; the output pollution level or index value only represents the state of the sampling location and cannot fully reflect the continuous spatial distribution and gradual change patterns of pollution. In mining areas with complex topography and strong human interference, the process of pollutants migrating with surface runoff and accumulating with topography has obvious spatial heterogeneity. Relying solely on interpolation of limited sampling point data can easily affect the results due to sampling density, point layout, and interpolation methods, which may be insufficient to support risk zoning and spatial warning at the whole-area scale.

[0003] More importantly, existing evaluation systems have limitations in their ability to integrate multi-source spatial information. Traditional remote sensing methods for retrieving heavy metals from soil often directly extract vegetation indices, surface temperature, or reflectance in certain bands as feature variables, or simply overlay auxiliary factors such as topography and land use into the evaluation model. This approach treats pollution source input, pollutant migration processes, soil adsorption capacity, and surface ecological response as independent, parallel variables, affecting the complete geochemical process chain of heavy metals from anthropogenic emissions to environmental media accumulation and then to surface spectral response. For example, pollution diffusion caused by mining activities is constrained by topographic slope and hydrological conditions; the content of clay minerals and iron oxides in the soil determines the adsorption and fixation capacity of heavy metals; and these underground accumulation characteristics further affect surface spectral absorption characteristics and vegetation stress performance. Due to the lack of consideration for the mechanisms linking these processes, existing methods struggle to clearly explain the driving mechanisms of risk formation, resulting in limited generalization ability and insufficient spatial interpretability of evaluation models in complex mining environments.

[0004] Furthermore, the strong reliance of existing technologies on ground sampling data constitutes a significant technical obstacle in practical applications. In the initial stages of environmental surveys in mining areas or in large-scale general surveys, there are often situations where sampling points are sparse or even no actual samples are available. Existing evaluation models, lacking sufficient training samples or calibration benchmarks, cannot output reliable spatially continuous risk results, making it difficult for management departments to quickly identify key investigation areas at low cost. Although the industry has attempted to compensate for data gaps by increasing sampling density, this undoubtedly significantly increases survey costs and timelines, and still cannot solve the problem of initial screening under sample-free conditions. How to achieve initial risk screening using existing remote sensing, topographic, and socioeconomic data under sample-free conditions, and achieve high-precision calibration after obtaining samples, to construct a hierarchical evaluation technology system that balances efficiency and accuracy, has always been a technical challenge that urgently needs to be overcome in this field. The above-mentioned deficiencies collectively result in shortcomings in the spatial continuity, mechanism integration, and scenario adaptability of existing mine area soil heavy metal risk assessment technologies, making it difficult to fully meet the actual needs of current refined management and restoration decision-making in the ecological environment of mining areas. Summary of the Invention

[0005] In view of this, the present invention aims to propose a multi-evidence fusion remote sensing heavy metal pollution risk index construction method to solve the problem that existing evaluation models cannot output reliable spatial continuous risk results due to the lack of sufficient training samples or calibration benchmarks, which makes it difficult for management departments to quickly identify key investigation areas under low cost.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index, the method comprising: Obtain heavy metal content data and corresponding spatial coordinates of soil sampling points in the mining area. Calculate the Nemerow Comprehensive Pollution Index (PN) of each sampling point based on the heavy metal content data and soil background value or soil environmental quality standard value. Divide the sampling points into warning risk points and non-warning risk points based on whether the PN is greater than a preset risk threshold. Based on the generation, migration, occurrence and response process of heavy metal pollution in mining area soil, a risk warning feature set with process correlation is constructed. The multi-evidence chain includes pollution source driving evidence layer D, pollution migration regulation evidence layer T, soil occurrence medium evidence layer M, remote sensing spectral response evidence layer S and ecological stress response evidence layer V. The evidence layers are organized in a cascade relationship of F={D→T→M→(S,V)}, where F is the process cascade multi-evidence feature set. When ground sampling data is available, the risk state mapping model is trained using the multi-evidence chain risk warning feature set as the input feature and the risk state of the sample points divided by PN as the target variable. The predicted probability of the sample points belonging to the warning risk state output by the model is defined as the sample risk warning index RWI. When there is no ground sampling data, the risk direction of each evidence layer feature in the multi-evidence chain risk warning feature set is unified, normalized and weighted within the layer in sequence, and then weighted between the layers to obtain the sample-free basic risk warning index RWIn. Based on RWI or RWIn, the mining area is classified and zoned for risk, and risk warning results are output.

[0007] Furthermore, a preferred approach is proposed, wherein the pollution source driving evidence layer D selects nighttime light brightness value and road density as features to characterize the input driving force of heavy metal pollution generated by mining, smelting, and transportation activities; The pollution migration regulation evidence layer T is based on a digital elevation model to extract elevation, slope, aspect, and topographic humidity index, and combines them with water body index to characterize the regulatory effect of topographic relief and hydrological conditions on the migration, dilution, and deposition of pollutants with surface runoff. The soil-borne medium evidence layer M is based on soil indices, iron oxide indices, clay content indices, and alteration indices retrieved from hyperspectral remote sensing images, and is used to characterize the ability of soil active adsorption components to store and accumulate heavy metals. The remote sensing spectral response evidence layer S extracts the original reflectance, statistical characteristics of the first and second derivatives of the spectrum, and morphological parameters of the depth, area, center position, and width of the spectral absorption band in the visible-near-infrared-shortwave infrared band. The evidence layer V for the ecological stress response selects the normalized vegetation index, soil-regulated vegetation index, enhanced vegetation index, green light normalized vegetation index, and red-edged vegetation index to characterize the indirect stress response of heavy metal pollution to vegetation growth.

[0008] Furthermore, an optimal method is proposed: when calculating the Nemerow Composite Pollution Index (PN), first follow the formula... Calculate the single-factor pollution index P for each heavy metal element. i ,in, C i The measured content of heavy metal i is given. B i The soil background or standard value for heavy metal i; then according to the formula Calculate the Nemerow composite pollution index, where, P imax This represents the maximum value of the single-factor pollution index. This represents the average value of the single-factor pollution index. The preset risk threshold is set to 2.0. Sampling points with PN>2.0 are classified as warning risk sampling points, and sampling points with PN≤2.0 are classified as non-warning risk sampling points.

[0009] Furthermore, an optimal approach is proposed: when training the risk state mapping model, the input features are standardized, and class weights are set according to the difference in the number of warning risk samples and non-warning risk samples, so that the risk category with fewer samples receives higher training weights; a hierarchical cross-validation method is used to train and optimize the candidate model, and the predicted probability that the spatial unit output by the optimized model belongs to the warning risk state is defined as the sampled risk warning index RWI, where the value of RWI ranges from 0 to 1.

[0010] Furthermore, an optimal method is proposed: when calculating the sample-less basic risk warning index RWIn, the j-th original feature in the k-th evidence layer is... Normalize: For positive indicators where an increase in value signifies increased risk, according to... Perform positive normalization; For negative indicators where an increase in value indicates a decrease in risk, according to... Perform negative normalization; Calculate the sub-index of the k-th evidence layer:

[0011] in, For the k-th evidence layer, the j-th indicator is the minimum value of the j-th indicator across the entire study area; For the k-th evidence layer, the j-th indicator is the maximum value of the j-th indicator across the entire study area; The index is the normalized value; For the weights within the layer, This represents the number of features in this layer of evidence; Recalculate ,in, K represents the interlayer weights of the evidence layers, and K is the set of evidence layers that actually participate in the fusion. When data of a certain evidence layer is missing, that evidence layer is omitted from the set K and the weights of the remaining evidence layers are renormalized.

