Rapid monitoring method for heavy metals in soil based on spectral analysis

Abstract

This invention discloses a rapid monitoring method for heavy metals in soil based on spectral analysis, belonging to the fields of environmental monitoring and spectral analysis technology. The method includes: establishing a digital spectral sampling plane for initial data acquisition to obtain raw three-dimensional spectral data; separating independent soil matrix, heavy metal absorption, and fluorescence spectral components through data unwrapping; constructing a dynamic sampling cycle decision network to calculate the adaptive sampling frequency for each point using these components as input, thereby generating a dynamically encrypted sampling grid for secondary data acquisition to obtain refined spectral data; and outputting the heavy metal type and content distribution through a multispectral collaborative analysis network. This method effectively eliminates matrix interference through signal decoupling and dynamically optimizes the sampling layout based on spectral characteristics, achieving high-precision and high-efficiency monitoring of heavy metals in soil.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring and spectral analysis technology, specifically a rapid monitoring method for heavy metals in soil based on spectral analysis. Background Technology

[0002] Current mainstream methods for monitoring heavy metals in soil based on spectroscopic techniques typically involve directly modeling and analyzing the raw, mixed spectral data. This raw spectral data is a superposition of signals from various physicochemical processes, including soil matrix scattering, heavy metal absorption, and potential fluorescence effects. Due to the significant spatial heterogeneity of soil matrix factors such as mineral composition, organic matter content, and moisture, its spectral characteristics often mask or interfere with the weak absorption characteristics of trace heavy metals, leading to unstable model prediction accuracy and insufficient detection capability for low-concentration heavy metals. Furthermore, to obtain regionally representative data, existing technologies generally employ pre-defined fixed-density grids for uniform sampling. This approach fails to consider the potential spatial unevenness of heavy metal pollution within the monitoring area, easily resulting in insufficient sampling in critical polluted areas leading to omissions, or excessive sampling in unpolluted areas causing resource waste and inefficiency.

[0003] Effectively extracting features that purely reflect heavy metal information from complex, superimposed spectral signals is a key bottleneck in improving monitoring accuracy. Simultaneously, enabling sampling strategies to intelligently respond to the actual spatial distribution differences of soil pollution presents another challenge for achieving efficient and accurate monitoring. Existing uniform sampling and direct spectral analysis techniques struggle to simultaneously address the two core issues of spectral signal decoupling and optimized allocation of sampling resources. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] Therefore, this invention proposes a rapid monitoring method for heavy metals in soil based on spectral analysis, comprising:

[0006] A digital spectral sampling plane is established for the soil area to be tested, and equally spaced sampling coordinate points are generated;

[0007] The drive spectral acquisition device is moved sequentially to each sampling coordinate point of the digital spectral sampling plane, and multi-mode spectral acquisition is performed at each sampling coordinate point to obtain the original three-dimensional spectral dataset.

[0008] The original three-dimensional spectral dataset is subjected to spectral data unwrapping operation to separate the matrix spectral components that independently characterize the soil matrix properties, the absorption spectral components that independently characterize the heavy metal absorption properties, and the fluorescence spectral components that independently characterize the heavy metal fluorescence properties.

[0009] A dynamic sampling cycle decision network is constructed, and the matrix spectral components, absorption spectral components, and fluorescence spectral components are used as inputs to the dynamic sampling cycle decision network. The adaptive sampling frequency of the current soil monitoring point is calculated through the dynamic sampling cycle decision network.

[0010] Based on the adaptive sampling frequency, the distribution density of sampling coordinate points on the digital spectral sampling plane is re-planned to form a dynamically encrypted sampling grid.

[0011] Secondary spectral acquisition is performed according to the coordinate point layout of the dynamically encrypted sampling grid to obtain a refined three-dimensional spectral dataset;

[0012] The refined three-dimensional spectral dataset is input into a pre-trained multispectral collaborative analysis network to obtain the types and content distribution results of various heavy metal elements in the soil.

[0013] Furthermore, establishing the digital spectral sampling plane includes:

[0014] Use a spatial coordinate calibration system to obtain the boundary contour of the soil area to be tested;

[0015] The region inside the boundary contour is projected onto a horizontal reference plane to form a rectangular reference projection region;

[0016] Using the lower left corner of the reference projection area as the origin, set mutually perpendicular horizontal and vertical coordinate axes;

[0017] According to the preset initial sampling interval, the horizontal and vertical coordinate axes are divided equally to generate an initial set of sampling coordinate points;

[0018] The digital spectral sampling plane is defined by the reference projection area, the origin of the coordinate system, the horizontal coordinate axis, the vertical coordinate axis, and the initial set of sampling coordinate points.

[0019] Furthermore, the multi-mode spectral acquisition performed at each sampling coordinate point includes:

[0020] The original three-dimensional spectral dataset includes near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data;

[0021] At a single sampling coordinate point, the near-infrared light source of the spectral acquisition device is controlled to irradiate the soil surface at a preset intensity, and the diffuse reflection light signal is collected. After spectral dispersion and photoelectric conversion, the near-infrared reflection spectral data is generated.

[0022] At the same sampling coordinate point, switch to mid-infrared light source illumination and adjust the spectral acquisition device to transmission absorption mode to obtain the light signal that passes through the thin layer of soil. After data processing, generate the mid-infrared absorption spectral data.

[0023] While keeping the same sampling coordinate point unchanged, the laser light source is switched again to excite the soil sample to produce Raman scattering and fluorescence effects. Scattered light signals in a specific band are collected, and the Raman scattering spectrum data is generated after analysis.

[0024] The near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data obtained from the same sampling coordinate point are correlated and encapsulated in time sequence to form a three-dimensional spectral data unit of the sampling coordinate point.

[0025] The three-dimensional spectral data units of all sampling coordinate points are arranged and combined sequentially to form the original three-dimensional spectral dataset.

[0026] Furthermore, the spectral data unwrapping operation performed on the original three-dimensional spectral dataset includes:

[0027] Construct an unwrapped network model containing three parallel processing branches, wherein one processing branch is dedicated to extracting the matrix spectral components, another processing branch is dedicated to extracting the absorption spectral components, and the third processing branch is dedicated to extracting the fluorescence spectral components.

