A soil environment background content statistical method and system based on multi-source data fusion

Abstract

This invention relates to the field of soil environmental technology, specifically to a method and system for statistical analysis of soil environmental background content based on multi-source data fusion. The method includes: S1, multi-source data acquisition: by setting up monitoring points, surface and deep soil samples are collected to obtain information on heavy metal and new pollutant content, physicochemical parameters, soil type, and parent material. Simultaneously, satellite remote sensing imagery, meteorological monitoring data, historical monitoring databases, administrative division data, and land use data are used to form a multi-source dataset; S2, data preprocessing; S3, spatial feature extraction; S4, temporal feature extraction; and S5, deep fusion modeling. Convolutional Neural Networks (CNNs) are used to extract spatial features, Long Short-Term Memory Networks (LSTMs) are used to extract temporal features, and deep fusion is performed using Multilayer Perceptrons (MLPs) to effectively reflect the spatial distribution patterns and temporal trends of soil elements.

Description

Technical Field

[0001] This invention relates to the field of soil environmental technology, specifically to a method and system for statistical analysis of soil environmental background content based on multi-source data fusion. Background Technology

[0002] Soil environmental quality is directly related to ecological security, agricultural production, and public health. With the acceleration of industrialization and urbanization, the pollution problems of heavy metals and emerging pollutants in soil are becoming increasingly prominent, making accurate assessment of soil background concentrations a fundamental task for environmental monitoring and management. Traditional methods for statistically analyzing soil background concentrations mainly rely on fixed-point sampling and single data sources, which have the following shortcomings: traditional methods often rely on manual sampling or a limited number of monitoring points, making it difficult to cover large-scale areas, leading to spatial biases in background concentration estimation; they cannot fully consider the influence of topography, parent material type, land use, and the surrounding environment on soil element distribution. Existing methods are usually based on single or short-term sampling data, making it difficult to reflect the trends, periodic characteristics, and abrupt changes in soil element concentrations over time; they lack systematic analysis of long-term trends and short-term fluctuations, easily leading to inaccurate background concentration estimations. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method and system for statistical analysis of soil environmental background content based on multi-source data fusion.

[0004] This invention provides a method for statistical analysis of soil environmental background content based on multi-source data fusion, comprising the following steps:

[0005] S1. Multi-source data acquisition: By setting up monitoring points, surface and deep soil samples are collected to obtain information on heavy metal and new pollutant content, physicochemical parameters, soil type and parent material. At the same time, satellite remote sensing images, meteorological monitoring data, historical monitoring databases, administrative division data and land use data are used to form a multi-source dataset.

[0006] S2. Data preprocessing: The multi-source data is imputed for missing values, screened for outliers, and standardized and normalized. A unified coding system for the multi-source data is established, and the preprocessed standardized multi-source data is output.

[0007] S3. Spatial feature extraction: The standardized multi-source data is input into a convolutional neural network (CNN). The spectral distribution features, spatial texture features and land cover distribution features of the remote sensing image are extracted in the convolutional layer. The spatial gradient and spatial aggregation features of soil element content in the region are extracted in the pooling layer. The spatial feature vector is output in the fully connected layer.

[0008] S4. Temporal Feature Extraction: Input the spatial feature vector obtained in step S3 and the time series monitoring data into a Long Short-Term Memory (LSTM) network. Extract the trend features, periodic features and mutation point features of soil element concentration over time through memory units. Extract the difference information between long-term trends and short-term fluctuations through a gating structure and output the temporal feature vector.

[0009] S5. Deep Fusion Modeling: The spatial feature vector output in step S3 is concatenated with the temporal feature vector output in step S4, and the feature is fused through a multilayer perceptron to output a multidimensional fused feature representation.