[0012] Furthermore, a preferred method is proposed, wherein the intra-layer weights Interlayer weights The entropy weight method can be used to determine the objective value, or the analytic hierarchy process can be used to calibrate the value when there is prior knowledge.

[0013] Furthermore, a preferred approach is proposed: when performing risk stratification and partitioning based on RWI or RWIn, a recall constraint method based on the receiver operating characteristic curve is used to determine the stratification threshold, dividing the mining area into: Safe zone: RWI < safety threshold; Concern Area: Safety threshold ≤ RWI < Warning threshold; Warning zone: RWI ≥ warning threshold.

[0014] Furthermore, a preferred approach is proposed, in which the method further includes verifying the RWI: Using PN as a pollution risk reference standard, the mean PN, median PN, proportion of sample points with PN greater than the preset risk threshold, and risk enrichment factor were statistically analyzed under different RWI levels. Correlation analysis or receiver operating characteristic (ROC) curve analysis was used to verify the consistency between RWI and PN risk status.

[0015] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a method for constructing a multi-evidence fusion remote sensing heavy metal pollution risk index according to any one of the preceding claims.

[0016] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of a method for constructing a multi-evidence fusion remote sensing heavy metal pollution risk index as described in any one of the above.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a novel technical concept: a multi-evidence fusion remote sensing heavy metal pollution risk index construction approach based on a five-level process cascade evidence chain of "source-pathway-sink-exploration-effect". At the same time, it establishes a dual-mode evaluation system that unifies sample-supervised modeling and sample-free weighted fusion, filling the gap in the field of integrated technology for rapid screening and evaluation of heavy metal risks in mining area soils.

[0018] The core technology of this invention lies in reconstructing the correlation logic between multi-source environment, remote sensing features, and soil heavy metal pollution risk, breaking through the cognitive limitations of existing evaluation methods that arbitrarily stack various features. The occurrence and development of heavy metal pollution in mining area soils follows a complete natural process chain. First, human activities such as mining, smelting, and transportation generate pollution input, constituting the initial driving conditions for pollution formation. After pollutants enter the environment, they undergo spatial migration and redistribution under the regulation of conditions such as topographic relief, surface runoff, and hydrological connectivity. This process affects the spatial diffusion range of pollution. The migrated heavy metals are further adsorbed and fixed by active components such as iron oxides, clay minerals, and organic matter in the soil, forming a stable pollution occurrence state, which is the core reason why pollution can persist for a long time. When pollution accumulates to a certain extent, it will cause changes in soil physicochemical properties, distortion of spectral absorption characteristics, and vegetation growth stress, ultimately forming a surface response signal that can be captured by remote sensing. Based on this, the present invention constructs a five-level cascaded evidence chain of pollution source driving, pollution migration regulation, soil occurrence medium, remote sensing spectral response, and ecological stress response. Each evidence layer is interconnected in the pollution formation process, rather than being independent feature variables, so that the feature changes at each level can correspond to the specific links in the pollution process, and the matching of features and risk essence is achieved at the mechanism level.

[0019] Based on this, this invention uses the Nemerow Integrated Pollution Index as a risk calibration benchmark. This index is a universally accepted and mature indicator in the field of soil pollution assessment, possessing widespread reliability. When ground sampling data is available, a risk mapping model is trained using the five-level evidence chain features as input and the risk status of sample points classified by the Nemerow index as the target. The warning risk probability output by the model is defined as the sample-based risk warning index, realizing the transformation from discrete sample point pollution judgment to continuous risk characterization across the entire region. When ground sampling data is unavailable, based on the same set of five-level evidence chain features, various indicators are first uniformly normalized according to risk direction, and then a sample-free basic risk warning index is generated through hierarchical weighted fusion. This reflects the relative risk level of different areas without relying on measured data. The two types of indices are constructed based on a common feature system, enabling seamless connection between sample-free initial screening and sample-based calibration, adapting to the needs of different stages of mining area investigation.

[0020] The interpretability of the evaluation logic in this invention is significantly improved. Because the five-level chain of evidence strictly corresponds to the pollution evolution process, the evaluation results can clearly trace the causes of risk. For example, a high risk in a certain area can be traced back to specific factors such as dense mining activities in the surrounding area, flat terrain conducive to pollutant deposition, and high clay content in the soil with strong adsorption capacity. This avoids the problem of existing technologies lacking mechanistic support in their evaluation results, providing clear guidance for pollution source tracing and the formulation of remediation plans. Secondly, the method proposed in this invention has significantly enhanced scenario adaptability. The dual-mode design solves the pain point of traditional methods' over-reliance on ground sampling data. In the general survey stage of mining areas where no measured samples are available, potential high-risk areas can be quickly identified through the sample-free index, guiding the optimized layout of sampling points and avoiding the waste of resources from uniform sampling across the entire area. After obtaining sampling data in the detailed survey stage, high-precision risk assessment can be achieved through the sample-based index, balancing the efficiency requirements of large-scale surveys with the accuracy requirements of key areas. Furthermore, the continuity of risk characterization avoids information loss. Traditional methods output discrete pollution levels or hard classification results that lose a significant amount of risk gradient information. The continuous risk index output by this invention can quantify the relative risk level of any spatial unit within the 0-1 range, supporting high-precision comparisons of risk levels in different regions and providing more refined support for risk classification and control. In addition, the risk classification threshold is determined using a recall-constrained method, which aligns with the practical need for large-scale risk screening to avoid false alarms rather than false negatives, effectively reducing the probability of missed detections in high-risk areas. The weight determination method combines objectivity and flexibility, objectively reflecting the information differences in the data itself through entropy weighting, or combining prior knowledge and expert experience through analytic hierarchy process (AHP) calibration, adapting to the application needs of different mining areas. The entire method adopts a standardized data processing and calculation process, is reproducible, and has high result stability, making it widely applicable to soil heavy metal risk assessment, ecological restoration priority zone delineation, and environmental management decision support in different types of mining areas. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a method for constructing a multi-evidence fusion remote sensing heavy metal pollution risk index according to the present invention. Figure 2 This is a schematic diagram of the spatial distribution of RWI as described in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other, and the described embodiments are only some embodiments of the present invention, not all embodiments.

