A method for monitoring environmental impact of national land space planning

Abstract

The present invention relates to the technical field of land and space planning, and discloses a method for monitoring the environmental impact of land and space planning, comprising the following steps: performing radiation calibration and atmospheric correction on original remote sensing images, converting them into surface reflectance, screening heavy metal sensitive bands and eliminating scattering noise, estimating the noise covariance matrix between bands; constructing an end-member spectral library and optimizing the end-members through non-negative matrix decomposition, inverting abundance using an alternating direction multiplier algorithm and satisfying physical constraints, calculating the covariance of unmixing residuals; and designing a deep neural network model, inputting abundance and reflectance features. The present invention improves the timeliness of pollution diffusion warnings by constructing a multi-source data collaborative governance platform, dynamically tracking pollution migration paths, and using intelligent prediction models and real-time solution algorithms. Furthermore, a spectral analysis method breaks through the spectral interference bottleneck of multi-metal composite pollution, providing decision support for land and space environmental governance.

Description

Technical Field

[0001] The present invention relates to the field of national land space planning, and more specifically, to a method for monitoring the environmental impact of national land space planning. Background Art

[0002] Under the national land space planning system, controlling the migration of heavy metal pollution after the closure of mineral resource development zones is a core challenge in ecological restoration.

[0003] Traditional monitoring methods have many technical bottlenecks:

[0004] Data silos: Soil, groundwater, and remediation project data are scattered across different departments, making it impossible to dynamically track pollution migration paths.

[0005] Traceability dilemma: Data during the repair process is easily tampered with, and a broken supervision chain makes it difficult to determine responsibility;

[0006] Delayed response: Pollution spread prediction relies on manual experience and cannot achieve proactive risk intervention;

[0007] Multi-metal composite pollution interference: In the scenario of multi-metal composite pollution interference, there is multi-metal composite pollution, which causes the spectral absorption peaks to overlap during monitoring;

[0008] The above technical bottlenecks make it impossible for traditional monitoring methods to provide precise supervision tools for the implementation of national land space planning. Therefore, a national land space planning environmental impact monitoring method based on multi-source heterogeneous data fusion is constructed to achieve precise monitoring, dynamic early warning and reliable traceability of heavy metal pollution migration, and solve the spectral overlap interference problem of multi-metal composite pollution. Summary of the Invention

[0009] The present invention provides a method for monitoring the environmental impact of national land space planning, which solves the technical problem in related technologies that heavy metal pollution migration dynamic monitoring and reliable control cannot be driven by multi-source heterogeneous data.

[0010] The present invention provides a method for monitoring the environmental impact of land space planning, comprising the following steps:

[0011] S100 performs radiometric calibration and atmospheric correction on the original remote sensing image, converts it into surface reflectance, filters heavy metal-sensitive bands, eliminates scattered noise, and estimates the noise covariance matrix between bands;

[0012] S200, constructs an endmember spectral library and optimizes the endmembers through non-negative matrix factorization, uses the alternating direction multiplier method algorithm to invert the abundance and meet physical constraints, and calculates the covariance of the unmixing residuals;

[0013] S300: Design a deep neural network model, input abundance and reflectance features; optimize hyperparameters and train the network through Bayesian optimization to output the spatial distribution of heavy metal concentrations;

[0014] S400, generates concentration probability distribution based on Monte Carlo regularization, combines reflectivity noise with abundance residual propagation error, calculates the total standard deviation and generates a 95% confidence interval spatial map;

[0015] S500, precision verification and iterative optimization: Cross-validation is used to calculate the root mean square error and determination coefficient of the predicted concentration, and the accuracy is determined based on the threshold. If it does not meet the standard, the parameters are adjusted, and S300-S400 are optimized cyclically. The final result is output after the standard is met.

[0016] Furthermore, in S100, the following steps are specifically included:

[0017] S110, radiometric calibration: converts the sensor's original digital quantization value into a radiometric brightness value to eliminate sensor hardware errors;

[0018] S120, atmospheric correction: eliminates the effects of atmospheric scattering and absorption on surface reflectivity;

[0019] S130, wavelet threshold denoising: noise is separated by wavelet transform and noise is suppressed by soft threshold function.