[0010] Preferably, the specific steps of S1 are as follows:

[0011] S1.1 Divide the monitoring area into a preset grid, and select a representative plot in each grid to set up one monitoring point to achieve uniform distribution;

[0012] S1.2 Collect soil physicochemical property data through sensor network, including but not limited to soil pH, soil moisture content, soil electrical conductivity and soil organic matter content;

[0013] S1.3. Obtain the content of heavy metal elements (such as lead, cadmium, mercury, and chromium) and new pollutants (such as zinc, iron, and manganese) in the soil through manual monitoring;

[0014] S1.4 Utilize satellite remote sensing to obtain data such as land cover information, residential areas, roads and industrial and mining pollution sources, and vegetation index;

[0015] S1.5. Soil environmental data of the target area are collected by using a variety of means, such as sensor networks, satellite remote sensing, meteorological monitoring equipment and geographic information systems (GIS), combined with the grid method.

[0016] Preferably, the missing value filling in S2 adopts the K-nearest neighbor interpolation method or the multiple interpolation method, and the anomaly point screening adopts the Z-score method or the interquartile range method.

[0017] Preferably, step S3 specifically includes:

[0018] S3.1, Standardize the multi-source data matrix Input to a convolutional neural network (CNN), where Indicates the number of pixels. This represents the spectral and environmental feature dimensions corresponding to each pixel;

[0019] S3.2 In the convolutional layer, through the convolution kernel Perform convolution operations on the input data:

[0020]

[0021] in, This represents the activation function. Indicates the first The feature maps after layer convolution are used to extract spectral distribution features, spatial texture features, and ground cover distribution features;

[0022] S3.3 In the pooling layer, the convolution results are downsampled, and the spatial gradient and spatial aggregation characteristics of soil element content within the region are calculated:

[0023]

[0024] in, Indicates the number of pixels in the pooling region. Indicates the first The gradient vector of each pixel The mean of the pooling region, Represents spatial gradient characteristics. Indicates spatial clustering characteristics;

[0025] S3.4 In the fully connected layer, convolutional features, spatial gradient features, and spatial clustering features are fused to output a unified spatial feature vector:

[0026]

[0027] in, Flatten represents the feature concatenation operation. Indicates the flattening operation. This represents the nonlinear mapping function of the fully connected layer, and the final output is... This is the spatial feature vector.

[0028] Preferably, step S4 requires converting the spatial feature vector obtained in step S3 into... With time series monitoring data Input to a Long Short-Term Memory (LSTM) network specifically includes:

[0029] S4.1 Input concatenation:

[0030]

[0031] in, This indicates a feature concatenation operation. For a moment The input vector;

[0032] S4.2, LSTM memory cell update:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] in: Forget gate: Used to control the retention of long-term trend characteristics; : Input gate, used to capture new changes in soil element concentration; Candidate memory units; : Memory unit state, including trend characteristics With periodic characteristics ; Output gate, combined with Output the hidden state; :time The hidden state represents the characteristics of local mutation points. S4.3, Integration of Time Features:

[0040]

[0041] in: : Characterizes short-term fluctuations and local changes; The final state of memory units, representing long-term trends and periodicity; Fully connected mapping functions; : The final output time-series feature vector.

[0042] Preferably, step S5 includes the following steps:

[0043] S5.1 Feature splicing:

[0044]

[0045] in: : Represents vector concatenation operation; Z: Joint feature vector input to the multilayer perceptron, containing spatial and temporal features;

[0046] S5.2, Multilayer Perceptron Mapping:

[0047]

[0048]

[0049]

[0050]

[0051] in: : No. Hidden layer output; : No. Layer weights and bias parameters; ReLU is a non-linear activation function.

[0052] S5.1, Feature Fusion Output:

[0053]

[0054] in: The final multi-dimensional fusion feature representation comprehensively includes spatial gradient and clustering features, temporal trend and periodic features, and abrupt change point features.

[0055] Preferably, the multilayer perceptron is used to fuse spatial feature vectors. With time series feature vectors Feature fusion methods include at least one or more of the following:

[0056] 1. Standard Multilayer Perceptron Fusion: Nonlinear feature fusion is achieved by sequentially mapping the input feature vector through multiple fully connected layers;

[0057] 2. Residual connection fusion: Residual connections are added between layers of a multilayer perceptron to directly add the input features to the output of subsequent layers, thereby alleviating gradient vanishing and enhancing feature propagation;

[0058] 3. Attention Mechanism Fusion: Applying attention weights to the concatenated feature vectors. To highlight important features, the formula is as follows:

[0059]

[0060] in, Calculated through an attention network, representing the [number]th [unit]. The importance of each feature;

[0061] 4. Multi-layer structure fusion: Two or more multi-layer perceptron layers can be used to perform deep mapping of input features, thereby improving the expressive power of fused features.

