Hyperspectral-VOCs underway combined traceability method based on deep learning

Abstract

The invention relates to the technical field of hyper-spectral satellite remote sensing, in particular to a hyper-spectral-VOCs underway combined tracing method based on deep learning, which screens abnormal concentration through underway monitoring, extracts multiple indexes to judge pollution phase change, identifies boundaries in combination with remote sensing images, extracts pixel identification components and calculates probability to screen a tracing emission source. According to the method, space-time linkage screening is carried out on VOCs concentration threshold exceeding behaviors by constructing a concentration abnormal response task flow, a high-sensitivity mark of pollution response is effectively formed, and on the basis of conjoint analysis of three groups of indexes of concentration time sequence difference, space standard deviation and local fluctuation ratio, by means of standardization processing and jump trend identification means, the VOCs concentration threshold exceeding behaviors are identified, and the VOCs concentration threshold exceeding behaviors are identified. And multi-dimensional dynamic discrimination of the pollution phase change phenomenon is realized, so that the response capability of sudden or hidden emission events is enhanced. The boundary of a suspected pollution area in a remote sensing image is finely depicted through a gray gradient change trend, and the stability of the boundary is judged in combination with time sequence displacement and local gradient density change, so that artifact disturbance factors are effectively eliminated.

Description

Technical Field

[0001] The present invention relates to the field of hyperspectral satellite remote sensing technology, and in particular to a hyperspectral-VOCs cruise-based combined tracing method based on deep learning. Background Art

[0002] The field of hyperspectral satellite remote sensing technology encompasses a technical system that utilizes hyperspectral imaging equipment to collect and analyze high-spectral resolution data on the Earth's surface, atmosphere, and water bodies. The core of this field is the precise identification and quantitative inversion of various material compositions, spatial distribution characteristics, and dynamic change processes through detailed observations of continuous bands in the electromagnetic spectrum. The overall technical system encompasses a variety of observation methods, including hyperspectral satellite remote sensing, ground-based horizontal remote sensing, and targeted imaging remote sensing. It is capable of quantitatively monitoring and spatially tracing pollutants such as PM2.5, O3, NO2, and VOCs in the atmosphere at varying spatial and temporal resolutions. It is widely used in urban environmental monitoring, air quality assurance for major events, pollution source identification, and regulatory assessment.

[0003] The hyperspectral-VOC cruise-based source tracing method combines hyperspectral remote sensing data with cruise-based VOC monitoring data to identify high-value pollutant sources and analyze emission traceability. This method addresses the identification of volatile organic compound emission sources, encompassing technical aspects such as simultaneous data acquisition, hyperspectral feature extraction, screening for abnormal VOC concentrations, and intelligent identification of high-value pollution source areas.

[0004] When faced with the problem of VOCs emission source tracing, existing technologies generally rely on plane matching of static data and human experience judgment, resulting in a lack of dynamic response in time and space, and are particularly difficult to deal with non-continuous and irregular emission behaviors. Although hyperspectral remote sensing has the ability to cover surface sources, in the process of identifying the source area boundary, it often has deviations in the judgment of the contour of the polluted area due to the complex grayscale distribution of pixels and the fuzzy gradient changes. It is easy to mistakenly identify differences in illumination or ground objects as pollution boundaries. VOCs cruise data lacks systematic screening standards in the identification of abnormal concentrations, and there is a problem of high-frequency fluctuations mixed with real anomalies, resulting in the accumulation of errors in task flow generation. At the same time, the identification of pollutant components relies on conventional feature comparison, ignoring the correlation of band structure and the difference in reflectivity, resulting in limited species classification accuracy. In the process of tracing the source, the emission points and components are often processed with simple matching or distance constraints, which makes it difficult to reflect the multi-factor relationship between component abundance and source point characteristics, and is prone to redundant item misassociation or missed detection. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a hyperspectral-VOCs navigation combined tracing method based on deep learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a hyperspectral-VOCs navigation combined tracing method based on deep learning, comprising the following steps:

[0007] S1: Use a mobile monitoring vehicle to obtain the VOCs concentration response sequence in the monitoring area, screen out abnormal VOCs concentrations, and build a concentration abnormality response task flow;

[0008] S2: Obtaining the concentration time series difference parameter, spatial distribution standard deviation, and local concentration fluctuation anomaly rate of the concentration anomaly response task flow and performing standardization processing, determining the pollution phase change point based on the three sets of indicators, and obtaining a pollution phase change determination label;

[0009] S3: Obtain a remote sensing image of the pollution phase change determination label monitoring area for a corresponding period of time, select the boundary position of the suspected pollution source area from the remote sensing image, perform gradient change judgment on the position, and output a boundary enhancement response map;

[0010] S4: Obtaining the hyperspectral fixed-point pixels of the suspected pollution source area in the boundary enhancement response map, and inputting them into the convolutional neural network model to perform VOCs compound component identification to obtain component identification results;

[0011] S5: Based on the component identification results, the posterior probability of the correspondence between the emission sources and the VOCs components is calculated by the Bayesian inversion method, and the pollution source tracing list is obtained by sorting and screening according to the probability values.

[0012] As a further solution of the present invention, the concentration anomaly response task flow includes the concentration anomaly moment, abnormal concentration amplitude, and abnormal duration; the pollution phase change judgment label includes the phase change occurrence time, phase change spatial position, and phase change type; the boundary enhancement response map specifically includes the boundary position coordinates, gradient change direction, and boundary continuity index; the component identification results include VOCs species type, reflectivity characteristic vector, and relative abundance value; the pollution source tracing list specifically refers to the emission source coordinates, emission source type, and posterior probability value.

[0013] As a further solution of the present invention, the steps for obtaining the abnormal concentration response task flow are specifically as follows:

[0014] S111: Obtain the VOCs concentration response sequence in the monitoring area through the mobile monitoring vehicle, collect the TVOCs continuous concentration data stream at a frequency of one second from the single mass spectrometer, and arrange the continuously collected TVOCs concentration values ​​in chronological order to generate a concentration time series;

[0015] S112: Based on the TVOCs value at each time point in the concentration time series, using a concentration threshold of 1 ppm as a judgment criterion, recording all time indexes that match the judgment criterion, and obtaining an over-threshold concentration index sequence;

[0016] S113: According to the time points recorded in the above-threshold concentration index sequence, the corresponding TVOCs concentration values ​​and occurrence times are extracted to construct a concentration anomaly response task flow.

[0017] As a further solution of the present invention, the steps for obtaining the pollution phase change determination tag are specifically as follows:

[0018] S211: Acquire VOCs concentration data for a time segment corresponding to the concentration anomaly response task flow, calculate concentration time series difference, spatial distribution standard deviation, and local concentration fluctuation anomaly rate, and obtain three change indicators;

[0019] S212: normalizing the three change indicators using the Z-score method to obtain a normalized indicator sequence;

[0020] S213: Setting a double sliding time window based on the standardized indicator sequence, identifying the jump trend of the local concentration fluctuation abnormality rate, marking the location area where the jump occurs, and generating a pollution phase change judgment label.

[0021] As a further solution of the present invention, the step of obtaining the boundary enhancement response map is specifically as follows:

[0022] S311: Acquire remote sensing image data corresponding to the monitoring area and time period based on the pollution phase change determination label, extract the grayscale change of each pixel in the image, identify the boundary lines of continuous pixels with prominent gradient changes, screen the edge area positions of suspected pollution sources, and obtain a set of candidate boundary positions;

[0023] S312: Extracting the position coordinates, gradient values, and gradient point density of the boundary candidate position set in three consecutive frames of images, calculating the position offset, gradient change rate, and density increase or decrease of each point in the time dimension, determining whether there is a continuous upward trend, and obtaining a continuously enhanced boundary point set;

[0024] S313: Input the continuously enhanced boundary point set into the attention mechanism module in the convolutional neural network structure, dynamically assign spatial weights according to the gradient strength and change trend of each point, perform response enhancement processing on the boundary area, and output a boundary enhancement response map.