[0020] Furthermore, in S200, the following steps are specifically included:

[0021] S210, Hybrid Spectrum Modeling: Construct a mathematical model of multi-metal composite spectra based on a linear mixed model to quantify the contribution of each heavy metal end member to the mixed spectrum;

[0022] S220, Regularization Constraint Construction: Design sparse constraints and spatial smoothness constraints to solve the unmixing problem caused by overlapping absorption peaks;

[0023] S230, ADMM iterative solution: The optimization problem is decomposed using the alternating direction multiplier method, and the abundance, auxiliary variables and Lagrange multipliers are updated alternately.

[0024] Furthermore, the mathematical model of the multi-metal composite spectrum is as follows:

[0025] ,

[0026] in, is the pixel reflectance vector, which is the surface reflectance after denoising output from S130. is the endmember spectrum matrix, is the abundance vector to be solved, is the observation noise vector, is the number of spectral bands, is the number of end members, is the endmember index, is the abundance value of the i-th end member, is the spectral vector of the i-th end member.

[0027] Furthermore, in S300, the following steps are specifically included:

[0028] S310, Build a deep neural network model: Design a fully connected neural network to map abundance vectors and environmental auxiliary variables to heavy metal concentrations;

[0029] S320, Bayesian Optimization Hyperparameters: Optimizing Network Structure and Learning Rate , maximize the Nash efficiency coefficient;

[0030] S330, end-to-end training and prediction: uses the Adam optimizer to minimize the mean square error and output a global concentration distribution map.

[0031] Furthermore, the deep neural network model is as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] ;

[0036] in, is the input vector, with dimension , is the number of environment variables, is the abundance vector with dimension , is the number of end members, is the environment variable vector with dimension , For the layer weight matrices, is the first layer weight matrix, is the first layer bias vector, For the layer bias vector, is the first hidden layer, For the The output of the hidden layer, is the total number of neural network layers, For the The number of neurons in the layer, is the rectified linear unit activation function, defined as , is the predicted heavy metal concentration.

[0037] Furthermore, in S400, the following steps are specifically included:

[0038] S410, Monte Carlo Dropout model perturbation: Enable the Dropout layer in the trained deep neural network model to perform A random forward propagation to generate the probability distribution of concentration prediction;

[0039] S420, Error Propagation Computation: Quantifying the impact of reflectivity noise and abundance unmixing errors on concentration predictions, first-order error propagation based on Taylor expansion;

[0040] S430, Confidence Interval Generation: Integrate the model perturbation and error propagation results to generate a 95% confidence interval for the predicted concentration of each pixel.

[0041] Furthermore, the probability distribution of concentration prediction is as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] in, is the input vector, is the abundance vector, is the environment variable vector, Deep neural network trained for S300, For the The network weights of the Dropout samples, is the number of Monte Carlo sampling, To predict the mean concentration, the range is determined by the specific heavy metal. The model's own variance, For the The predicted concentration of the sub-sampling is determined by the specific heavy metal. is the sampling index, is the number of end members, The number of environment variables.

[0046] Furthermore, the error propagation calculation formula is as follows:

[0047] ;

[0048] in, is the gradient of concentration versus reflectivity, is the reflectivity noise covariance matrix, is the gradient of concentration versus abundance, is the abundance unmixing covariance matrix, is the model's own variance, is the total variance, is the number of spectral bands, is the number of end members, Predicted concentration, is the reflectivity vector, is the abundance vector, where is the reflectivity error, is the abundance error.

[0049] Furthermore, in S500, the following steps are specifically included:

[0050] S510, cross-validation accuracy evaluation: k-fold cross-validation is used to evaluate the concentration prediction accuracy in S300, and the root mean square error and determination coefficient are calculated;

[0051] S520, iterative optimization determination: determine whether the model needs to be optimized based on the accuracy threshold, and adjust key parameters;

[0052] S530, closed-loop iteration and result output: Based on the determination result of S520, the optimization is executed cyclically or the final product is output.