[0023] Implementation Method 1, see [link] Figure 1 This embodiment describes a method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index, the method comprising: Obtain heavy metal content data and corresponding spatial coordinates of soil sampling points in the mining area. Calculate the Nemerow Comprehensive Pollution Index (PN) of each sampling point based on the heavy metal content data and soil background value or soil environmental quality standard value. Divide the sampling points into warning risk points and non-warning risk points based on whether the PN is greater than a preset risk threshold. Based on the generation, migration, occurrence and response process of heavy metal pollution in mining area soil, a risk warning feature set with process correlation is constructed. The multi-evidence chain includes pollution source driving evidence layer D, pollution migration regulation evidence layer T, soil occurrence medium evidence layer M, remote sensing spectral response evidence layer S and ecological stress response evidence layer V. The evidence layers are organized in a cascade relationship of F={D→T→M→(S,V)}, where F is the process cascade multi-evidence feature set. When ground sampling data is available, the risk state mapping model is trained using the multi-evidence chain risk warning feature set as the input feature and the risk state of the sample points divided by PN as the target variable. The predicted probability of the sample points belonging to the warning risk state output by the model is defined as the sample risk warning index RWI. When there is no ground sampling data, the risk direction of each evidence layer feature in the multi-evidence chain risk warning feature set is unified, normalized and weighted within the layer in sequence, and then weighted between the layers to obtain the sample-free basic risk warning index RWIn. Based on RWI or RWIn, the mining area is classified and zoned for risk, and risk warning results are output.

[0024] In this embodiment, the pollution source driving evidence layer D selects nighttime light brightness value and road density as features to characterize the input driving force of heavy metal pollution generated by mining, smelting and transportation activities. The pollution migration regulation evidence layer T is based on a digital elevation model to extract elevation, slope, aspect, and topographic humidity index, and combines them with water body index to characterize the regulatory effect of topographic relief and hydrological conditions on the migration, dilution, and deposition of pollutants with surface runoff. The soil-borne medium evidence layer M is based on soil indices, iron oxide indices, clay content indices, and alteration indices retrieved from hyperspectral remote sensing images, and is used to characterize the ability of soil active adsorption components to store and accumulate heavy metals. The remote sensing spectral response evidence layer S extracts the original reflectance, statistical characteristics of the first and second derivatives of the spectrum, and morphological parameters of the depth, area, center position, and width of the spectral absorption band in the visible-near-infrared-shortwave infrared band. The evidence layer V for the ecological stress response selects the normalized vegetation index, soil-regulated vegetation index, enhanced vegetation index, green light normalized vegetation index, and red-edged vegetation index to characterize the indirect stress response of heavy metal pollution to vegetation growth.

[0025] In this embodiment, when calculating the Nemerow Composite Pollution Index (PN), first follow the formula... Calculate the single-factor pollution index P for each heavy metal element. i ,in, C i The measured content of heavy metal i is given. B i The soil background or standard value for heavy metal i; then according to the formula Calculate the Nemerow composite pollution index, where, P imax This represents the maximum value of the single-factor pollution index. This represents the average value of the single-factor pollution index. The preset risk threshold is set to 2.0. Sampling points with PN>2.0 are classified as warning risk sampling points, and sampling points with PN≤2.0 are classified as non-warning risk sampling points.

[0026] In this embodiment, when training the risk state mapping model, the input features are standardized, and the category weights are set according to the difference in the number of warning risk samples and non-warning risk samples, so that the risk category with fewer samples gets higher training weights. The candidate model is trained and optimized using a hierarchical cross-validation method, and the predicted probability that the spatial unit output by the optimized model belongs to the warning risk state is defined as the sampled risk warning index RWI, where the value of RWI ranges from 0 to 1.

[0027] In this embodiment, when calculating the risk warning index RWIn without a sample base, the j-th original feature in the k-th evidence layer is... Normalize: For positive indicators where an increase in value signifies increased risk, according to... Perform positive normalization; For negative indicators where an increase in value indicates a decrease in risk, according to... Perform negative normalization; Calculate the sub-index of the k-th evidence layer:

[0028] in, For the k-th evidence layer, the j-th indicator is the minimum value of the j-th indicator across the entire study area; For the k-th evidence layer, the j-th indicator is the maximum value of the j-th indicator across the entire study area; The value of the j-th feature in the k-th evidence layer after risk direction unification and normalization; Intra-layer weights; This represents the number of features in this layer of evidence; Recalculate ,in, K represents the interlayer weights of the evidence layers, and K is the set of evidence layers that actually participate in the fusion. When data of a certain evidence layer is missing, that evidence layer is omitted from the set K and the weights of the remaining evidence layers are renormalized.

[0029] Furthermore, the intra-layer weights Interlayer weights The entropy weight method can be used to determine the objective value, or the analytic hierarchy process can be used to calibrate the value when there is prior knowledge.

[0030] In this embodiment, when performing risk stratification and partitioning based on RWI or RWIn, the recall constraint method based on the receiver operating characteristic curve is used to determine the stratification threshold, dividing the mining area into: Safe zone: RWI < safety threshold; Concern Area: Safety threshold ≤ RWI < Warning threshold; Warning zone: RWI ≥ warning threshold.

[0031] In this embodiment, the method further includes verifying the RWI: Using PN as a pollution risk reference standard, the mean PN, median PN, proportion of sample points with PN greater than the preset risk threshold, and risk enrichment factor were statistically analyzed under different RWI levels. Correlation analysis or receiver operating characteristic (ROC) curve analysis was used to verify the consistency between RWI and PN risk status.

[0032] Implementation Method 2, see below Figure 2 This embodiment describes a complete implementation of the multi-evidence fusion-based remote sensing heavy metal pollution risk index construction method described in Embodiment 1, including: Step 1: Soil sample collection and spatial information recording in the mining area.

[0033] Based on the topography of the mining area and the research objectives, soil sampling points were established within the mining area. A five-point quincunx sampling method was used to collect soil samples, and the latitude and longitude coordinates of each sampling point were recorded simultaneously. Each sampling point was uniquely numbered to ensure a one-to-one correspondence between the soil sample, its spatial location, and subsequent experimental results.

[0034] Step 2: Soil sample pretreatment and heavy metal content detection.

[0035] The collected soil samples were brought back to the laboratory for pretreatment, including natural air drying, removal of stones and plant debris, grinding, sieving, and mixing. The pretreated soil samples were then sent to the laboratory for heavy metal content determination, obtaining data on the content of heavy metal elements such as Cr, Ni, Cu, Zn, Cd, Pb, As, and Hg in the soil at each sampling point. This heavy metal content data, together with the sampling point number and latitude and longitude coordinates, constitutes the basic dataset of heavy metals in the mining area soil.

[0036] Step 3: Calculation of Nemerow Comprehensive Pollution Index (PN) and risk status classification.

[0037] Based on the soil heavy metal content data obtained in step two, and combined with the corresponding soil environmental quality standard values ​​or regional background values, the single-factor pollution index of each heavy metal element is calculated, and the Nemerow comprehensive pollution index (PN) of each sampling point is further calculated. The pollution risk status of the sampling points is classified according to a preset pollution risk threshold. Preferably, when PN is greater than 2.0, the sampling point is defined as a warning risk sampling point; when PN is less than or equal to 2.0, the sampling point is defined as a non-warning risk sampling point.

[0038] Step 4: Construct a risk warning feature set with multiple evidence chains.

[0039] To overcome the problem that existing remote sensing risk assessment methods for heavy metals in soil are characterized by the independence of different environmental factors and their inability to reflect the formation and evolution of pollution, this invention constructs a multi-evidence chain risk warning feature set with process correlation based on the evolution mechanism of heavy metal pollution in mining area soil from generation input, migration and diffusion, soil accumulation to remote sensing response and ecological impact. This multi-evidence feature set serves as a unified feature base for both sample-based modeling and sample-free fusion. It can be used as input variables for risk warning models and for risk index synthesis under sample-free conditions, achieving consistent expression between sample-based supervised modeling and sample-free risk assessment.