[0028] The original three-dimensional spectral dataset is simultaneously input into the three parallel processing branches of the unwrapped network model;

[0029] In the processing branch dedicated to extracting matrix spectral components, a low-frequency feature filter is used to filter the input data, retaining low-frequency spectral features related to soil mineral composition and organic matter content, and outputting the matrix spectral components.

[0030] In the processing branch dedicated to extracting absorption spectral components, a characteristic absorption band recognizer is used to process the input data, lock the spectral band corresponding to the characteristic absorption peak of heavy metal ions and extract its contour information, and output the absorption spectral components.

[0031] In the processing branch dedicated to extracting fluorescence spectral components, a fluorescence peak separator is used to process the input data, remove background fluorescence interference, separate the characteristic fluorescence peak signals generated by specific heavy metal elements, and output the fluorescence spectral components.

[0032] Furthermore, the construction of the dynamic sampling period decision network includes:

[0033] Design a neural network with a multilayer perceptron structure, the neural network including an input layer, intermediate hidden layers and an output layer;

[0034] The matrix spectral components, absorption spectral components, and fluorescence spectral components are vectorized to form matrix feature vectors, absorption feature vectors, and fluorescence feature vectors, respectively.

[0035] The matrix feature vector, absorption feature vector, and fluorescence feature vector are concatenated to form a combined feature vector, and the combined feature vector is input into the input layer of the neural network.

[0036] The intermediate hidden layer performs nonlinear transformation and feature mapping on the input combined feature vector;

[0037] The output layer converts the data processed by the intermediate hidden layer into a scalar value, which is then mapped to the adaptive sampling frequency after undergoing a preset linear transformation.

[0038] Furthermore, the formation of the dynamic encrypted sampling grid includes:

[0039] Obtain the initial set of sampling coordinate points on the digital spectral sampling plane;

[0040] The sampling density adjustment coefficient is calculated based on the adaptive sampling frequency.

[0041] The initial sampling interval is scaled using the sampling density adjustment coefficient to obtain a new dynamic sampling interval;

[0042] Using the new dynamic sampling interval, the grid is re-divided within the reference projection area of ​​the digital spectral sampling plane to generate a new set of sampling coordinate points;

[0043] The non-uniformly distributed grid formed by the new set of sampling coordinate points in space is the dynamic encrypted sampling grid.

[0044] Furthermore, the secondary spectral acquisition according to the coordinate point layout of the dynamically encrypted sampling grid includes:

[0045] The control spectral acquisition device moves into position sequentially according to the order of the new sampling coordinate points in the dynamically encrypted sampling grid;

[0046] At each new sampling coordinate point, the multi-mode spectral acquisition process is repeated to obtain a refined three-dimensional spectral data unit of the sampling coordinate point;

[0047] The refined three-dimensional spectral data units corresponding to all new sampling coordinate points on the dynamic encrypted sampling grid are summarized and arranged in spatial order to construct the refined three-dimensional spectral dataset.

[0048] Furthermore, the multispectral collaborative analysis network includes a deep fusion module for spectral features and a heavy metal analysis module;

[0049] The spectral feature deep fusion module receives the refined three-dimensional spectral dataset and performs cross-modal feature alignment and fusion on the near-infrared reflectance spectral data, mid-infrared absorption spectral data and Raman scattering spectral data in the refined three-dimensional spectral dataset to generate a unified multispectral deep feature map.

[0050] The heavy metal analysis module receives the multispectral depth feature map and extracts the discriminative features associated with different heavy metal elements layer by layer through multiple cascaded analysis layers. Finally, it generates the type and content distribution results of multiple heavy metal elements in the soil in parallel at the output layer.

[0051] Furthermore, the cross-modal feature alignment and fusion of the near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data in the refined three-dimensional spectral dataset includes:

[0052] Perform continuous wavelet transform on the near-infrared reflectance spectral data to generate a near-infrared time-spectrum image;

[0053] The mid-infrared absorption spectrum data is differentiated to enhance the absorption peak characteristics and generate enhanced absorption spectrum data.

[0054] Baseline correction and noise suppression are performed on the Raman scattering spectral data to generate clean Raman spectral data;

[0055] The near-infrared time-spectrum, enhanced absorption spectral data, and pure Raman spectral data are projected into a shared latent feature space, so that the spectral data of different modes have a consistent representation dimension in the shared latent feature space.

[0056] Within the shared latent feature space, the projected features from the three modalities are summed element-wise to complete feature fusion and output the unified multispectral depth feature map.

[0057] Furthermore, the layer-by-layer extraction of discriminative features associated with different heavy metal elements includes:

[0058] The initial analysis layer of the heavy metal analysis module performs a convolution operation on the input multispectral depth feature map to extract basic spectral spatial features.

[0059] Each subsequent parsing layer performs deeper convolution and non-linear activation operations on the feature map output by the previous layer, gradually abstracting higher-level semantic features.

[0060] In the final parsing layer, the network branches into sub-networks equal in number to the number of heavy metal elements to be monitored. Each sub-network focuses on parsing the content information of a specific heavy metal element from the high-level semantic features.

[0061] The outputs of all subnetworks together constitute the results of the type and content distribution of various heavy metal elements in the soil.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] By performing spectral data unwrapping on the original three-dimensional spectral dataset, the previously coupled soil matrix scattering signal, heavy metal absorption signal, and fluorescence emission signal can be separated. This operation allows subsequent analysis to focus on the pure components that independently characterize the absorption and fluorescence properties of heavy metals, eliminating spectral interference from complex soil backgrounds. Directly utilizing these unwrapped independent components for modeling enhances the ability to identify weak characteristic signals of heavy metals, thereby improving the accuracy and reliability of qualitative identification and quantitative inversion of multiple heavy metal elements in soil, especially low-concentration pollutants.

[0064] A dynamic sampling cycle decision network was constructed, using the unwrapped matrix, absorption, and fluorescence spectral components as inputs. This enabled the system to analyze the complexity of the spectral characteristics and the intensity of potential pollution indicators at each initial sampling point in real time. Based on this, the network calculated an adaptive sampling frequency reflecting the required level of attention for that point, driving the sampling system to automatically increase the sampling density in areas with complex features or suspected pollution, while reducing sampling in areas with uniform background and no significant pollution characteristics. This dynamic, dense sampling grid, formed based on real-time feedback of initial spectral information, changes the traditional fixed-grid sampling mode, allowing limited sampling resources to be optimally allocated according to the actual spatial heterogeneity of the monitored objects. While ensuring or even improving overall monitoring accuracy, it reduces unnecessary sampling and data acquisition workload, improving the efficiency of large-scale regional monitoring. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the steps of the rapid monitoring method for heavy metals in soil based on spectral analysis described in this invention.