[0053] The beneficial effects of the present invention are:

[0054] By constructing a multi-source data collaborative governance platform, the present invention effectively breaks through the "data island" of land and space environmental monitoring and realizes the dynamic tracking of pollution migration paths; through intelligent prediction models and real-time solution algorithms, it significantly improves the timeliness of pollution spread warnings; at the same time, it successfully breaks through the spectral interference bottleneck of multi-metal composite pollution using spectral analysis methods, providing accurate and efficient decision-making support for land and space environmental governance. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a general flow chart of a method for monitoring environmental impacts of land space planning according to the present invention;

[0056] Figure 2 The present invention Figure 1 Flowchart of S100;

[0057] Figure 3 The present invention Figure 1 Flowchart of S200;

[0058] Figure 4 The present invention Figure 1 Flowchart of S300;

[0059] Figure 5 The present invention Figure 1 Flowchart of S400;

[0060] Figure 6 The present invention Figure 1 Flowchart of S500. DETAILED DESCRIPTION

[0061] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. In addition, features described with respect to some examples may also be combined in other examples.

[0062] like Figures 1-6 As shown, a method for monitoring the environmental impact of land space planning includes the following steps:

[0063] S100, spectral preprocessing and feature extraction: perform radiometric calibration and atmospheric correction on the original remote sensing image, convert it into surface reflectance, screen heavy metal sensitive bands and eliminate scattered noise, and estimate the noise covariance matrix between bands;

[0064] In one embodiment of the present invention, the following steps are specifically included:

[0065] S110, radiometric calibration: converts the sensor's original digital quantization value (DN value) into a radiometric brightness value (physical quantity) to eliminate sensor hardware errors;

[0066] formula: ;

[0067] in, For band The radiance, The original image in the band The pixel value of For band The gain coefficient, For band The offset of is the wavelength;

[0068] S120, atmospheric correction: eliminates the effects of atmospheric scattering (Rayleigh scattering, aerosol scattering) and absorption (water vapor, ozone) on surface reflectivity;

[0069] formula: ;

[0070] in, For band The surface reflectivity, For band The sensor receives the radiance, For band Atmospheric radiation, is the correction factor for the distance between the Sun and the Earth, For band The solar irradiance at the top of the atmosphere is is the solar zenith angle, range , is pi, approximately equal to 3.14159;

[0071] S130, wavelet threshold denoising: separates noise (high-frequency signals) through wavelet transform and suppresses noise using soft threshold function;

[0072] formula:

[0073] 1. Wavelet decomposition:

[0074] ;

[0075] 2. Soft threshold processing:

[0076] ;

[0077] ;

[0078] 3. Wavelet reconstruction:

[0079] ;

[0080] 4. Adaptive threshold:

[0081] ;

[0082] in, is a discrete wavelet transform operator that decomposes the signal into low-frequency and high-frequency components. The input is a reflectivity vector and the output is a coefficient matrix. is the inverse discrete wavelet transform operator, which reconstructs the signal. The input is the coefficient matrix and the output is the reconstructed reflectivity vector. is the approximate coefficient matrix, is the detail coefficient matrix, is the detail coefficient matrix after threshold processing, is the adaptive threshold, For band The noise standard deviation is estimated by the MAD method, is the total number of image pixels, is a sign function, with values ​​of {-1, 0, 1}, is the absolute value operator, is the surface reflectance after denoising;

[0083] S200, multi-metallic abundance unmixing: construct an endmember spectral library and optimize the endmembers through non-negative matrix factorization (NMF), use the ADMM algorithm to invert the abundance and meet physical constraints, and calculate the covariance of the unmixing residuals;

[0084] In one embodiment of the present invention, the following steps are specifically included:

[0085] S210, Hybrid Spectral Modeling: Construct a mathematical model of multimetallic composite spectra based on the linear mixed model (LMM) to quantify the contribution of each heavy metal end member to the hybrid spectrum;

[0086] formula: ;

[0087] in, is the pixel reflectivity vector, output from S130 , is the endmember spectrum matrix, column vector For the Standard spectra of metals, is the abundance vector to be solved, is the observation noise vector, is the number of spectral bands, is the number of end members, is the end member index (in the range of ), is the abundance value of the i-th end member, is the spectral vector of the i-th end member;

[0088] S220, Regularization Constraint Construction: Designing Sparse Constraints ( -norm) and spatial smoothness constraint (TV-norm) to solve the unmixing ill-conditioned problem caused by the overlap of absorption peaks;

[0089] formula: ;