[0062] Preferably, for the multi-dimensional fusion feature representation output by S5, the following operations are also required:

[0063] S6. Point re-evaluation: The fused feature representation obtained in step S5 is combined with statistical methods and machine learning models to identify and classify outliers for each monitoring point. If it is determined to be a high background point, it is retained; if it is determined to be a point of human interference, it is removed.

[0064] S7. Dataset Evaluation and Back Calculation: Based on the effective monitoring points processed in step S6, and combined with the statistical unit division and variation coefficient control method, a background content statistical model is established to back calculate the background content of the soil environment in the target area.

[0065] S8. Dynamic Update: New multi-source data acquired in subsequent monitoring are processed through steps S2 to S7 to update the background content statistics.

[0066] A soil environmental background content statistical system based on multi-source data fusion includes:

[0067] The data acquisition module is used to collect soil, remote sensing, meteorological, and geographic information data.

[0068] The data preprocessing module is used for missing value imputation, outlier screening, format standardization, and normalization.

[0069] The spatial feature extraction module is used to extract spatial texture, spectral distribution, and spatial gradient features based on CNN.

[0070] The time feature extraction module is used to extract trend features, periodic features, and abrupt change features based on LSTM.

[0071] The deep fusion modeling module is used to stitch and fuse spatial and temporal features;

[0072] The point re-evaluation module is used for outlier detection and identification of points with high background / human interference.

[0073] The statistical evaluation module is used to establish a statistical model for background content and perform inverse calculations.

[0074] The dynamic update module is used to receive new data and update model parameters through transfer learning.

[0075] Compared with the prior art, the advantages of this invention are:

[0076] (1) This scheme combines sensor networks, satellite remote sensing, meteorological monitoring, geographic information system (GIS), historical monitoring data and land use information to achieve unified preprocessing and fusion of multi-source data; it uses convolutional neural network (CNN) to extract spatial features and long short-term memory network (LSTM) to extract temporal features, and performs deep fusion through multilayer perceptron (MLP) to effectively reflect the spatial distribution pattern and temporal change trend of soil elements.

[0077] (2) By combining statistical methods with machine learning models, outlier determination and classification of monitoring points can be carried out, and natural high background areas and areas of human interference can be distinguished. By using historical data, on-site verification, cumulative coefficients and pollution source buffers, hierarchical identification and processing of outliers can be achieved, ensuring the scientificity and reliability of statistical results.

[0078] (3) The scheme extracts the spectral distribution, texture and land cover distribution features of remote sensing images in the convolutional layer, and extracts the spatial gradient and aggregation features of soil elements in the pooling layer, making the background content statistics more spatially continuous and regionally representative; the long short-term memory network can capture the trend, periodicity and abrupt change of element concentration over time, and realize the effective analysis of long-term changes and short-term fluctuations. Attached Figure Description

[0079] Figure 1 This is a flowchart of a soil environmental background content statistical method based on multi-source data fusion proposed in this invention. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0081] Example 1:

[0082] Reference Figure 1 A statistical method for soil environmental background content based on multi-source data fusion includes the following steps:

[0083] S1. Multi-source data acquisition: By setting up monitoring points, surface and deep soil samples are collected to obtain information on heavy metal and new pollutant content, physicochemical parameters, soil type and parent material. At the same time, satellite remote sensing images, meteorological monitoring data, historical monitoring databases, administrative division data and land use data are used to form a multi-source dataset.