[0066] Figure 2 A flowchart for performing multi-mode spectral acquisition at each sampling coordinate point;

[0067] Figure 3 A line graph showing the dimensionality distribution of the feature vectors;

[0068] Figure 4 A line graph comparing multi-mode spectral acquisition data;

[0069] Figure 5A line graph showing the weight distribution of feature fusion in multispectral collaborative analysis. Detailed Implementation

[0070] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] See Figure 1 The present invention provides a rapid monitoring method for heavy metals in soil based on spectral analysis. The overall implementation scheme is as follows: For the soil area to be tested, a digital spectral sampling plane is established. This plane defines the two-dimensional spatial reference for sampling operations, and a series of equally spaced sampling coordinate points are generated within the plane. A movable spectral acquisition device is driven to move sequentially to each of the above sampling coordinate points, and multi-mode spectral acquisition is performed at each coordinate point, acquiring a raw three-dimensional spectral dataset containing information on multiple spectral modes.

[0072] A spectral data unwrapping operation was performed on the original 3D spectral dataset to separate the matrix spectral components (independently characterizing soil matrix properties), absorption spectral components (independently characterizing heavy metal absorption properties), and fluorescence spectral components (independently characterizing heavy metal fluorescence properties) from the mixed signal. A dynamic sampling period decision network was constructed, using the separated matrix, absorption, and fluorescence spectral components as inputs. The network calculated and output the adaptive sampling frequency for the current soil monitoring point. Based on the calculated adaptive sampling frequency, the distribution density of sampling coordinate points on the digital spectral sampling plane was replanned, forming a spatially non-uniform dynamic densified sampling grid. A second spectral acquisition was performed according to the coordinate point layout of the dynamic densified sampling grid to obtain a refined 3D spectral dataset. The refined 3D spectral dataset was then input into a pre-trained multispectral collaborative analysis network, which processed and output the type and content distribution results of various heavy metal elements in the soil.

[0073] In one embodiment of the present invention, see [reference] Figure 2In practice, the establishment of the digital spectral sampling plane is achieved through the following method: The spatial coordinate calibration system employs a measuring device combining a global satellite navigation system receiver and a total station. Operators use the measuring device to walk along the boundary of the soil area to be measured or set measuring points, acquiring a series of three-dimensional coordinate data of the boundary points. The processing software within the measuring device connects and fits these boundary points to form the boundary profile of the soil area to be measured. In the measuring device or its connected host computer software, the inner region of the fitted boundary profile is vertically projected onto a virtual horizontal reference plane. This horizontal reference plane is typically defined as the local horizontal plane or a custom reference plane. After projection, a minimum bounding rectangle that completely covers the inner region of the boundary profile is obtained; this rectangle is defined as the reference projection area.

[0074] The software automatically sets the lower left corner of the reference projection area as the origin of the coordinate system. Starting from the origin, it sets the horizontal and vertical coordinate axes along the two adjacent sides of the rectangle. The operator inputs the preset initial sampling interval into the host computer software. The initial sampling interval is a length value preset according to the monitoring requirements. Based on the value of the initial sampling interval, the software divides the length and width of the reference projection area equally along the horizontal and vertical coordinate axes. The intersection of the division points with the coordinate axes is the theoretical sampling position. The set of all these theoretical sampling positions constitutes the initial set of sampling coordinate points. At this point, the digital spectral sampling plane defined by the reference projection area, the origin of the coordinate system, the horizontal coordinate axis, the vertical coordinate axis, and the initial set of sampling coordinate points is constructed in the software.

[0075] In practical implementation, the multi-mode spectral acquisition process at each sampling coordinate point follows a clear timing and mode switching logic. The spectral acquisition device is a mobile platform integrating a near-infrared light source, a mid-infrared light source, a laser light source, and corresponding spectral detectors. In practice, when the spectral acquisition device is controlled to move to a designated sampling coordinate point on the digital spectral sampling plane, the control unit first activates the near-infrared light source. The near-infrared light source illuminates the soil surface below the coordinate point at a preset constant intensity. The diffuse reflection light signal from the soil surface is collected by the light collector and transmitted to the near-infrared spectrometer. The grating within the near-infrared spectrometer splits the light signal, and the array detector receives light intensities of different wavelengths and converts them into electrical signals. After analog-to-digital conversion and preliminary correction, the near-infrared reflectance spectral data for that sampling coordinate point is generated. In some embodiments, the operating wavelength range of the near-infrared light source is set between 780 nm and 2500 nm.

[0076] At the same sampling coordinate point, the control unit turns off the near-infrared light source and switches to the mid-infrared light source. At the same time, the mechanical structure adjusts the optical path of the spectral acquisition device to the transmission absorption mode. The mid-infrared light source illuminates the soil thin-section sample that was collected and prepared at the point in advance. The light signal transmitted through the soil thin-section sample is received by the mid-infrared detector. Due to the absorption of mid-infrared light by the soil components, the signal received by the detector contains absorption characteristics. After processing, the signal generates mid-infrared absorption spectral data.

[0077] While maintaining the relative position of the spectral acquisition device and the sampling coordinates, the control unit switches the light source to a laser source of a specific wavelength. The laser beam excites the soil material to produce Raman scattering and fluorescence effects. The Raman spectrometer in the spectral acquisition device collects the scattered light signal in a specific wavelength band, which is detected by a grating spectral divider and a charge-coupled device detector. The Raman shift and fluorescence intensity are analyzed to generate Raman scattering spectral data. The data management module in the control unit associates the near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data acquired from the same sampling coordinates according to the acquisition time sequence, and encapsulates them into a structured data package containing spatial coordinate identifiers and three types of spectral data blocks. This structured data package constitutes the three-dimensional spectral data unit of the sampling coordinates. The spectral acquisition device traverses all points in the initial set of sampling coordinates on the digital spectral sampling plane, repeating the above acquisition and encapsulation process. Finally, all the obtained three-dimensional spectral data units are arranged and combined according to the spatial order of the sampling coordinates or the acquisition time sequence to form a complete original three-dimensional spectral dataset.