[0090] in, is the sparse regularization strength, is the spatial smoothing intensity, is the spatial gradient operator, For the The abundance plane of the metal, is the image height, is the image width, is the metal type index (in the range of ), for norm, for norm, is the spectral reconstruction residual, To ensure data fidelity, is a sparse constraint, represents the total variation regularization;

[0091] S230, ADMM iterative solution: The optimization problem is decomposed using the alternating direction multiplier method (ADMM), and the abundance, auxiliary variables, and Lagrange multipliers are updated alternately;

[0092] formula:

[0093] 1. Variable splitting: ;

[0094] 2. Augmented Lagrangian function:

[0095] ;

[0096] 3. Alternate update: ;

[0097] ;

[0098] ;

[0099] in, is an auxiliary variable with dimension , is the number of end members, is the Lagrange multiplier with dimension , is the penalty coefficient, is a soft threshold operator, defined as , is the iteration index, is the total variation operator, To minimize the operator, is a symbolic function, is the absolute value operator, is the value of the augmented Lagrangian function, is the variable value of the tth iteration;

[0100] S300, Concentration Nonlinear Mapping: Design a deep neural network (DNN) model, inputting abundance and reflectance features; train the network through Bayesian hyperparameter optimization, and output the spatial distribution of heavy metal concentrations;

[0101] In one embodiment of the present invention, the following steps are specifically included:

[0102] S310, build a deep neural network model: design a fully connected neural network (FCN) to transform the abundance vector ( ) and environmental auxiliary variables (pH, organic matter, etc.) are mapped to heavy metal concentrations;

[0103] formula:

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] in, is the input vector, with dimension , is the number of environment variables, is the abundance vector with dimension , is the number of end members, is the environment variable vector with dimension , For the layer weight matrices, is the first layer weight matrix, is the first layer bias vector, For the layer bias vector, is the first hidden layer, For the The output of the hidden layer, is the total number of neural network layers, For the The number of neurons in the layer, is the rectified linear unit activation function, defined as , is the predicted heavy metal concentration.

[0109] S320, Bayesian optimization hyperparameters: Optimizing network structure (number of layers) , number of neurons ) and the learning rate , maximize the Nash efficiency coefficient (NSE);

[0110] formula:

[0111] 1. Objective function:

[0112] ;

[0113] 2. Gaussian process modeling:

[0114] ;

[0115] 3. Expected improvement criteria:

[0116] ;

[0117] ;

[0118] in, is the hyperparameter combination vector, For the The measured concentration of the samples, For the The predicted concentration of samples, is the mean measured concentration, is the current optimal NSE value, is the number of training samples, is the mean function, is the kernel function, is another value of the hyperparameter vector, To optimize the number of iterations, is the expectation operator, is the sample index (in the range of ), is a Gaussian process, is a Gaussian process function, is the hyperparameter vector for the t+1th iteration;

[0119] S330, end-to-end training and prediction: uses the Adam optimizer to minimize the mean square error (MSE) and outputs a global concentration distribution map;

[0120] formula:

[0121] 1. Loss function:

[0122] ;

[0123] 2. Weight update:

[0124] ;

[0125] 3. Global concentration distribution prediction:

[0126] ;

[0127] in, is the total loss function value, is the L2 regularization coefficient, is the learning rate, is the concatenation vector of all weight matrices, is the L2 norm, is the loss function for the The gradient of the layer weights, is the number of training samples, is the sample index, is the network layer index, is the global concentration distribution matrix, whose dimension is H×W, is the global abundance matrix, is the global environment variable matrix, Input feature matrix for the whole domain;

[0128] S400, uncertainty quantification: Generate concentration probability distribution based on Monte Carlo Dropout (regularization), jointly propagate errors of reflectivity noise and abundance residuals, calculate total standard deviation and generate 95% confidence interval spatial map;

[0129] In one embodiment of the present invention, the following steps are specifically included:

[0130] S410, Monte Carlo Dropout model perturbation: Enable the Dropout layer in the trained DNN (S300) and perform A random forward propagation to generate the probability distribution of concentration prediction;

[0131] formula:

[0132] ;

[0133] ;

[0134] ;

[0135] in,, is the input vector, is the abundance vector, is the environment variable vector, Deep neural network trained for S300, For the The network weights of the Dropout samples, is the number of Monte Carlo sampling, To predict the mean concentration, the range is determined by the specific heavy metal. The model's own variance, For the The predicted concentration of the sub-sampling is determined by the specific heavy metal. is the sampling index (the range is ), is the number of end members, The number of environment variables.