[0084] The specific steps of S1 are as follows:

[0085] S1.1 Divide the monitoring area into a preset grid, and select a representative plot in each grid to set up one monitoring point to achieve uniform distribution;

[0086] S1.2 Collect soil physicochemical property data through sensor network, including but not limited to soil pH, soil moisture content, soil electrical conductivity and soil organic matter content;

[0087] S1.3. Obtain the content of heavy metal elements (such as lead, cadmium, mercury, and chromium) and new pollutants (such as zinc, iron, and manganese) in the soil through manual monitoring;

[0088] S1.4 Utilize satellite remote sensing to obtain data such as land cover information, residential areas, roads and industrial and mining pollution sources, and vegetation index;

[0089] S1.5. Soil environmental data of the target area are collected by using a variety of means, such as sensor networks, satellite remote sensing, meteorological monitoring equipment and geographic information systems (GIS), combined with the grid method.

[0090] S2. Data preprocessing: The multi-source data is imputed for missing values, screened for outliers, standardized and normalized, and a unified coding system for multi-source data is established to output the preprocessed standardized multi-source data. In S2, missing value imputation adopts K-nearest neighbor interpolation or multiple interpolation, and outlier screening adopts Z-score method or interquartile range method.

[0091] The steps in S3 specifically include:

[0092] S3.1, Standardize the multi-source data matrix Input to a convolutional neural network (CNN), where Indicates the number of pixels. This represents the spectral and environmental feature dimensions corresponding to each pixel;

[0093] S3.2 In the convolutional layer, through the convolution kernel Perform convolution operations on the input data:

[0094]

[0095] in, This represents the activation function. Indicates the first The feature maps after layer convolution are used to extract spectral distribution features, spatial texture features, and ground cover distribution features;

[0096] S3.3 In the pooling layer, the convolution results are downsampled, and the spatial gradient and spatial aggregation characteristics of soil element content within the region are calculated:

[0097]

[0098] in, Indicates the number of pixels in the pooling region. Indicates the first The gradient vector of each pixel The mean of the pooling region, Represents spatial gradient characteristics. Indicates spatial clustering characteristics;

[0099] S3.4 In the fully connected layer, convolutional features, spatial gradient features, and spatial clustering features are fused to output a unified spatial feature vector:

[0100]

[0101] in, Flatten represents the feature concatenation operation. Indicates the flattening operation. This represents the nonlinear mapping function of the fully connected layer, and the final output is... This is the spatial feature vector.

[0102] S3. Spatial feature extraction: The standardized multi-source data is input into a convolutional neural network (CNN). The spectral distribution features, spatial texture features and land cover distribution features of the remote sensing image are extracted in the convolutional layer. The spatial gradient and spatial aggregation features of soil element content in the region are extracted in the pooling layer. The spatial feature vector is output in the fully connected layer.

[0103] S4. Temporal Feature Extraction: Input the spatial feature vector obtained in step S3 and the time series monitoring data into a Long Short-Term Memory (LSTM) network. The network extracts the trend, periodicity, and abrupt change features of soil element concentration over time through memory units. A gating structure is used to extract the difference between long-term trends and short-term fluctuations, outputting a time series feature vector. In step S4, the spatial feature vector obtained in step S3 needs to be... With time series monitoring data Input to a Long Short-Term Memory (LSTM) network specifically includes:

[0104] S4.1 Input concatenation:

[0105]

[0106] in, This indicates a feature concatenation operation. For a moment The input vector;

[0107] S4.2, LSTM memory cell update:

[0108]

[0109]

[0110]

[0111]

[0112]

[0113]

[0114] in: Forget gate: Used to control the retention of long-term trend characteristics; : Input gate, used to capture new changes in soil element concentration; Candidate memory units; : Memory unit state, including trend characteristics With periodic characteristics ; Output gate, combined with Output the hidden state; :time The hidden state represents the characteristics of local mutation points. S4.3, Integration of Time Features:

[0115]

[0116] in: : Characterizes short-term fluctuations and local changes; The final state of memory units, representing long-term trends and periodicity; Fully connected mapping functions; : The final output time-series feature vector.

[0117] S5. Deep Fusion Modeling: The spatial feature vector output in step S3 is concatenated with the temporal feature vector output in step S4, and the feature is fused through a multilayer perceptron to output a multidimensional fused feature representation.