[0078] In some embodiments, the determination of the preset initial sampling interval can take into account the area of ​​the soil region to be tested and the initial resolution requirements of the monitoring. The value of the initial sampling interval is related to the size of the reference projection area. One optional method of determination is to calculate the ratio based on the diagonal length of the reference projection area, and the formula can be expressed as:

[0079]

[0080] in: This represents a calculated reference value. Represents the length of the reference projection area. Represents the width of the reference projection area. This represents a pre-set segmentation coefficient based on experience; the actual initial sampling interval can be taken as... Similar integer values. It can be understood that the above formula provides a quantitative reference method that links region size with sampling density. It can also be understood that the accuracy of the boundary contour obtained by the spatial coordinate calibration system directly affects the accuracy of the reference projection region, and thus affects the spatial correspondence between the initial set of sampling coordinate points and the actual region to be measured.

[0081] Optionally, after generating the initial set of sampling coordinate points, their coordinate information can be imported into the control system of the spectral acquisition device. The control system then generates a movement path plan for the spectral acquisition device based on the coordinate sequence. Optionally, when a sampling coordinate point is inaccessible in the actual terrain, the control software can automatically mark the point as invalid and select a nearby accessible point to replace it, while simultaneously recording coordinate correction information. In practical implementation, the switching between the three modes of multi-mode spectral acquisition requires ensuring the stability of the light source and the collimation of the detector. The optical path switching mechanism inside the spectral acquisition device uses electrically controlled reflectors or rotating filter wheels to ensure that light emitted from different light sources accurately illuminates the same soil micro-area and to ensure the consistency of the field of view of signals received by different detectors.

[0082] In one embodiment of the present invention, the spectral data unwrapping operation is accomplished by deploying a dedicated unwrapping network model. This unwrapping network model is a computational architecture containing three parallel processing branches, which operate independently within the model. The first processing branch is dedicated to extracting matrix spectral components from the mixed spectral data; the second processing branch is dedicated to extracting absorption spectral components; and the third processing branch is dedicated to extracting fluorescence spectral components. In this specific implementation, the original three-dimensional spectral dataset is simultaneously input into the entry points of the three parallel processing branches of the unwrapping network model. Each processing branch receives the same copy of the spectral data and processes it independently. In the first processing branch, dedicated to extracting matrix spectral components, a low-frequency feature filter performs a convolutional filtering operation on the input copy of the original three-dimensional spectral dataset. The low-frequency feature filter is designed based on the broad spectral response characteristics of soil matrix components such as silica, clay minerals, and organic matter. Its core principle is to allow low-frequency spectral information to pass through while suppressing high-frequency noise and sharp features. The transfer function of the low-frequency feature filter can be expressed as:

[0083]

[0084] in: Indicates frequency response, Represents the angular frequency variable. It is a preset cutoff angular frequency. It is an imaginary unit. The low-frequency feature filter processes the frequency domain representation of the spectral data through the above function, retains the low-frequency broad peak spectral features related to soil mineral composition and organic matter content, and finally outputs pure matrix spectral components.

[0085] In some embodiments, the untangled network model is deployed as software on a cloud server or local computing workstation, and the original three-dimensional spectral dataset is transmitted to the computing unit via a network or data cable. In the second processing branch, which is dedicated to extracting absorption spectral components, the feature absorption band recognizer operates on the input spectral data copy. The feature absorption band recognizer has a built-in database of the positions and contours of the characteristic absorption peaks of heavy metal ions. The feature absorption band recognizer performs sliding correlation calculations or second derivative analysis on the input spectral data and the standard absorption bands in the database within a preset wavelength range, locks the concave regions in the spectral curve whose matching degree with the database exceeds a preset threshold, extracts the contour information such as the center wavelength, depth, and half-width at half-maximum of these regions, and reconstructs these feature information into independent absorption spectral components for output. The absorption spectral components mainly contain the characteristic absorption information caused by electronic transitions or molecular bond vibrations of heavy metal ions.

[0086] In practical implementation, in the third processing branch, specifically dedicated to extracting fluorescence spectral components, the fluorescence peak separator processes the input spectral data copy. The separator first estimates and subtracts the broad background fluorescence signal generated by soil organic matter, etc. The background fluorescence signal is estimated using polynomial fitting or an adaptive iterative weighted least squares algorithm. The residual signal after background subtraction contains sharp Raman scattering peaks and possible heavy metal characteristic fluorescence peaks. The fluorescence peak separator further identifies and separates the characteristic fluorescence peak signals generated by laser excitation of specific heavy metal elements from the residual signal through peak shape matching or deconvolution algorithms, ultimately outputting independent fluorescence spectral components. It can be understood that the outputs of the three parallel processing branches—matrix spectral components, absorption spectral components, and fluorescence spectral components—are non-overlapping in the spectral dimensions, each carrying physicochemical information from different sources in the original data.

[0087] Optionally, the low-frequency feature filter can be implemented as a digital finite-length unit impulse response filter or an infinite-length unit impulse response filter, with the cutoff frequency set based on the typical frequency range of soil matrix spectral variations. Optionally, the database in the feature absorption band identifier can be customized and updated and expanded according to the target heavy metal species of the monitoring task. In some embodiments, the fluorescence peak separator performs smoothing preprocessing on the input spectral data before background subtraction to reduce noise interference with background estimation. It can be understood that the purpose of the unwrapping operation is to decouple complex mixed spectral signals into independent components. The three parallel processing branches of the unwrapping network model need to be trained with a large amount of known component spectral data to optimize the parameters within the low-frequency feature filter, feature absorption band identifier, and fluorescence peak separator, ensuring the accuracy and reliability of the unwrapping.

[0088] In one embodiment of the present invention, in a specific implementation, the construction process of the dynamic sampling period decision network begins with the design and implementation of the network structure. The neural network with a multilayer perceptron structure is composed of an input layer, an intermediate hidden layer and an output layer connected in sequence. The number of neurons in the input layer matches the dimension of the input combination feature vector. The intermediate hidden layer contains at least one fully connected layer connected by a nonlinear activation function. The output layer contains a linearly activated neuron for generating scalar values.