[0025] As a further solution of the present invention, the steps for obtaining the component identification results are specifically as follows:

[0026] S411: Obtaining the location of a hotspot region in the boundary enhancement response map where the brightness enhancement exceeds the boundary enhancement response threshold, selecting corresponding hyperspectral fixed-point pixels and image space position indexes from the hotspot region locations to obtain a hotspot pixel location set;

[0027] S412: Extracting the reflectance value of each pixel in multiple bands from the hotspot pixel position set, obtaining spectral features in the 3.3μm, 6.2μm, and 9.6μm band channels, constructing a continuous spectral reflectance sequence and encapsulating it as an input tensor to obtain a spectral feature input sequence;

[0028] S413: Input the spectral feature input sequence into the convolutional neural network classification model, call the convolution extraction layer and the fully connected discriminant layer to identify and compare the spectral angle structure, output the corresponding VOCs compound name label, and obtain the component identification result.

[0029] As a further solution of the present invention, the steps for obtaining the pollution source tracing list are specifically as follows:

[0030] S511: extracting the VOCs type and relative abundance information based on the component identification results, and merging similar items according to the standard format of component names to obtain a VOCs component feature set;

[0031] S512: Based on the VOCs component feature set, the spatial range marked in the pollution phase change label is obtained, and the species information and coordinate data of all emission points in the target area are extracted from the spatial emission database to construct a regional emission source list;

[0032] S513: Based on the regional emission source list and the VOCs component feature set, the Bayesian inversion method is called to calculate the posterior probability value of the component corresponding to each emission point, the probability results are sorted, and a pollution source tracing list is generated.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are:

[0034] In this invention, a concentration anomaly response task flow is constructed to conduct spatiotemporal linkage screening of VOCs concentration exceeding thresholds, effectively forming a highly sensitive marker for pollution response. Based on the joint analysis of three groups of indicators: concentration time series difference, spatial standard deviation, and local volatility, with the help of standardization processing and jump trend identification methods, multi-dimensional dynamic discrimination of pollution phase change phenomena is achieved, thereby enhancing the response capability of sudden or hidden emission events. The boundaries of suspected polluted areas in remote sensing images are finely characterized by the grayscale gradient change trend, and their stability is judged by combining time series displacement and local gradient density changes, effectively eliminating artifact disturbance factors and enhancing the robustness of pollution source boundary identification. In the selection of hyperspectral pixels, the hot spot area is focused based on the brightness enhancement threshold to ensure that the acquired spectral data has typical characteristics. The spectral input sequence is constructed using multi-band reflectance, and the type and abundance of VOCs are identified in combination with a deep convolution structure, effectively improving the accuracy of component identification and anti-interference ability. In the link between components and emission sources, an emission database under spatial constraints is introduced, and the emission source list is probability-sorted in the form of a posteriori probability evaluation, so that the traceability results have significant data support and quantitative reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the main steps of the present invention;

[0036] Figure 2 This is a flow chart of step S1 of the present invention;

[0037] Figure 3 This is a flow chart of step S2 of the present invention;

[0038] Figure 4 This is a flow chart of step S3 of the present invention;

[0039] Figure 5 This is a flow chart of step S4 of the present invention;

[0040] Figure 6 This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0042] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0043] See also Figure 1 The present invention provides a technical solution: a hyperspectral-VOCs navigation combined tracing method based on deep learning, comprising the following steps:

[0044] S1: Use a mobile monitoring vehicle to obtain the VOCs concentration response sequence in the monitoring area, screen out abnormal VOCs concentrations, and build a concentration abnormality response task flow;

[0045] S2: Obtain and standardize three sets of indicators for the concentration anomaly response task flow: concentration time series difference parameters, spatial distribution standard deviation, and local concentration fluctuation anomaly rate. Determine the pollution phase transition point based on the three sets of indicators and obtain the pollution phase transition judgment label.