[0118] S5 includes the following steps:

[0119] S5.1 Feature splicing:

[0120]

[0121] in: : Represents vector concatenation operation; Z: Joint feature vector input to the multilayer perceptron, containing spatial and temporal features;

[0122] S5.2, Multilayer Perceptron Mapping:

[0123]

[0124]

[0125]

[0126]

[0127] in: : No. Hidden layer output; : No. Layer weights and bias parameters; ReLU is a non-linear activation function.

[0128] S5.1, Feature Fusion Output:

[0129]

[0130] in: The final multi-dimensional fusion feature representation comprehensively includes spatial gradient and clustering features, temporal trend and periodic features, and abrupt change point features.

[0131] The multilayer perceptron is used to fuse spatial feature vectors. With time series feature vectors Feature fusion methods include at least one or more of the following:

[0132] 1. Standard Multilayer Perceptron Fusion: Nonlinear feature fusion is achieved by sequentially mapping the input feature vector through multiple fully connected layers;

[0133] 2. Residual connection fusion: Residual connections are added between layers of a multilayer perceptron to directly add the input features to the output of subsequent layers, thereby alleviating gradient vanishing and enhancing feature propagation;

[0134] 3. Attention Mechanism Fusion: Applying attention weights to the concatenated feature vectors. To highlight important features, the formula is as follows:

[0135]

[0136] in, Calculated through an attention network, representing the [number]th [unit]. The importance of each feature;

[0137] 4. Multi-layer structure fusion: Two or more multi-layer perceptron layers can be used to perform deep mapping of input features, thereby improving the expressive power of fused features.

[0138] S6. Location Re-evaluation: The fused feature representation obtained in step S5 is combined with statistical methods and machine learning models to identify and classify outliers for each monitoring point. Points identified as having high background are retained, while points identified as being caused by human interference are removed. The specific steps are as follows:

[0139] S6.1 Preliminary outlier identification: Based on the fused feature vector obtained in step S5... For each monitoring point, the degree of deviation was calculated using statistical methods (Z-score, interquartile range).

[0140]

[0141] in, For the first The integrated characteristic index values ​​of each monitoring point (such as element concentration or spatial characteristic comprehensive index). and The mean and standard deviation are for the same region; if If it is, then it is marked as a potential outlier.

[0142] S62, Machine Learning Classification and Probability Definition

[0143] Fusing feature vectors of potential outlier locations Input a pre-trained classification model (such as random forest, support vector machine, lightweight graph neural network), and the model outputs the classification probability for each point:

[0144]

[0145]

[0146] in Both represent the probability that a location is judged as being caused by human interference or having a high background, respectively.

[0147] S6.3, Distinguishing between high background and artificial interference. If it is a high background point, it is determined to be retained; if If a point is identified as a potential source of interference, it will be marked as such and proceed to the subsequent verification and removal steps. Spatial clustering analysis and comparison with neighboring points can serve as input features for the model, helping to improve the accuracy of the judgment.

[0148] S6.4 Historical Data Verification and On-site Verification

[0149] For potential points of human interference, the following factors are considered: historical land use records; historical remote sensing data and monitoring databases; and on-site survey information (whether there are any sudden pollution incidents, sampling operation procedures, etc.) are used to calculate the cumulative coefficient. Determining whether element concentrations are accumulated due to human activities helps in verifying the authenticity of anomalies.

[0150] S6.5 Interference Source Buffer and Hierarchical Identification

[0151] Buffer zones should be established for pollution sources of different types, sizes, and emission intensities. The scope of impact is indicated; based on on-site verification and historical data, areas with small or no substantial interference are not excluded, while points corresponding to medium or small interference areas are excluded; the final set of excluded man-made interference points is obtained.

[0152] S6.6, Output

[0153] Output a set of valid monitoring points, including: retained high-background points; removed points of human interference; outlier determination results and machine learning classification probabilities for each point. and

[0154] S7. Dataset Evaluation and Back Calculation: Based on the effective monitoring points processed in step S6, and combined with the statistical unit division and variation coefficient control method, a background content statistical model is established to back calculate the background content of the soil environment in the target area.