[0040] Unlike existing technologies that directly input remote sensing spectral features, topographic factors, land use factors, and vegetation indices as independent variables into the evaluation model, this implementation method points out that soil heavy metal risk is not determined by a single environmental factor, but rather by the combined effects of pollution source input, spatial migration conditions, soil media storage capacity, and surface ecological response. Specifically, human activities such as mining, smelting, and transportation in mining areas are the primary driving force for heavy metal pollution input. After entering the environment, pollutants migrate, diffuse, and redistribute spatially under the influence of factors such as topographic relief, surface runoff, hydrological conditions, and spatial connectivity. The migrated heavy metals are further adsorbed, fixed, and accumulated by active components such as soil mineral composition, iron oxides, clay minerals, and organic matter, forming spatially stable pollution enrichment areas. As the degree of pollution increases, soil physicochemical properties, surface color characteristics, spectral response characteristics, and vegetation growth status change, ultimately forming spatial response information that can be captured by remote sensing technology.

[0041] Therefore, this implementation constructs a multi-level evidence chain of "source-pathway-sink-detection-effect" to achieve a correlation mapping between the pollution formation process and remote sensing spatial representation. The process is expressed as: F={D→T→M→(S,V)}. Here, arrows indicate the correlation and transmission relationship of pollution risk evidence along the pollution formation and spatial response processes, and parentheses indicate that the remote sensing response and ecological response formed after pollution accumulation are parallel results. It should be noted that the five evidence layers D, T, M, S, and V in this implementation are not independent parallel variables, but rather a cascaded evidence chain constructed based on the generation, migration, occurrence, and response processes of heavy metal pollution in mining area soil.

[0042] The multi-evidence fusion process constructed in this implementation reflects a cascading mechanism of "pollution source-driven - migration regulation - media occurrence - remote sensing response - ecological response." There are correlations and transmission relationships between each evidence layer based on the pollution evolution process, rather than simple data stacking. This mechanism can establish a correspondence between pollution formation mechanisms and remote sensing spatial representations, helping to improve the geoscientific interpretation and spatial risk identification capabilities of risk indices. This is a key technical feature that distinguishes it from existing evaluation methods based solely on pollution concentration, single spectral anomalies, or disordered multi-factor fusion.

[0043] Among them, the source corresponds to the pollution source driving evidence layer D, which is used to characterize the driving force of heavy metal pollution input generated by human activities such as mining, smelting, and transportation, and is the starting condition for the formation of pollution risk; the path corresponds to the pollution migration regulation evidence layer T, which is used to describe the influence of topography, hydrological conditions and spatial connectivity on the migration of pollutants from the source area to the surrounding area, representing the pollution migration regulation process, rather than a single physical migration path, and is an important control factor affecting the range and spatial distribution of pollution diffusion. The evidence layer M, corresponding to the soil storage medium, describes the adsorption, fixation, and accumulation capacity of pollutants after they enter the soil, and is a key factor affecting the persistence and spatial stability of pollution. The evidence layer S, corresponding to the remote sensing spectral response, describes the remotely observable response after heavy metal accumulation causes changes in soil color, mineral composition, and spectral absorption characteristics. The evidence layer V, corresponding to the ecological stress response, describes the indirect response after heavy metal pollution affects vegetation growth and ecosystem function.

[0044] By constructing the multi-level evidence chain as described above, the method proposed in this embodiment unifies the pollution formation mechanism, environmental migration process, soil occurrence characteristics, and remote sensing response information into the same risk warning feature system, providing a process-related feature foundation for subsequent intelligent model training with samples and risk index calculation without samples.

[0045] Step 5: Establish a risk warning model for heavy metals in the soil of the mining area.

[0046] Based on the feature set of the multi-evidence chain constructed by the pollution formation mechanism, a risk state mapping model is established to realize the transformation from multi-dimensional environmental evidence to the risk state of soil heavy metal pollution.

[0047] The risk status, determined by the Nemerow Comprehensive Pollution Index (PN), is used as the target variable in the model. When PN exceeds a preset risk threshold, the corresponding sample point is defined as a warning risk sample point; when PN is less than or equal to the preset risk threshold, the corresponding sample point is defined as a non-warning risk sample point. The obtained multi-evidence chain risk warning features are used as model input variables. These input features include one or more features from the following layers: pollution source-driven evidence layer D, pollution migration and regulation evidence layer T, soil-borne medium evidence layer M, remote sensing spectral response evidence layer S, and ecological stress response evidence layer V. A risk warning model is established using a statistical learning model, machine learning model, or other predictive model capable of outputting risk category probabilities. During model training, the model learns the correspondence between different combinations of evidence layer features and the actual pollution risk status, establishing a mapping relationship from "pollution source-driven - migration and regulation - medium-borne - remote sensing response - ecological response" to pollution risk status.

[0048] Unlike existing evaluation methods that rely solely on a single pollution indicator or multivariate inputs without mechanistic constraints, the risk model constructed in this implementation uses multiple evidence chains as its input basis. This allows the model to comprehensively utilize spatial environmental information corresponding to different pollution evolution stages, thereby improving the geoscientific interpretability of the risk assessment results. After model training, parameter optimization, and performance evaluation, the candidate model with relatively superior evaluation indicators is obtained, and the predicted probability value of each spatial unit belonging to a warning risk state is output for subsequent construction of the Risk Warning Index (RWI).

[0049] Step 6: Construct the Risk Warning Index (RWI) for heavy metals in the mining area soil.

[0050] The obtained risk state prediction probabilities are converted into a continuous risk warning index (RWI) to achieve the transformation from discrete risk classification results to continuous spatial risk representation results. That is, the predicted probability that a sample point or spatial unit output by the model belongs to a warning risk category is defined as the risk warning index. RWI=P(Y=1|F) Where Y=1 represents the warning risk state when PN exceeds the preset risk threshold, and F represents the multi-evidence chain feature.

[0051] Since this probability value comprehensively reflects the degree to which pollution source input, migration conditions, soil storage capacity, and surface response information support the risk status, it can be used as a probabilistic characterization index for regional soil heavy metal pollution risk. The RWI value ranges from 0 to 1. The higher the value, the higher the consistency between the multi-evidence chain information and the high pollution risk status in the region, and the greater the potential pollution risk; the lower the value, the lower the correlation between the region and the warning risk status.

[0052] Compared with traditional pollution level assessment methods based on fixed thresholds, the method proposed in this embodiment reduces the information loss problem in discrete classification by expressing risk probability continuously. It can realize continuous comparison of risk levels between different spatial units, providing a basis for subsequent risk zoning and key area identification.

[0053] Step 7: Divide the mining area into risk zones based on the risk warning index.

[0054] To achieve tiered warnings of soil heavy metal pollution risks, and to meet the core principle of "better to report false positives than to miss reports" in large-scale rapid screening, this implementation method adopts the recall constraint method based on the receiver operating characteristic curve to determine the grading threshold of the risk warning index RWI and construct a three-level warning system of "safe zone - concern zone - warning zone".

[0055] Step 8: Verify the risk warning index.