[0089] In practical implementation, the matrix spectral components, absorption spectral components, and fluorescence spectral components obtained from the unwrapped network model first need to be converted into vector form. The matrix spectral component, as a two-dimensional data matrix, is expanded into a one-dimensional long sequence and mapped to a fixed-length matrix feature vector through a fully connected coding layer. The absorption spectral components and fluorescence spectral components undergo the same vectorization transformation process, generating absorption feature vectors and fluorescence feature vectors of the same dimension, respectively. The generated matrix feature vector, absorption feature vector, and fluorescence feature vector are then concatenated along their feature dimensions. Concatenation refers to combining the three vectors end-to-end to form a longer vector, which constitutes the combined feature vector.

[0090] In some embodiments, the intermediate hidden layer may contain two fully connected layers. The first fully connected layer performs a dimensionality increase or decrease transformation on the input combined feature vector, and the second fully connected layer further performs a nonlinear mapping on the features. In a specific implementation, the combined feature vector is input into the input layer of the neural network. The input layer distributes the vector data to each neuron in the intermediate hidden layer. Each neuron in the intermediate hidden layer performs a weighted summation of the received data and applies a nonlinear activation function. The nonlinear activation function can be a rectified linear unit function or a hyperbolic tangent function. After the nonlinear transformation and feature mapping of the intermediate hidden layer, the input combined feature vector is transformed into a set of abstract high-level feature representations. The high-level feature representations are passed to the output layer, where a single neuron performs a final linear weighted summation of these high-level features to generate an unscaled original scalar value. In a specific implementation, this original scalar value needs to undergo a preset linear transformation to map to a physically meaningful adaptive sampling frequency. The formula for the linear transformation can be expressed as:

[0091]

[0092] in: The adaptive sampling frequency represents the final output. Represents the original scalar value generated by the output layer. It is a preset scaling factor. It is a preset translation factor, which can be adjusted by changing the scaling factor. Translation coefficient The value can be used to convert the original scalar value. The numerical range is mapped to a frequency range that matches the performance of the actual sampling equipment and the monitoring requirements.

[0093] Optionally, the vectorization transformation of the matrix spectral components, absorption spectral components, and fluorescence spectral components can be implemented using the flattening operation in a convolutional neural network in conjunction with fully connected layers. Optionally, the combined feature vectors can be pre-processed with standardization before being input into the neural network to eliminate differences in the dimensions and numerical ranges of different feature vectors. In some embodiments, dropout layers can be introduced into the intermediate hidden layers to prevent overfitting during neural network training. It is understood that the scaling factor... Translation coefficient The specific values ​​need to be determined along with the internal weights of the neural network during the training phase of the dynamic sampling period decision network. The training process uses historical sampling data and corresponding optimal sampling frequency labels as supervision signals. It can be understood that the overall function of the dynamic sampling period decision network is to map the three independent components characterizing the spectral properties of the soil into a frequency value that determines the spatial density of subsequent sampling. The calculation of the adaptive sampling frequency is entirely data-driven and does not require manual intervention to set fixed rules.

[0094] See Figure 3 This is a line graph showing the dimensionality distribution of feature vectors, corresponding to the "combined feature vector construction" stage in the soil heavy metal spectral monitoring process. It primarily displays the number of dimensions of different feature vectors. The concatenation of three independent feature vectors integrates multimodal information on soil matrix, heavy metal absorption, and fluorescence, providing comprehensive data support for subsequent dynamic sampling frequency calculations. The dimension of the combined feature vectors directly determines the number of neurons in the input layer of the dynamic sampling period decision network, a crucial factor in ensuring network performance. By using fixed-length vector representations, features of different spectral components can be processed by the neural network within a unified dimension, embodying the core design principle of "data-driven sampling frequency."

[0095] In one embodiment of the present invention, the process of forming a dynamic encrypted sampling grid begins with an initial set of sampling coordinate points on the digital spectral sampling plane. The adaptive sampling frequency output by the dynamic sampling period decision network is sent to the grid planning module. The grid planning module calculates the sampling density adjustment coefficient based on the adaptive sampling frequency. The formula for calculating the sampling density adjustment coefficient is as follows:

[0096]

[0097] in: Represents the sampling density adjustment coefficient. The adaptive sampling frequency represents the output of the dynamic sampling period decision network. This represents a preset reference frequency parameter. This is related to the setting of the initial sampling interval. The grid planning module uses the calculated sampling density adjustment coefficient to scale the preset initial sampling interval. The scaling operation is to divide the value of the initial sampling interval by the sampling density adjustment coefficient. This results in a new dynamic sampling interval, the value of which is less than or equal to the initial sampling interval value.

[0098] In practical implementation, the grid planning module uses a new dynamic sampling interval to re-grid the reference projection area of ​​the digital spectral sampling plane. The grid division process starts from the origin and divides the horizontal coordinate axis into equal parts with the new dynamic sampling interval as the step size, and also divides the vertical coordinate axis into equal parts with the same new dynamic sampling interval as the step size. The intersections of the dividing lines and the lines parallel to the coordinate axes constitute a new set of sampling coordinate points. This new set of sampling coordinate points forms a non-uniformly distributed grid in space, i.e., a dynamically densified sampling grid. The point density in different regions of the dynamically densified sampling grid is determined by the corresponding adaptive sampling frequency. Regions with higher adaptive sampling frequencies have a higher sampling density adjustment coefficient. A value greater than 1 results in a new dynamic sampling interval smaller than the initial sampling interval, leading to a denser sampling point density. The grid planning module outputs the coordinate layout information of the dynamically densified sampling grid to the motion control system of the spectral acquisition device.

[0099] In some embodiments, the reference frequency parameter The setting corresponds to the initial spatial sampling frequency represented by the initial sampling interval, so that when the adaptive sampling frequency... Equal to reference frequency parameter At that time, the sampling density adjustment coefficient When the value is 1, the new dynamic sampling interval equals the initial sampling interval. Optional, adaptive sampling frequency. With sampling density adjustment factor The mapping relationship can be defined by a pre-defined lookup table, which stores a series of discrete frequency values ​​and their corresponding adjustment coefficients. See Table 1, which illustrates a possible mapping relationship between adaptive sampling frequency and sampling density adjustment coefficients.