[0136] S420, Error Propagation Calculation: Quantify the impact of reflectivity noise (S100) and abundance unmixing error (S200) on concentration predictions, using first-order error propagation based on Taylor expansion;

[0137] formula:

[0138] ;

[0139] in, is the gradient of concentration versus reflectivity, is the reflectivity noise covariance matrix, is the gradient of concentration versus abundance, is the abundance unmixing covariance matrix, is the model's own variance, is the total variance, is the total standard deviation, is the number of spectral bands, is the number of end members, Predicted concentration, is the reflectivity vector, is the abundance vector, where is the reflectivity error, is the abundance error.

[0140] S430, confidence interval generation: integrating the model perturbation and error propagation results to generate a 95% confidence interval for the predicted concentration of each pixel;

[0141] formula: ;

[0142] in, is the mean concentration, is the total standard deviation, is the 95% confidence interval, ranging from , is the 97.5% quantile of the standard normal distribution;

[0143] S500, accuracy verification and iterative optimization: Cross-validation is used to calculate the RMSE and R² of the predicted concentration. A threshold is used to determine whether the accuracy meets the standard. If not, parameters (such as regularization coefficients or end members) are adjusted and optimization is repeated through S300-S400. Once the standard is met, the final result is output.

[0144] In one embodiment of the present invention, the following steps are specifically included:

[0145] S510, cross-validation accuracy evaluation: k-fold cross-validation (k=5) was used to evaluate the concentration prediction accuracy of S300, and the root mean square error (RMSE) and coefficient of determination (R²) were calculated;

[0146] formula: ;

[0147] ;

[0148] in, is the number of samples in the validation set, which is 20% of the total number of samples, such as 200. For the The measured heavy metal concentrations of samples were For the The predicted concentration of samples, is the mean measured concentration, is the sample index, is the root mean square error, is the coefficient of determination;

[0149] S520, iterative optimization determination: determine whether the model needs to be optimized based on the accuracy threshold, and adjust key parameters;

[0150] Judgment conditions and operations: If , then output the final result, otherwise, adjust the parameters according to the priority and perform the following operations:

[0151] Overfitting: Enter S220 and adjust ;

[0152] Unmixing error: Return to step 210 and optimize the endmember spectrum matrix and weights ;

[0153] Noise impact: Return to step 120 and optimize the adaptive threshold ;

[0154] in, is the RMSE threshold, is the R² threshold, is the logical AND operator;

[0155] S530, closed-loop iteration and result output: Based on the determination result of S520, the optimization is executed cyclically or the final product is output;

[0156] Iteration logic:

[0157] Loop execution: Step 300 (concentration mapping) Step 400 (Uncertainty Quantification) Step 510, until satisfied , or the maximum number of iterations is reached ;

[0158] Output content: Concentration distribution map is (S330);

[0159] It uses the deep neural network trained in S330 Input features for the entire domain The final concentration distribution matrix obtained by prediction has an output dimension of H×W, and each pixel corresponds to a concentration value;

[0160] The uncertainty diagram is (S420 );

[0161] Calculate the total standard deviation of each pixel based on the error propagation formula of S420 , considering the three sources of reflectance error, abundance error and model variance, the output dimension is H×W, and each pixel corresponds to a standard deviation value;

[0162] The confidence interval plot is (S430);

[0163] Accuracy is reported as value (S510);

[0164] in, is the image height, is the image width, is the maximum number of iterations, is the final concentration distribution matrix, is the uncertainty distribution matrix, is a 95% confidence interval, and the two channels represent the upper and lower bounds respectively. is the symbol for the set of real numbers, It is a symbol that represents the dimension and numerical type of the matrix / tensor.