[0056] Using PN results as a pollution risk reference standard, statistical analysis was conducted on the mean PN, median PN, proportion of sample points with PN greater than the preset threshold, and risk enrichment factor under different RWI levels to verify whether RWI can reflect the spatial variation characteristics of heavy metal pollution risk in mining area soil. Further verification methods could include correlation analysis, receiver operating characteristic (ROC) curve analysis, or consistency analysis.

[0057] Step 9: Output the heavy metal risk warning results for the mining area soil.

[0058] The output includes RWI values ​​at the sample point scale, a risk warning statistics table, and a spatial distribution map of heavy metal risk in the mining area soil. These results can be used for identifying soil pollution risks in mining areas, delineating priority areas for ecological restoration, zoning pollution control, and making environmental management decisions.

[0059] Step 10: Constructing the Risk Warning Index RWIn without a Sample Base.

[0060] "No sample" refers to the absence of actual measured samples of heavy metals in ground soil, but the availability of spatial environmental data such as remote sensing images, DEM, roads, and nighttime lighting.

[0061] To address the issue of insufficient ground sampling data and difficulty in conducting monitoring modeling during the initial stages of mining area investigations, this implementation method further provides a sample-free risk screening method based on multi-evidence chain feature fusion. This method does not rely on measured heavy metal content data; instead, it directly utilizes the multi-evidence chain risk warning feature set constructed in step four to comprehensively characterize the source-driven factors, migration conditions, media occurrence, and spatial response information in the pollution formation process, constructing a basic risk warning index, RWIn.

[0062] Specifically, it can be expressed as: RWIn=f(D,T,M,S,V) Where D, T, M, S, and V represent the evidence layers of pollution source driving, migration regulation, soil occurrence, remote sensing response, and ecological stress, respectively; f represents the multi-evidence fusion function.

[0063] The spatial features in each layer of evidence are first subjected to risk direction unification and normalization. For indicators whose numerical increase indicates stronger risk evidence, positive normalization is used; for indicators whose numerical increase indicates weaker risk evidence, negative normalization is used; for indicators that have a non-monotonic relationship with risk, a preset membership function can be used to convert them into risk characterization values ​​in the range of 0 to 1.

[0064] Positive risk indicators:

[0065] Negative risk indicators:

[0066] Among them, the k The first layer of evidence j The original indicators are After normalization, it becomes .

[0067] No. k Sub-index of each layer of evidence The calculation is as follows:

[0068] in, For the first k The first layer of evidence j The numerical values ​​of each feature after risk direction unification and normalization For the corresponding feature, the intra-layer weights, This represents the number of features contained in this layer of evidence.

[0069] The basic risk warning index RWIn is calculated as follows:

[0070] in, For the first k Interlayer weights of each layer of evidence K This is the set of evidence layers that actually participate in the fusion. When data from a certain evidence layer is missing, that evidence layer is omitted from set K, and the weights of the remaining evidence layers are renormalized. RWIn ranges from 0 to 1; the higher the value, the stronger the comprehensive potential risk evidence presented by that spatial unit. RWIn is used for initial risk screening under conditions where there are no measured heavy metal samples on the ground, and is not directly equivalent to the actual heavy metal content or the measured degree of pollution.

[0071] Each layer of evidence is first normalized based on the spatial characteristics of its corresponding category, and then sub-indices of the evidence layer are obtained through weighted fusion. The weights of each evidence layer and its features can be objectively determined using the entropy weighting method to preserve the information differences within the data to the greatest extent possible. In areas with prior knowledge or experienced experts, the analytic hierarchy process (AHP) can also be used for weighting. Finally, a basic risk warning index, RWIn, is obtained through multi-evidence weighted fusion, with a value ranging from 0 to 1. The higher the RWIn value, the higher the risk characteristics exhibited by multiple risk evidence layers in the area, and the higher the potential risk of heavy metal pollution.

[0072] Unlike RWI, which is calibrated based on measured pollution data under sample conditions, RWIn is mainly used for large-scale rapid risk screening under sample-free conditions. It can be used for identifying potentially high-risk areas, optimizing the layout of sampling points, and determining subsequent refined evaluation areas. Therefore, the method proposed in this embodiment forms a two-stage application process of "rapid initial screening without samples - calibration and evaluation with samples". Risk zoning serves as the output link of the two-stage evaluation results, improving the applicability of the method under different mining area survey stages and data conditions.

[0073] Implementation Method 3: This implementation method is a specific application example of the multi-evidence fusion remote sensing heavy metal pollution risk index construction method described in Implementation Method 1. This embodiment uses a typical coal-mining county in Northeast my country as the study area. This region has long been disturbed by mining activities, resulting in a significant risk of heavy metal accumulation in the soil. The soil types are mainly black soil and dark brown soil. The specific implementation steps of this method are as follows: Step S1: Sample collection and spatial data acquisition in the study area Within the study area, soil sampling points were established based on the distribution pattern of mining areas, land use types, and topographic conditions. A five-point quincunx sampling method was used for mixed sampling, with a sampling depth of 0-20 cm, resulting in a total of 200 surface soil sampling points. The latitude, longitude, and elevation information of each sampling point were simultaneously acquired using equipment, and each sampling point was uniquely numbered.

[0074] Step S2: Sample pretreatment and heavy metal detection After natural air drying, the collected soil samples were cleaned of stones and plant debris, ground, sieved, and mixed thoroughly for later use. Inductively coupled plasma mass spectrometry (ICP-MS) was used for analysis from September 24th to September 28th, 2025, ultimately obtaining data on the content of eight heavy metals (Cr, Ni, Cu, Zn, Cd, Pb, As, and Hg) in all samples.

[0075] Step S3: Calculation of Nemerow Integrated Pollution Index and Risk Classification This study uses the median values ​​of heavy metals in layer A soil of Heilongjiang Province from the statistical results of "Background Values ​​of Soil Elements in China" as the regional natural background values.

[0076] The formula for calculating the single-factor pollution index is as follows:

[0077] In the formula, P i This represents the single-factor pollution index value; C i This indicates the measured content (mg / kg) of a certain heavy metal element. B i This indicates the soil background value (mg / kg) for a certain heavy metal element.

[0078] The formula for calculating the Nemerow Comprehensive Pollution Index is as follows:

[0079] In the formula, PN represents the Nemerow Comprehensive Pollution Index value; P imax This represents the maximum value in a single-factor index; This represents the average value of a single-factor index.

[0080] After outlier screening and data cleaning of the original 200 samples, 144 valid samples were retained, and 56 samples were not included in subsequent modeling. Based on the PN calculation results, a risk warning threshold of PN=2.0 was used for classification. PN>2.0: 51 points were identified as high-risk warning points, accounting for 35.4% of the valid samples; PN≤2.0: These were identified as non-alert risk sampling points, totaling 93, accounting for 64.6% of the valid samples.

[0081] Step S4: Construction of Multi-Source Risk Features Unlike existing technologies that utilize only a single pollution concentration or remote sensing anomaly, this implementation method does not simply superimpose multiple environmental variables. Instead, it establishes a multi-level evidence chain with process correlation based on the formation process of heavy metal pollution, from pollution source, migration and diffusion, mineral adsorption and fixation, surface response to ecological feedback. By constructing a multi-level evidence chain of "source-pathway-sink-detection-effect," the correlation mapping between the pollution formation process and remote sensing spatial representation is achieved. This can be expressed as: F={D→T→M→(S,V)}.