[0100] Table 1: Mapping Table of Adaptive Sampling Frequency and Sampling Density Adjustment Coefficients

[0101]

[0102] It is understandable that using a lookup table can simplify calculations and keep the adjustment coefficients within a certain range, thus avoiding unrealistic extreme sampling intervals.

[0103] In practice, the process of secondary spectral acquisition according to the coordinate point layout of the dynamically encrypted sampling grid is executed by the motion control system. The motion control system reads the new set of sampling coordinate points in the dynamically encrypted sampling grid and plans the optimal movement path of the spectral acquisition device, driving the spectral acquisition device to move to each new sampling coordinate point in sequence. At each new sampling coordinate point, the spectral acquisition device repeatedly executes the multi-mode spectral acquisition process. The multi-mode spectral acquisition process includes switching to a near-infrared light source to irradiate the soil surface to collect diffuse reflection light signals and generate near-infrared reflectance spectral data, switching to a mid-infrared light source to acquire light signals from a thin layer of soil in transmission absorption mode to generate mid-infrared absorption spectral data, and switching to a laser light source to excite the soil sample to collect scattered light signals and generate Raman scattering spectral data. After completing the multi-mode spectral acquisition at each new sampling coordinate point, the control unit encapsulates the acquired near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data into a refined three-dimensional spectral data unit for that point. The refined three-dimensional spectral data unit contains the location identifier of the current sampling coordinate point.

[0104] In some embodiments, the motion control system plans a movement path following the shortest traversal principle to reduce the total acquisition time. Optionally, during the secondary spectral acquisition process, if a new sampling coordinate point cannot be reached due to physical obstacles, the control system can select an alternative point in the local neighborhood of the dynamically encrypted sampling grid according to preset rules and update the coordinate record. After the spectral acquisition device traverses all new sampling coordinate points on the dynamically encrypted sampling grid, the data management module summarizes the refined three-dimensional spectral data units corresponding to all new sampling coordinate points. The summarization process arranges and indexes the data according to the spatial position order of the new sampling coordinate points or the time sequence of acquisition completion, ultimately constructing a structured refined three-dimensional spectral dataset. It can be understood that the spatial sampling density of the refined three-dimensional spectral dataset is higher than that of the initial sampling area in areas with significant potential heavy metal characteristics in the soil. Its data composition is similar to that of the original three-dimensional spectral dataset, but the spatial resolution is adaptively adjusted according to the judgment of the dynamic sampling period decision network.

[0105] See Figure 4This is a line graph comparing multi-mode spectral acquisition data, used to compare the signal intensity performance of the initial sampling point and the denser sampling point in three spectral modes during soil heavy metal spectral monitoring. The overall spectral signal intensity of the denser sampling point is higher than that of the initial sampling point, reflecting the effectiveness of the dynamic dense sampling strategy. In areas with higher heavy metal content or more significant spectral characteristics, the system automatically increases the sampling density, thereby obtaining a stronger spectral response. The near-infrared reflectance spectrum shows the most significant signal improvement, reaching a peak at "dense sampling point 3," indicating that the soil matrix or heavy metal components at this sampling point have the most prominent reflectance characteristics in the near-infrared band. This graph visually verifies the value of the dynamic dense sampling mechanism: by adaptively adjusting the sampling density, it not only improves the quality of the spectral signal but also provides a more reliable data foundation for subsequent multispectral fusion and heavy metal analysis.

[0106] In one embodiment of the present invention, the operation of the multispectral collaborative analysis network begins with the reception and processing of the refined three-dimensional spectral dataset by the spectral feature deep fusion module. The spectral feature deep fusion module first preprocesses and transforms the near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data in the refined three-dimensional spectral dataset to prepare for cross-modal feature alignment and fusion. Specifically, for the near-infrared reflectance spectral data, the spectral feature deep fusion module uses continuous wavelet transform for processing. The continuous wavelet transform uses a mother wavelet function to convolve the spectral signal at multiple scales, generating a near-infrared time-spectrum map that simultaneously reflects the characteristics of the wavelength and scale domains. For the mid-infrared absorption spectral data, the spectral feature deep fusion module performs differential processing, typically using first- or second-order derivative algorithms. By calculating the slope change of the spectral curve at each wavelength point, the contour of the absorption peak is sharpened and the baseline drift effect is partially eliminated, thereby generating enhanced absorption spectral data. For Raman scattering spectral data, the spectral feature deep fusion module performs baseline correction and noise suppression operations. Baseline correction uses an adaptive iterative weighted least squares algorithm to fit and subtract fluorescence background, while noise suppression uses a wavelet threshold denoising method to finally generate clean Raman spectral data.

[0107] The spectral feature deep fusion module internally pre-defines a shared feature latent space, which is an abstract vector space with a fixed dimension. In implementation, the module inputs the generated near-infrared time-frequency spectrogram, enhanced absorption spectrum data, and pure Raman spectrum data into three independent encoder sub-networks. The near-infrared time-frequency spectrogram encoder sub-network typically employs a two-dimensional convolutional neural network structure to extract time-frequency features, the enhanced absorption spectrum data encoder sub-network employs a one-dimensional convolutional neural network structure to extract spectral shape features, and the pure Raman spectrum data encoder sub-network also employs a one-dimensional convolutional neural network structure. The three encoder sub-networks project their respective input data into the shared feature latent space, outputting projected feature vectors of consistent dimension, thus achieving cross-modal feature alignment. Within the shared feature latent space, the spectral feature deep fusion module performs an element-wise weighted summation operation on the projected features from the three modalities. The weighted summation formula is expressed as:

[0108]

[0109] in: This represents the unified multispectral depth feature map after fusion. The projection characteristics of the near-infrared time spectrum. The projection characteristics represent the enhanced absorption spectral data. Projection characteristics representing pure Raman spectral data, , , They are trainable weight coefficients and satisfy... Once the weighted summation is complete, a unified multispectral depth feature map is output to the heavy metal analysis module.