[0155] S7.1 Input of valid monitoring point data: Input the set of valid monitoring points output in step S6. The fused feature vector and corresponding soil element concentration data are input into the background content statistics module. The fused feature vector of each point is denoted as... The soil element concentration vector is denoted as .

[0156] S7.2 Statistical Unit Division: The target area is divided into multiple statistical units (grids or administrative regions), each unit containing several valid monitoring points. Statistical units can be divided based on spatial adjacency, geological type, or land use type to ensure that the background characteristics within the region are relatively consistent.

[0157] S7.3 Calculation of mean background content and coefficient of variation: For each statistical unit, calculate the mean background content of each soil element. With coefficient of variation :

[0158]

[0159] in, For the first The point is the [number]th [location] in this statistical unit. Concentration of elements This represents the standard deviation of the element within the cell.

[0160] S7.4 Outlier Removal and Intra-cell Correction: For points within a cell whose concentration deviates significantly from the mean or whose coefficient of variation is abnormal, removal or correction can be performed again (using the aforementioned cumulative coefficient method or machine learning model to assist) to ensure that the mean background content of the cell is not affected by extreme values.

[0161] S7.5 Background Content Back-Calculation: Based on the mean and coefficient of variation of effective points within the unit, a statistical model of background content is established, and spatial interpolation or regression back-calculation is performed to obtain the distribution of soil background content in the entire target area. For example, the weighted average method or Kriging interpolation method can be used for back-calculation.

[0162]

[0163] in, For position The Predicted background content values ​​of various elements, weights Based on the location and The distance and spatial correlation are determined to satisfy the following conditions. .

[0164] S7.6 Model Accuracy Evaluation and Adjustment: The back-calculation accuracy is evaluated through cross-validation, residual analysis, or coefficient of variation control methods. If the prediction error exceeds the preset threshold, the statistical unit division or weighting function can be adjusted to ensure the stability and reliability of the background content prediction. S7.7 Output: Output a distribution map and statistical table of soil element background content for the entire target area, for subsequent dynamic updates in S8.

[0165] S8. Dynamic Update: New multi-source data acquired in subsequent monitoring are processed through steps S2 to S7 to update the background content statistics.