[0082] (1) Pollution source-driven evidence layer D: used to characterize the generation and input pressure of heavy metal pollutants, and is the starting condition for risk formation. Two spatial indicators, nighttime light and road density, are selected to quantify the degree of industrial activity agglomeration in the mining area and the pollution diffusion potential brought by transportation, and to characterize the external input pressure of heavy metal pollution.

[0083] (2) Evidence layer T for pollution migration regulation: This layer describes the influence of topography, hydrology, and spatial connectivity on the migration of pollutants from the source area to the surrounding area, and is an important factor affecting the range of pollution diffusion. Based on the digital elevation model, elevation, slope, aspect, and topographic humidity index are extracted and combined with water body index to characterize the regulatory effect of topographic relief and hydrological conditions on the migration, dilution, and deposition of heavy metal elements with surface runoff.

[0084] (3) Soil storage medium evidence layer M: This layer describes the adsorption, fixation, and accumulation capacity of pollutants after they enter the soil, and is an important controlling factor affecting the persistence of pollution. Soil indices, iron oxide indices, clay content indices, and various alteration indices are obtained based on hyperspectral remote sensing images to characterize the relative content of active adsorbent components such as iron oxides and clay minerals in the soil. The above mineral components are the main carriers for the immobilization of heavy metals in the soil, and their spatial distribution affects the enrichment capacity and storage intensity of heavy metals.

[0085] (4) Remote sensing spectral response evidence layer S: This layer describes the remotely observable response of soil color, mineral composition, and spectral absorption characteristics after heavy metal accumulation. Three types of spectral features are extracted: first, the reflectance of six original bands in the visible-near-infrared-shortwave infrared range, reflecting the overall spectral reflectance characteristics of the soil; second, the statistical characteristics of the first and second derivatives of the spectrum, amplifying the details of local changes in the spectral curve; and third, the morphological parameters such as the depth, area, center position, and width of the spectral absorption band, capturing the characteristic absorption distortion caused by the interaction between heavy metals and soil minerals. These three types of features together constitute candidate features for characterizing spectral changes related to heavy metal risk in soil.

[0086] (5) Ecological Stress Response Evidence Layer V: This layer describes the indirect response to the impact of heavy metal pollution on vegetation growth and ecosystem function. Normalized Difference Vegetation Index (NDVE), Soil-Regulated Vegetation Index (SVEI), Enhanced Vegetation Index (EGI), and Red-Edge Vegetation Index (DEI) are selected to reflect the growth and physiological state of surface vegetation. Heavy metal pollution may stress vegetation growth, manifesting as a decrease in vegetation indices or other abnormal changes. Since vegetation status may also be affected by factors such as water, nutrients, climate, and human activities, the vegetation indices primarily serve as indirect auxiliary evidence of pollution risk.

[0087] The exponent is calculated as follows (ε=10 is added to the denominator for all ratio calculations). -6 (To avoid numerical anomalies) NDVI (Normalized Difference Vegetation Index):

[0088] Where N is the near-infrared reflectance and R is the red light reflectance.

[0089] SAVI Soil-Regulating Vegetation Index:

[0090] Where N is the near-infrared reflectance and R is the red light reflectance.

[0091] EVI (Enhanced Vegetation Index):

[0092] Where N is the near-infrared reflectance, R is the red light reflectance, and B is the blue light reflectance.

[0093] GNDVI (Gold Normalized Difference Vegetation Index):

[0094] Where N is the near-infrared reflectance and G is the green light reflectance.

[0095] NDWI_green_nir (Normalized Difference Water Index):

[0096] Where N is the near-infrared reflectance and G is the green light reflectance.

[0097] Near-infrared / red ratio:

[0098] Where N is the near-infrared reflectance and R is the red light reflectance.

[0099] BSI (Boiler Index):

[0100] Where N is the near-infrared reflectance, R is the red light reflectance, B is the blue light reflectance, and S1 is SWIR1.

[0101] Iron oxide index (iron staining):

[0102] Where R is the red light reflectance and B is the blue light reflectance.

[0103] Clay index (hydroxyl alteration):

[0104] Where S1 is SWIR1 and S2 is SWIR2.

[0105] Soil Index:

[0106] Where S1 is SWIR1 and N is near-infrared reflectance.

[0107] Carbonate alteration:

[0108] Where S1 is SWIR1 and S2 is SWIR2.

[0109] Spectral first derivative characteristics:

[0110] in, Let be the spectral reflectance corresponding to the (i+1)th band. This represents the spectral reflectance corresponding to the (i-1)th band.

[0111] First derivative mean:

[0112] in, n This represents the total number of spectral bands involved in the derivative characteristic calculation. Let be the value of the first-order spectral derivative at the i-th band.

[0113] Standard deviation of first derivative:

[0114] Maximum value of the first derivative:

[0115] Minimum value of the first derivative:

[0116] Spectral second derivative characteristics:

[0117] in, Let be the value of the first derivative of the spectrum at the (i+1)th band. It is the value of the first-order spectral derivative at the (i-1)th band.

[0118] Second derivative mean:

[0119] Second derivative standard deviation:

[0120] Maximum value of the second derivative:

[0121] Minimum value of the second derivative:

[0122] in, For the first spectral derivative, Let be the second-order spectral derivative, and Ri be the spectral value of the i-th band.

[0123] Continuum calculation:

[0124]

[0125] Where R1 is the first band and Rn is the last band. t i For the position of the i-th band, It is a continuum.

[0126] Absorption characteristics:

[0127]

[0128]

[0129]

[0130]

[0131]

[0132]

[0133] in, Absorption amount, For a continuum, H represents the maximum absorption location, and H represents the half-depth. , For absorption depth, For the absorption area, As an absorption center, For absorption width, denoted as the original spectral reflectance of the ground object corresponding to the i-th band.

[0134] Step S5: Data Preprocessing For heavy metal element content data, the interquartile range method was used to clean outliers in the heavy metal detection data. The effective value range of each indicator was determined by calculating the interquartile range; if any indicator at a sampling point exceeded this range, it was identified as an outlier and the entire record was removed to eliminate extreme value interference and ensure the statistical robustness of the dataset. Based on this, the median value of heavy metals in Layer A soil of Heilongjiang Province was used as the regional natural background value from the statistical results of "Background Values ​​of Soil Elements in China" to calculate the Nemerow pollution index.

[0135] This implementation uses the processed ZY1E hyperspectral imagery as spatial reference data, reading its projection coordinate system, affine transformation parameters, image size, and spatial resolution to establish a unified spatial reference grid. For DEM and nighttime light raster data, the WarpedVRT virtual reprojection method is employed, using the hyperspectral image spatial reference parameters as the target, and achieving pixel-by-pixel spatial matching through coordinate transformation and bilinear resampling. Road vector data is first transformed to the reference coordinate system, and then a binary road mask is generated using a rasterization method, where road-covered pixels are assigned a value of 1, and other pixels are assigned a value of 0, with the alltouched rule ensuring the complete preservation of linear road features. All input raster data undergoes invalid value processing, converting NoData, Inf, and non-finite values ​​to NaN, and adopting a NaN ignoring strategy in subsequent statistical calculations.