[0110] In some embodiments, the encoder subnetwork learns to map spectral data with different physical meanings to comparable and complementary feature representations during training by constraining a shared feature latent space. It can be understood that the purpose of cross-modal feature alignment and fusion is to integrate different aspects of soil composition and heavy metal information captured by near-infrared reflectance, mid-infrared absorption, and Raman scattering techniques to form a more comprehensive and discriminative feature representation.

[0111] The heavy metal analysis module receives a unified multispectral depth feature map as input and contains multiple cascaded analysis layers. In practice, the initial analysis layer performs a two-dimensional convolution operation on the input unified multispectral depth feature map. The convolution kernel slides along the spatial-spectral dimension of the feature map, extracting basic spectral spatial features such as edges, textures, and preliminary spectral banding patterns. The output feature map of the initial analysis layer is passed to subsequent analysis layers. Each subsequent analysis layer performs deeper convolutions and nonlinear activation operations on the feature map output from the previous layer. Deeper convolutions use smaller kernels or larger receptive fields to capture more complex patterns. The nonlinear activation function is typically a rectified linear unit function. Through processing by multiple cascaded analysis layers, the feature map is gradually abstracted, evolving from basic spectral spatial features into high-level semantic features containing information on the presence and species correlation of heavy metals in the soil.

[0112] Optionally, residual connections or dense connections can be introduced between cascaded parsing layers to facilitate gradient flow and feature reuse. In specific implementations, the network structure branches in the final parsing layer of the heavy metal parsing module. The number of branches is equal to the number of heavy metal elements to be monitored, and each branch constitutes an independent sub-network. Each sub-network focuses on parsing the content information of a specific heavy metal element from the high-level semantic features output by the final parsing layer. The sub-network is usually composed of fully connected layers, and its output is a numerical value representing the predicted content of that heavy metal element. The outputs of all sub-networks are collected in parallel and organized into a vector or list. This vector or list together constitutes the distribution results of the types and contents of various heavy metal elements in the soil, where the type of element corresponds to the sequential index of the sub-network, and the content is represented by the output value of each sub-network.

[0113] In some embodiments, the parsing layer of the heavy metal parsing module may include pooling operations to reduce the spatial dimensionality of the feature map and increase the translation invariance of the features. It can be understood that the heavy metal parsing module extracts features from shallow to deep layers progressively through cascaded convolutional layers, and achieves multi-task parallel parsing through branch sub-networks at the end of the network, ultimately outputting the content information of multiple heavy metals at once.

[0114] See Figure 5This is a line graph showing the weight distribution of feature fusion in multispectral collaborative analysis. It illustrates the weight coefficient distribution of near-infrared, mid-infrared, and Raman spectra when fusing features for five typical soil heavy metals (lead, cadmium, mercury, chromium, and arsenic) in a multispectral collaborative analysis network. The mid-infrared spectral weight reaches its highest value (approximately 0.42) for cadmium (Cd), and also shows significant performance for lead (Pb) and chromium (Cr), making it the spectral mode that contributes the most to most heavy metals. The near-infrared spectral weight is highest for mercury (Hg) (approximately 0.38), making it the second most important source of information overall. The Raman spectral weight is the lowest among all elements (approximately 0.25–0.27), but remains stable, providing supplementary molecular vibrational characteristic information. This graph clearly presents the differences in the discrimination ability of different spectra for different heavy metals, validating the necessity of multispectral collaborative analysis: a single spectrum cannot optimally identify all heavy metals, while a fusion strategy with dynamically allocated weights can integrate the advantages of each spectrum, achieving more accurate prediction of heavy metal content.

[0115] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for rapid monitoring of heavy metals in soil based on spectral analysis, characterized in that, include: A digital spectral sampling plane is established for the soil area to be tested, and equally spaced sampling coordinate points are generated; The drive spectral acquisition device is moved sequentially to each sampling coordinate point of the digital spectral sampling plane, and multi-mode spectral acquisition is performed at each sampling coordinate point to obtain the original three-dimensional spectral dataset. The original three-dimensional spectral dataset is subjected to spectral data unwrapping operation to separate the matrix spectral components that independently characterize the soil matrix properties, the absorption spectral components that independently characterize the heavy metal absorption properties, and the fluorescence spectral components that independently characterize the heavy metal fluorescence properties. A dynamic sampling cycle decision network is constructed, and the matrix spectral components, absorption spectral components, and fluorescence spectral components are used as inputs to the dynamic sampling cycle decision network. The adaptive sampling frequency of the current soil monitoring point is calculated through the dynamic sampling cycle decision network. Based on the adaptive sampling frequency, the distribution density of sampling coordinate points on the digital spectral sampling plane is re-planned to form a dynamically encrypted sampling grid. Secondary spectral acquisition is performed according to the coordinate point layout of the dynamically encrypted sampling grid to obtain a refined three-dimensional spectral dataset; The refined three-dimensional spectral dataset is input into a pre-trained multispectral collaborative analysis network to obtain the types and content distribution results of various heavy metal elements in the soil; The step of performing spectral data unwrapping on the original three-dimensional spectral dataset includes: Construct an unwrapped network model containing three parallel processing branches, wherein one processing branch is dedicated to extracting the matrix spectral components, another processing branch is dedicated to extracting the absorption spectral components, and the third processing branch is dedicated to extracting the fluorescence spectral components. The original three-dimensional spectral dataset is simultaneously input into the three parallel processing branches of the unwrapped network model; In the processing branch dedicated to extracting matrix spectral components, a low-frequency feature filter is used to filter the input data, retaining low-frequency spectral features related to soil mineral composition and organic matter content, and outputting the matrix spectral components. In the processing branch dedicated to extracting absorption spectral components, a characteristic absorption band recognizer is used to process the input data, lock the spectral band corresponding to the characteristic absorption peak of heavy metal ions and extract its contour information, and output the absorption spectral components. In the processing branch dedicated to extracting fluorescence spectral components, a fluorescence peak separator is used to process the input data, remove background fluorescence interference, separate the characteristic fluorescence peak signals generated by specific heavy metal elements, and output the fluorescence spectral components. The formation of the dynamic encrypted sampling grid includes: Obtain the initial set of sampling coordinate points on the digital spectral sampling plane; The sampling density adjustment coefficient is calculated based on the adaptive sampling frequency. The initial sampling interval is scaled using the sampling density adjustment coefficient to obtain a new dynamic sampling interval; Using the new dynamic sampling interval, the grid is re-divided within the reference projection area of ​​the digital spectral sampling plane to generate a new set of sampling coordinate points; The non-uniformly distributed grid formed by the new set of sampling coordinate points in space is the dynamic encrypted sampling grid.