[0166] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0167] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, 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. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for statistical analysis of soil environmental background content based on multi-source data fusion, characterized in that, Includes the following steps: S1. Multi-source data acquisition: By setting up monitoring points, surface and deep soil samples are collected to obtain information on heavy metal and new pollutant content, physicochemical parameters, soil type and parent material. At the same time, satellite remote sensing images, meteorological monitoring data, historical monitoring databases, administrative division data and land use data are used to form a multi-source dataset. S2. Data preprocessing: The multi-source data is imputed for missing values, screened for outliers, standardized and normalized, and a unified coding system for multi-source data is established to output the preprocessed standardized multi-source data. S3. Spatial feature extraction: The standardized multi-source data is input into a convolutional neural network. The spectral distribution features, spatial texture features and land cover distribution features of the remote sensing image are extracted in the convolutional layer. The spatial gradient and spatial aggregation features of soil element content in the region are extracted in the pooling layer. The spatial feature vector is output in the fully connected layer. S4. Temporal Feature Extraction: Input the spatial feature vector obtained in step S3 and the time series monitoring data into the Long Short-Term Memory Network. Extract the trend features, periodic features and mutation point features of soil element concentration over time through memory units. Extract the difference information between long-term trends and short-term fluctuations through the gating structure and output the temporal feature vector. S5. Deep Fusion Modeling: The spatial feature vector output in step S3 is concatenated with the temporal feature vector output in step S4, and feature fusion is performed through a multilayer perceptron to output a multi-dimensional fused feature representation. S6. Point re-evaluation: The fused feature representation obtained in step S5 is combined with statistical methods and machine learning models to identify and classify outliers for each monitoring point. If it is determined to be a high background point, it is retained; if it is determined to be a point of human interference, it is removed. S6.1 Preliminary outlier identification: The fused feature vector obtained in step S5... For each monitoring point in the data, the degree of deviation is calculated using statistical methods; S6.2 Machine Learning Classification and Probability Definition: Fusing feature vectors of potential outlier locations Input the trained classification model, and the model outputs the classification probability for each location: ; ; in These two terms represent the probability that a location is identified as being subject to human interference or having a high background, respectively. S6.3, Distinguishing between high background and artificial interference: If If so, it is determined to be a high background point and retained; like If it is, it will be marked as a potential point of human interference and will proceed to the subsequent verification and elimination steps. S6.4 Historical Data Verification and On-site Verification: For potential points of human interference, combine: historical land use records; historical remote sensing data and monitoring databases; and on-site survey information through cumulative coefficients. Determining whether element concentrations are accumulated due to human activities helps in verifying the authenticity of anomalies; S6.5 Interference Source Buffer Zones and Hierarchical Identification: Establish buffer zones for pollution sources of different types, sizes, and emission intensities. Indicate the scope of impact; based on on-site verification and historical data, areas with small or no substantial interference are not excluded, while corresponding points in medium-sized or interference areas are excluded. The final set of removed artificial interference points is obtained; S6.6 Output: Output the set of valid monitoring points, including: retained high background points; Human-caused interference points were removed; outlier identification results and machine learning classification probabilities for each point. and ; S7. Dataset Evaluation and Back Calculation: Based on the effective monitoring points processed in step S6, and combined with the statistical unit division and variation coefficient control method, a background content statistical model is established to back calculate the background content of the soil environment in the target area. S7.1 Input of valid monitoring point data: Input the set of valid monitoring points output in step S6. The fused feature vector and the corresponding soil element concentration data are input into the background content statistics module; S7.2 Statistical Unit Division: Divide the target area into multiple statistical units; S7.3 Calculation of mean background content and coefficient of variation: For each statistical unit, calculate the mean background content of each soil element. With coefficient of variation ; S7.4 Outlier Removal and Intra-cell Correction: Points within a cell whose concentration deviates significantly from the mean or whose coefficient of variation is abnormal can be removed or corrected again. S7.5 Background content back calculation: Based on the mean and coefficient of variation of effective points within the unit, establish a statistical model of background content, perform spatial interpolation or regression back calculation, and obtain the distribution of soil background content in the entire target area. S7.6 Model Accuracy Evaluation and Adjustment: Evaluate the back-calculation accuracy. If the prediction error exceeds the preset threshold, adjust the statistical unit division or weighting function. S7.7 Output: Output a distribution map and statistical table of soil element background content for the entire target area, for subsequent dynamic updates in S8; S8. Dynamic Update: New multi-source data acquired in subsequent monitoring are processed through steps S2 to S7 to update the background content statistics.

2. The method for statistical analysis of soil environmental background content based on multi-source data fusion according to claim 1, characterized in that, The specific steps of S1 are as follows: S1.1 Divide the monitoring area into a preset grid, and select a representative plot in each grid to set up one monitoring point to achieve uniform distribution; S1.2 Collect soil physicochemical property data through sensor networks, including but not limited to soil pH, soil moisture content, soil electrical conductivity and soil organic matter content; S1.

3. Obtain the content of heavy metal elements and new pollutants in the soil through manual monitoring; S1.4 Utilize satellite remote sensing to obtain information on land cover, residential areas, roads, industrial and mining pollution sources, and vegetation index data; S1.

5. Soil environmental data of the target area are collected by using a variety of means, including sensor networks, satellite remote sensing, meteorological monitoring equipment and geographic information systems, combined with the grid method.

3. The method for statistical analysis of soil environmental background content based on multi-source data fusion according to claim 1, characterized in that, The missing value filling in S2 adopts the K-nearest neighbor interpolation method or the multiple interpolation method, and the anomaly point screening adopts the Z-score method or the interquartile range method.