[0136] Six key bands—blue, green, red, near-infrared, shortwave infrared 1, and shortwave infrared 2—were selected from hyperspectral imagery as basic spectral features. Spectral indices such as NDVI, SAVI, EVI, GNDVI, and NDWI were further constructed. All ratio calculations included the addition of [missing information - likely a typo, should be 0] to the denominator. ε =10⁻⁶ to avoid numerical anomalies. Based on the complete hyperspectral band sequence, the first derivative D1 is calculated along the band direction using the numerical gradient method, and the second derivative D2 is obtained by further differentiating D1. Eight statistical features, including mean, standard deviation, maximum value, and minimum value, are extracted to describe the slope and curvature changes of the spectral curve.

[0137] In the absorption feature extraction process, the linear continuum construction method at the beginning and end of the spectrum is adopted. Based on the reflectance of the first and last bands of the spectral interval, the continuum value is obtained through linear interpolation. C i And calculate the difference absorption curve between the continuum and the original spectrum. A i = C i - R i Based on the absorption curves, parameters such as maximum absorption depth, cumulative absorption area, maximum absorption location, and absorption width are calculated as spectral absorption characteristics. Based on the matched DEM data, slope and aspect information are obtained by calculating the elevation gradient, and the local average elevation is calculated using an 11×11 neighborhood window. A humidity proxy variable is constructed using the local elevation difference, and the TWI index is simplified by combining it with slope calculations. For nighttime lights, the spatially registered brightness value is used as an indicator of human activity intensity. Road data is filtered using a 31×31 window mean to calculate the road pixel ratio, characterizing road distribution density and the degree of human interference.

[0138] Finally, the latitude and longitude of the soil sampling points are converted into a projection coordinate system consistent with the hyperspectral image. The row and column index of the raster where the sampling point is located is determined by the inverse operation of affine transformation, and the feature value of the corresponding pixel is read. Finally, a structured sample point feature set containing spectral, topographic and social environmental factors is formed for soil heavy metal pollution inversion modeling.

[0139] Step S6: Risk Model Construction and Assessment After obtaining the sample feature extraction data, samples with missing PN values ​​were first removed, and the missing values ​​were imputed using the median of each feature. Features with a missing rate not exceeding 30% were retained. Subsequently, low-variance features with variances no greater than 1×10⁻⁶ were removed using variance screening, and redundant features with absolute correlation coefficients greater than 0.90 were deleted using the correlation coefficient matrix. Features with absolute correlation coefficients no less than 0.03 with PN were retained. A random forest regression model was constructed using PN as the response variable. Features were ranked according to their importance, and those with the highest importance were selected as input variables for the PN risk identification model. This screening process reduces features with low variation, redundant information, or low correlation with PN, providing a feature foundation for the risk identification model.

[0140] Using a PN value of 2.0 as the risk assessment threshold, samples with a PN value no greater than 2.0 were classified as low-risk samples, and samples with a PN value greater than 2.0 were classified as high-risk samples. Based on the selected features, support vector machine, random forest, and logistic regression classification models were constructed respectively. The support vector machine used a radial basis function kernel function, and the input features were standardized before model training. All three classification models employed class balancing to reduce the impact of differences in the number of samples in different risk categories on model training.

[0141] A grid search method combined with hierarchical five-fold cross-validation was used to optimize the model parameters, and the cross-validation F1 score was used as the model comparison metric. For the support vector machine (SVM), the penalty coefficient C and kernel parameter γ were jointly searched within a preset parameter range. In this embodiment, 144 valid samples were obtained after data cleaning, including 93 low-risk samples and 51 samples with warning risks. The cross-validation F1 scores of SVM, random forest, and logistic regression were 0.531, 0.419, and 0.511, respectively. The cross-validation F1 score of SVM was relatively high, so it was selected as the risk identification model in this embodiment, with corresponding parameters C=2 and γ=0.1.

[0142] The determined support vector machine model was used for out-of-fold cross-validation prediction to obtain the predicted risk category of each sample and its probability of belonging to the warning risk category. The overall model accuracy was 0.674, the F1 score for the warning risk category was 0.535, the macro-average F1 score was 0.642, the Cohen's Kappa coefficient was 0.283, and the area under the receiver operating characteristic curve was 0.608. The confusion matrix results showed that 70 out of 93 low-risk samples were identified as low-risk, and 27 out of 51 warning risk samples were identified as warning risk. These results indicate that under the data and feature conditions of this embodiment, the constructed model can distinguish the PN risk categories to a certain extent and can provide probabilistic outputs for the calculation of the risk warning index.

[0143] Step S7: Construction of Risk Warning Index (RWI) and Risk Zoning During model training, a class balancing approach is used to assign weights to different risk categories. The category weights are automatically determined based on the number of samples in each category, giving higher training weights to the warning risk category, which has a relatively smaller sample size. In this embodiment, there are 93 low-risk samples and 51 warning risk samples, corresponding to category weights of approximately 0.774 and 1.412, respectively. The random forest feature importance calculated during the feature selection stage is only used for candidate feature ranking and selection and is not used as a fixed weighting coefficient for each input feature in the support vector machine model; the relationship between each input feature and the risk category is automatically determined during the support vector machine training process.

[0144] Based on the out-of-range prediction results of support vector machines, the predicted probability of each sample belonging to the warning risk category is extracted and defined as the PN risk warning index (RWI), expressed as: RWI = P(Y=1|F). The RWI value ranges from 0 to 1; the larger the value, the higher the probability that the corresponding sample is identified as a warning risk by the model. Different samples exhibit certain differences in RWI, which can be used to characterize the relative risk level among sample points in the study area and provide a reference for subsequent risk classification, spatial distribution analysis, and key area identification.

[0145] This study uses a recall-constrained method based on receiver operating characteristic (ROC) curves to determine the risk warning index thresholds, dividing the study area into three risk levels: Safe Zone: RWI < Safe Threshold. This area has a low risk of pollution, the risk of missed detection is controllable, and the risk of heavy metal pollution can be largely ruled out. There is no need to prioritize on-site verification.

[0146] Concern Zone: Safety threshold ≤ RWI < Warning threshold. This zone is a risk transition area with high uncertainty regarding pollution levels, requiring inclusion in long-term dynamic monitoring and regular risk tracking.

[0147] Warning Zone: RWI ≥ Warning Threshold. This area has a high risk of pollution and is a high-probability concentration area. It should be a key area for verification, and on-site sampling and detailed investigation should be prioritized.

[0148] In this case, the safe zone has a RWI < 0.317, the zone of interest has a RWI of 0.317 ≤ 0.345, and the warning zone has a RWI ≥ 0.345. Step S8: Risk Warning Index Verification Using the Nemerow Comprehensive Pollution Index (PN) as a reference standard, the RWI was validated and analyzed. The results are as follows: (1) The hierarchical statistical results show that the mean PN value in the warning area is 2.03, which is higher than that in the safe area (1.883); the proportion of samples with PN>2.0 in the warning area is 42.5%, which is higher than the average level of 35.4% in the whole area, showing a certain risk enrichment tendency; (2) ROC analysis results show that RWI has an AUC of 0.608 for identifying samples with PN>2.0, indicating a certain discriminative ability; (3) The consistency analysis results show that as the PN pollution level increases, the RWI value shows a synchronous upward trend and has a certain monotonic response relationship.