[0165] Based on the above method, the following multi-metal composite pollution spectral analysis process example is given (taking the monitoring of heavy metal pollution such as arsenic (As), lead (Pb), and cadmium (Cd) in a mining area as an example)

[0166] Step 1: Spectral preprocessing and feature extraction: Perform radiometric calibration and atmospheric correction on the mining area hyperspectral image to generate surface reflectance data; filter the characteristic bands of As, Pb, and Cd (such as the absorption peaks of As at 600nm, Pb at 950nm, and Cd at 1450nm), eliminate scattering noise, and estimate the noise covariance.

[0167] The key outputs of spectral preprocessing and feature extraction are shown in Table 1:

[0168] Table 1: Key outputs of spectral preprocessing and feature extraction

[0169]

[0170] Step 2: Multi-metal abundance unmixing: Construct end-member libraries for soil, vegetation, and water bodies and optimize the end-member spectra; use the ADMM algorithm to invert the abundance of As, Pb, and Cd in each pixel (satisfying abundance ≥ 0 and sum = 1), and calculate the covariance of the unmixing residuals.

[0171] The key outputs of polymetallic abundance unmixing are shown in Table 2:

[0172] Table 2: Key outputs of polymetallicity unmixing

[0173]

[0174] Step 3: Concentration nonlinear mapping: Design a DNN model (input: abundance + reflectance features); determine the network structure through Bayesian optimization, and output the concentration prediction map of As, Pb, and Cd after training.

[0175] The key outputs of the concentration nonlinear mapping are shown in Table 3:

[0176] Table 3: Key outputs of concentration nonlinear mapping

[0177]

[0178] Step 4: Uncertainty quantification: Monte Carlo Dropout is used to generate the concentration probability distribution; the total standard deviation is calculated by combining the noise and residuals to generate a 95% confidence interval space map.

[0179] The key outputs of uncertainty quantification are shown in Table 4:

[0180] Table 4: Key outputs of uncertainty quantification

[0181]

[0182] Step 5: Accuracy Verification and Iterative Optimization: Verify accuracy using ground sampling points (calculate RMSE / R²). If the Pb accuracy is found to be substandard, return to step 2, add "mining area pollution end members," and perform unmixing again, ultimately outputting the optimized results.

[0183] The iterative optimization process is shown in Table 5:

[0184] Table 5: Iterative optimization process

[0185]

[0186] Final products: Spatial distribution map of heavy metal concentrations, uncertainty quantification products (standard deviation map + confidence interval), and precision verification report (RMSE and R² for As / Pb / Cd);

[0187] Example summary: This process successfully inverts and verifies polymetallic contamination in mining soils. By iteratively optimizing endmembers and unmixing parameters, the prediction accuracy of Pb is significantly improved. The final output concentration distribution and uncertainty products provide a reliable basis for environmental remediation.

[0188] The above describes the embodiments of the present invention, but the present invention is not limited to the above specific implementation methods. The above specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make many forms under the guidance of the present invention, all of which are protected by the present invention.

Claims

1. A method for monitoring the environmental impact of land space planning, characterized in that: The following steps are involved: S100 performs radiometric calibration and atmospheric correction on the original remote sensing image, converts it into surface reflectance, filters heavy metal-sensitive bands, eliminates scattered noise, and estimates the noise covariance matrix between bands; S200, constructs an endmember spectral library and optimizes the endmembers through non-negative matrix factorization, uses the alternating direction multiplier method algorithm to invert the abundance and meet physical constraints, and calculates the covariance of the unmixing residuals; In S200, the following steps are specifically included: S210, Hybrid Spectrum Modeling: Construct a mathematical model of multi-metal composite spectra based on a linear mixed model to quantify the contribution of each heavy metal end member to the mixed spectrum; Among them, the mathematical model of multi-metal composite spectrum is as follows: ； in, is the pixel reflectance vector, which is the surface reflectance after denoising output from S130. is the endmember spectrum matrix, is the abundance vector to be solved, is the observation noise vector, is the number of spectral bands, is the number of end members, is the endmember index, is the abundance value of the i-th end member, is the spectral vector of the i-th end member; S220, Regularization Constraint Construction: Design sparse constraints and spatial smoothness constraints to solve the unmixing problem caused by overlapping absorption peaks; S230, ADMM iterative solution: The optimization problem is decomposed using the alternating direction multiplier method, and the abundance, auxiliary variables and Lagrange multipliers are updated alternately; S300: Design a deep neural network model, input abundance and reflectance features; optimize hyperparameters and train the network through Bayesian optimization to output the spatial distribution of heavy metal concentrations; S400, generates concentration probability distribution based on Monte Carlo regularization, combines reflectivity noise with abundance residual propagation error, calculates the total standard deviation and generates a 95% confidence interval spatial map; S500, precision verification and iterative optimization: Cross-validation is used to calculate the root mean square error and determination coefficient of the predicted concentration, and the accuracy is determined based on the threshold. If it does not meet the standard, the parameters are adjusted, and S300-S400 are optimized cyclically. The final result is output after the standard is met.