4. The method for statistical analysis of soil environmental background content based on multi-source data fusion according to claim 1, characterized in that, The steps in S3 specifically include: S3.1, Standardize the multi-source data matrix Input to a convolutional neural network, where Indicates the number of pixels. This represents the spectral and environmental feature dimensions corresponding to each pixel; S3.2 In the convolutional layer, through the convolution kernel Perform convolution operations on the input data: ; in, This represents the activation function. Indicates the first The feature maps after layer convolution are used to extract spectral distribution features, spatial texture features, and ground cover distribution features; S3.3 In the pooling layer, the convolution results are downsampled, and the spatial gradient and spatial aggregation characteristics of soil element content within the region are calculated: ; in, Indicates the number of pixels in the pooling region. Indicates the first The gradient vector of each pixel The mean of the pooling region, Represents spatial gradient characteristics. Indicates spatial clustering characteristics; S3.4 In the fully connected layer, convolutional features, spatial gradient features, and spatial clustering features are fused to output a unified spatial feature vector: ; in, Flatten represents the feature concatenation operation. Indicates the flattening operation. This represents the nonlinear mapping function of the fully connected layer, and the final output is... This is the spatial feature vector.

5. The method for statistical analysis of soil environmental background content based on multi-source data fusion according to claim 1, characterized in that, Step S4 requires the spatial feature vector obtained in step S3 to be... With time series monitoring data Input to a Long Short-Term Memory (LSTM) network, specifically including: S4.1 Input concatenation: ; in, This indicates a feature concatenation operation. For a moment The input vector; S4.2, LSTM memory cell update: ; in: Forget gate: Used to control the retention of long-term trend characteristics; : Input gate, used to capture new changes in soil element concentration; Candidate memory units; : Memory unit state, including trend characteristics With periodic characteristics ; Output gate, combined with Output the hidden state; :time The hidden state represents the characteristics of local mutation points. S4.3, Integration of Time Features: ; in: : Characterizes short-term fluctuations and local changes; The final state of memory units, representing long-term trends and periodicity; Fully connected mapping functions; : The final output time-series feature vector.

6. The method for statistical analysis of soil environmental background content based on multi-source data fusion according to claim 5, characterized in that, S5 includes the following steps: S5.1 Feature splicing: ; in: : Represents vector concatenation operation; Z: Joint feature vector input to the multilayer perceptron, containing spatial and temporal features; S5.2, Multilayer Perceptron Mapping: ; in: : No. Hidden layer output; : No. Layer weights and bias parameters; ReLU is a non-linear activation function. S5.1, Feature Fusion Output: ; in: The final multi-dimensional fusion feature representation comprehensively includes spatial gradient and clustering features, temporal trend and periodic features, and abrupt change point features.

7. The method for statistical analysis of soil environmental background content based on multi-source data fusion according to claim 6, characterized in that, The multilayer perceptron is used to fuse spatial feature vectors. With time series feature vectors Feature fusion methods include at least one or more of the following: Standard multilayer perceptron fusion: Nonlinear feature fusion is achieved by sequentially mapping the input feature vector through multiple fully connected layers; Residual connection fusion: Residual connections are added between layers of a multilayer perceptron to directly add the input features to the output of the subsequent layers, thereby alleviating gradient vanishing and enhancing feature propagation; Attention mechanism fusion: Applying attention weights to the concatenated feature vectors To highlight important features, the formula is as follows: ; in, Calculated through an attention network, representing the [number]th [unit]. The importance of each feature; Multi-layer structure fusion: Two or more multi-layer perceptron layers can be used to perform deep mapping of input features, thereby improving the expressive power of fused features.

8. A soil environmental background content statistical system based on multi-source data fusion, designed based on the soil environmental background content statistical method based on multi-source data fusion as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect soil, remote sensing, meteorological, and geographic information data. The data preprocessing module is used for missing value imputation, outlier screening, format standardization, and normalization. The spatial feature extraction module is used to extract spatial texture, spectral distribution, and spatial gradient features based on CNN. The time feature extraction module is used to extract trend features, periodic features, and abrupt change features based on LSTM. The deep fusion modeling module is used to stitch and fuse spatial and temporal features; The point re-evaluation module is used for outlier detection and identification of points with high background / human interference. The statistical evaluation module is used to establish a statistical model for background content and perform inverse calculations. The dynamic update module is used to receive new data and update model parameters through transfer learning.

Images