[0149] This embodiment verifies the feasibility of a risk warning method based on the fusion of Nemerow index and multi-source spatial features. It can realize the spatial expression and hierarchical warning of heavy metal risks in mining soil, and can provide technical support for the identification, control zoning and auxiliary decision-making of soil pollution risks in mining areas.

[0150] For mining area survey scenarios without ground sampling data, the multi-evidence feature set constructed in this embodiment can be reused, and the basic risk index RWIn can be calculated using an evidence accumulation fusion formula to quickly output the initial risk screening results for the entire region, which can be used to lock in potential high-risk areas and optimize the layout of sampling points. After the sampling data is obtained, the index calibration can be completed through the modeling process of this embodiment, realizing the full-process application of "initial screening-calibration-evaluation".

[0151] Implementation Method 3: This implementation method also proposes a computer device, including a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a method for constructing a multi-evidence fusion remote sensing heavy metal pollution risk index according to any one of Implementation Method 1 to Implementation Method 2.

[0152] Implementation Method 4: This implementation method also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of a multi-evidence fusion remote sensing heavy metal pollution risk index construction method as described in any one of Implementation Methods 1 to 2.

[0153] The specific embodiments of the present invention disclosed above are merely illustrative of the invention. These embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index, characterized in that, The method includes: Obtain heavy metal content data and corresponding spatial coordinates of soil sampling points in the mining area. Calculate the Nemerow Comprehensive Pollution Index (PN) of each sampling point based on the heavy metal content data and soil background value or soil environmental quality standard value. Divide the sampling points into warning risk points and non-warning risk points based on whether the PN is greater than a preset risk threshold. Based on the generation, migration, occurrence and response process of heavy metal pollution in mining area soil, a risk warning feature set with process correlation is constructed. The multi-evidence chain includes pollution source driving evidence layer D, pollution migration regulation evidence layer T, soil occurrence medium evidence layer M, remote sensing spectral response evidence layer S and ecological stress response evidence layer V. The evidence layers are organized in a cascade relationship of F={D→T→M→(S,V)}, where F is the process cascade multi-evidence feature set. When ground sampling data is available, the risk state mapping model is trained using the multi-evidence chain risk warning feature set as the input feature and the risk state of the sample points divided by PN as the target variable. The predicted probability of the sample points belonging to the warning risk state output by the model is defined as the sample risk warning index RWI. When there is no ground sampling data, the risk direction of each evidence layer feature in the multi-evidence chain risk warning feature set is unified, normalized and weighted within the layer in sequence, and then weighted between the layers to obtain the sample-free basic risk warning index RWIn. Based on RWI or RWIn, the mining area is classified and zoned for risk, and risk warning results are output.

2. The method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index according to claim 1, characterized in that, The pollution source driving evidence layer D selects nighttime light brightness and road density as features to characterize the input driving force of heavy metal pollution generated by mining, smelting, and transportation activities. The pollution migration regulation evidence layer T is based on a digital elevation model to extract elevation, slope, aspect, and topographic humidity index, and combines them with water body index to characterize the regulatory effect of topographic relief and hydrological conditions on the migration, dilution, and deposition of pollutants with surface runoff. The soil-borne medium evidence layer M is based on soil indices, iron oxide indices, clay content indices, and alteration indices retrieved from hyperspectral remote sensing images, and is used to characterize the ability of soil active adsorption components to store and accumulate heavy metals. The remote sensing spectral response evidence layer S extracts the original reflectance, statistical characteristics of the first and second derivatives of the spectrum, and morphological parameters of the depth, area, center position, and width of the spectral absorption band in the visible-near-infrared-shortwave infrared band. The evidence layer V for the ecological stress response selects the normalized vegetation index, soil-regulated vegetation index, enhanced vegetation index, green light normalized vegetation index, and red-edged vegetation index to characterize the indirect stress response of heavy metal pollution to vegetation growth.

3. The method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index according to claim 1, characterized in that, When calculating the Nemerow Comprehensive Pollution Index (PN), first follow the formula... Calculate the single-factor pollution index P for each heavy metal element. i ,in, C i The measured content of heavy metal i is given. B i The soil background or standard value for heavy metal i; then according to the formula Calculate the Nemerow composite pollution index, where, P imax This represents the maximum value of the single-factor pollution index. This represents the average value of the single-factor pollution index. The preset risk threshold is set to 2.

0. Sampling points with PN>2.0 are classified as warning risk sampling points, and sampling points with PN≤2.0 are classified as non-warning risk sampling points.

4. The method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index according to claim 1, characterized in that, When training the risk state mapping model, the input features are standardized, and the class weights are set according to the difference in the number of warning risk samples and non-warning risk samples, so that the risk class with fewer samples gets higher training weights. The candidate model is trained and optimized using a hierarchical cross-validation method. The predicted probability that the spatial unit output by the optimized model belongs to the warning risk state is defined as the sampled risk warning index RWI, where the value of RWI ranges from 0 to 1.

5. The method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index according to claim 1, characterized in that, When calculating the no-sample-based risk warning index RWIn, the j-th original feature in the k-th evidence layer is... Normalize: For positive indicators where an increase in value signifies increased risk, according to... Perform positive normalization; For negative indicators where an increase in value indicates a decrease in risk, according to... Perform negative normalization; Calculate the sub-index of the k-th evidence layer: in, For the k-th evidence layer, the j-th indicator is the minimum value of the j-th indicator across the entire study area; For the k-th evidence layer, the j-th indicator is the maximum value of the j-th indicator across the entire study area; The index is the normalized value; For the weights within the layer, This represents the number of features in this layer of evidence; Recalculate ,in, K represents the interlayer weights of the evidence layers, and K is the set of evidence layers that actually participate in the fusion. When data of a certain evidence layer is missing, that evidence layer is omitted from the set K and the weights of the remaining evidence layers are renormalized.

6. The method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index according to claim 5, characterized in that, Intra-layer weights Interlayer weights The entropy weight method can be used to determine the objective value, or the analytic hierarchy process can be used to calibrate the value when there is prior knowledge.

7. The method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index according to claim 1, characterized in that, When performing risk stratification and partitioning based on RWI or RWIn, the recall constraint method based on the receiver operating characteristic curve is used to determine the stratification threshold, and the mining area is divided into: Safe zone: RWI < safety threshold; Concern Area: Safety threshold ≤ RWI < warning threshold; Warning zone: RWI ≥ warning threshold.

8. The method for constructing a multi-evidence fusion-based remote sensing heavy metal pollution risk index according to claim 1, characterized in that, The method also includes verification of RWI: Using PN as a pollution risk reference standard, the mean PN, median PN, proportion of sample points with PN greater than the preset risk threshold, and risk enrichment factor were statistically analyzed under different RWI levels. Correlation analysis or receiver operating characteristic (ROC) curve analysis was used to verify the consistency between RWI and PN risk status.

9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a method for constructing a multi-evidence fusion remote sensing heavy metal pollution risk index according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a method for constructing a multi-evidence fusion remote sensing heavy metal pollution risk index as described in any one of claims 1-8.

Images