2. A method for monitoring the environmental impact of land space planning according to claim 1, characterized in that: In S100, the following steps are specifically included: S110, radiometric calibration: converts the sensor's original digital quantization value into a radiometric brightness value to eliminate sensor hardware errors; S120, atmospheric correction: eliminates the effects of atmospheric scattering and absorption on surface reflectivity; S130, wavelet threshold denoising: noise is separated by wavelet transform and noise is suppressed by soft threshold function.

3. The method for monitoring the environmental impact of land space planning according to claim 1, characterized in that: In S300, the following steps are specifically included: S310, Build a deep neural network model: Design a fully connected neural network to map abundance vectors and environmental auxiliary variables to heavy metal concentrations; S320, Bayesian Optimization Hyperparameters: Optimizing Network Structure and Learning Rate , maximize the Nash efficiency coefficient; S330, end-to-end training and prediction: uses the Adam optimizer to minimize the mean square error and output a global concentration distribution map.

4. A method for monitoring the environmental impact of land space planning according to claim 3, characterized in that: The deep neural network model is as follows: ； ； ； ； in, is the input vector, with dimension , is the number of environment variables, is the abundance vector with dimension , is the number of end members, is the environment variable vector with dimension , For the layer weight matrices, is the first layer weight matrix, is the first layer bias vector, For the layer bias vector, is the first hidden layer, For the The output of the hidden layer, is the total number of neural network layers, For the The number of neurons in the layer, is the rectified linear unit activation function, defined as , is the predicted heavy metal concentration.

5. The method for monitoring the environmental impact of land space planning according to claim 1, characterized in that: In S400, the following steps are specifically included: S410, Monte Carlo Dropout model perturbation: Enable the Dropout layer in the trained deep neural network model to perform A random forward propagation to generate the probability distribution of concentration prediction; S420, Error Propagation Computation: Quantifying the impact of reflectivity noise and abundance unmixing errors on concentration predictions, first-order error propagation based on Taylor expansion; S430, Confidence Interval Generation: Integrate the model perturbation and error propagation results to generate a 95% confidence interval for the predicted concentration of each pixel.

6. A method for monitoring the environmental impact of land space planning according to claim 5, characterized in that: The probability distribution of concentration prediction is as follows: ； ； ； in, is the input vector, is the abundance vector, is the environment variable vector, Deep neural network trained for S300, For the The network weights of the Dropout samples, is the number of Monte Carlo sampling, To predict the mean concentration, the range is determined by the specific heavy metal. is the model's own variance, For the The predicted concentration of the sub-sampling is determined by the specific heavy metal. is the sampling index, is the number of end members, The number of environment variables.

7. A method for monitoring the environmental impact of land space planning according to claim 6, characterized in that: The error propagation calculation formula is as follows: ； in, is the gradient of concentration versus reflectivity, is the reflectivity noise covariance matrix, is the gradient of concentration versus abundance, is the abundance unmixing covariance matrix, is the model's own variance, is the total variance, is the number of spectral bands, is the number of end members, To predict the concentration, is the reflectivity vector, is the abundance vector, where is the reflectivity error, is the abundance error.

8. The method for monitoring the environmental impact of land space planning according to claim 1, characterized in that: In S500, the following steps are specifically included: S510, cross-validation accuracy evaluation: k-fold cross-validation is used to evaluate the concentration prediction accuracy in S300, and the root mean square error and determination coefficient are calculated; S520, iterative optimization determination: determine whether the model needs to be optimized based on the accuracy threshold, and adjust key parameters; S530, closed-loop iteration and result output: Based on the determination result of S520, the optimization is executed cyclically or the final product is output.

Images