2. The method for rapid monitoring of heavy metals in soil based on spectral analysis as claimed in claim 1, wherein, The establishment of the digital spectral sampling plane includes: Use a spatial coordinate calibration system to obtain the boundary contour of the soil area to be tested; The region inside the boundary contour is projected onto a horizontal reference plane to form a rectangular reference projection region; Using the lower left corner of the reference projection area as the origin, set mutually perpendicular horizontal and vertical coordinate axes; According to the preset initial sampling interval, the horizontal and vertical coordinate axes are divided equally to generate an initial set of sampling coordinate points; The digital spectral sampling plane is defined by the reference projection area, the origin of the coordinate system, the horizontal coordinate axis, the vertical coordinate axis, and the initial set of sampling coordinate points.

3. The method for rapid monitoring of heavy metals in soil based on spectral analysis as claimed in claim 1, wherein, The multi-mode spectral acquisition performed at each sampling coordinate point includes: The original three-dimensional spectral dataset includes near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data; At a single sampling coordinate point, the near-infrared light source of the spectral acquisition device is controlled to irradiate the soil surface at a preset intensity, and the diffuse reflection light signal is collected. After spectral dispersion and photoelectric conversion, the near-infrared reflection spectral data is generated. At the same sampling coordinate point, switch to mid-infrared light source illumination and adjust the spectral acquisition device to transmission absorption mode to obtain the light signal that passes through the thin layer of soil. After data processing, generate the mid-infrared absorption spectral data. While keeping the same sampling coordinate point unchanged, the laser light source is switched again to excite the soil sample to produce Raman scattering and fluorescence effects. Scattered light signals in a specific band are collected, and the Raman scattering spectrum data is generated after analysis. The near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data obtained from the same sampling coordinate point are correlated and encapsulated in time sequence to form a three-dimensional spectral data unit of the sampling coordinate point. The three-dimensional spectral data units of all sampling coordinate points are arranged and combined sequentially to form the original three-dimensional spectral dataset.

4. The method for rapid monitoring of heavy metals in soil based on spectral analysis as claimed in claim 3, wherein, The construction of the dynamic sampling period decision network includes: Design a neural network with a multilayer perceptron structure, the neural network including an input layer, intermediate hidden layers and an output layer; The matrix spectral components, absorption spectral components, and fluorescence spectral components are vectorized to form matrix feature vectors, absorption feature vectors, and fluorescence feature vectors, respectively. The matrix feature vector, absorption feature vector, and fluorescence feature vector are concatenated to form a combined feature vector, and the combined feature vector is input into the input layer of the neural network. The intermediate hidden layer performs nonlinear transformation and feature mapping on the input combined feature vector; The output layer converts the data processed by the intermediate hidden layer into a scalar value, which is then mapped to the adaptive sampling frequency after undergoing a preset linear transformation.

5. The method for rapid monitoring of heavy metals in soil based on spectral analysis as claimed in claim 4, wherein, The secondary spectral acquisition according to the coordinate point layout of the dynamically encrypted sampling grid includes: The control spectral acquisition device moves into position sequentially according to the order of the new sampling coordinate points in the dynamically encrypted sampling grid; At each new sampling coordinate point, the multi-mode spectral acquisition process is repeated to obtain a refined three-dimensional spectral data unit of the sampling coordinate point; The refined three-dimensional spectral data units corresponding to all new sampling coordinate points on the dynamic encrypted sampling grid are summarized and arranged in spatial order to construct the refined three-dimensional spectral dataset.

6. The method for rapid monitoring of heavy metals in soil based on spectral analysis as claimed in claim 1, wherein, The multispectral collaborative analysis network includes a spectral feature deep fusion module and a heavy metal analysis module; The spectral feature deep fusion module receives the refined three-dimensional spectral dataset and performs cross-modal feature alignment and fusion on the near-infrared reflectance spectral data, mid-infrared absorption spectral data and Raman scattering spectral data in the refined three-dimensional spectral dataset to generate a unified multispectral deep feature map. The heavy metal analysis module receives the multispectral depth feature map and extracts the discriminative features associated with different heavy metal elements layer by layer through multiple cascaded analysis layers. Finally, it generates the type and content distribution results of multiple heavy metal elements in the soil in parallel at the output layer.

7. The method for rapid monitoring of heavy metals in soil based on spectral analysis as claimed in claim 6, wherein, The cross-modal feature alignment and fusion of near-infrared reflectance spectral data, mid-infrared absorption spectral data, and Raman scattering spectral data in the refined three-dimensional spectral dataset includes: Perform continuous wavelet transform on the near-infrared reflectance spectral data to generate a near-infrared time-spectrum image; The mid-infrared absorption spectrum data is differentiated to enhance the absorption peak characteristics and generate enhanced absorption spectrum data. Baseline correction and noise suppression are performed on the Raman scattering spectral data to generate clean Raman spectral data; The near-infrared time-spectrum, enhanced absorption spectral data, and pure Raman spectral data are projected into a shared latent feature space, so that the spectral data of different modes have a consistent representation dimension in the shared latent feature space. Within the shared latent feature space, the projected features from the three modalities are summed element-wise to complete feature fusion and output the unified multispectral depth feature map.

8. The method for rapid monitoring of heavy metals in soil based on spectral analysis as claimed in claim 6, wherein, The layer-by-layer extraction of discriminative features associated with different heavy metal elements includes: The initial analysis layer of the heavy metal analysis module performs a convolution operation on the input multispectral depth feature map to extract basic spectral spatial features. Each subsequent parsing layer performs deeper convolution and non-linear activation operations on the feature map output by the previous layer, gradually abstracting higher-level semantic features. In the final parsing layer, the network branches into sub-networks equal in number to the number of heavy metal elements to be monitored. Each sub-network focuses on parsing the content information of a specific heavy metal element from the high-level semantic features. The outputs of all subnetworks together constitute the results of the types and content distribution of various heavy metal elements in the soil.

Images