[0046] S3: Obtain remote sensing images of the pollution phase change determination label monitoring area for the corresponding period, select the boundary position of the suspected pollution source area from the remote sensing image, perform gradient change judgment on the position, and output the boundary enhancement response map;

[0047] S4: Obtain the hyperspectral fixed-point pixels of the suspected pollution source area in the boundary enhancement response map and input them into the convolutional neural network model to identify the VOCs compound components and obtain the component identification results;

[0048] S5: Based on the component identification results, the posterior probability of the corresponding relationship between emission sources and VOCs components is calculated through the Bayesian inversion method, and the pollution source tracing list is obtained by sorting and screening according to the probability value;

[0049] The concentration anomaly response task flow includes the concentration anomaly moment, abnormal concentration amplitude, and anomaly duration. The pollution phase change judgment label includes the phase change occurrence time, phase change spatial position, and phase change type. The boundary enhancement response map specifically includes the boundary position coordinates, gradient change direction, and boundary continuity index. The component identification results include VOCs species type, reflectivity characteristic vector, and relative abundance value. The pollution source tracing list specifically refers to the emission source coordinates, emission source type, and posterior probability value.

[0050] See also Figure 2,The specific steps for obtaining the concentration anomaly response task flow are:

[0051] S111: Obtain the VOCs concentration response sequence in the monitoring area through the mobile monitoring vehicle, collect the TVOCs continuous concentration data stream at a frequency of one second from the single mass spectrometer, and arrange the continuously collected TVOCs concentration values ​​in chronological order to generate a concentration time series;

[0052] Acquiring a VOC concentration response sequence within the monitoring area using a mobile monitoring vehicle requires continuous monitoring along a fixed route. The vehicle's single mass spectrometer collects total VOC concentration values ​​in the air once per second, recording three items during the sampling process: sampling time, sampling location, and TVOC value. The sampling device automatically samples in seconds, generating a concentration data stream corresponding to the duration and frequency. After receiving each second of sampling information, the device first arranges the information in chronological order and stores it in a cache, forming a sampling timeline. Each time point corresponds to a sampled concentration value. During continuous driving, for example, if a monitoring mission is set to a driving time of 30 minutes, or 1800 seconds, the collected concentration response data will also be 1800 sets. This data structure is maintained using a linear index, with time point indices paired with concentration values, such as 1.03 ppm for the first second and 0.96 ppm for the second second. After initial sampling, the sequence is organized into a one-dimensional time series for subsequent filtering operations. The resulting time series represents a mapping of TVOC concentrations against time.

[0053] S112: Based on the TVOCs value at each time point in the concentration time series, using the 1 ppm concentration threshold as a judgment benchmark, record all time indexes that match the judgment benchmark to obtain an over-threshold concentration index sequence;

[0054] After the concentration time series is established, the system performs conditional screening based on a set concentration threshold. In this example, the threshold is set at 1 ppm, meaning any concentration value greater than 1 ppm per second is considered an outlier. The system traverses all data points in the concentration time series and performs individual comparisons, using the "if current value > 1 ppm, then an outlier" logic for judgment. For time points that meet the criteria, their indexes in the time series are recorded, and the corresponding original concentration values ​​are retained. During system execution, if the concentration values ​​collected at positions such as the 12th, 30th, and 85th seconds are 1.14 ppm, 1.23 ppm, and 1.07 ppm, respectively, these positions are marked as outliers. All outlier point indices are aggregated into a sequence structure and passed to the next stage of analysis. It is important to note that to mitigate sampling errors, the system requires at least three outliers within a consecutive time interval to constitute a valid segment. Isolated outlier points are discarded. The final record retains the indices of all sampling points that exceed the threshold and meet the continuity criteria, forming a sequence of over-threshold concentration indices.

[0055] S113: According to the time points recorded in the threshold concentration index sequence, the corresponding TVOCs concentration values ​​and occurrence times are extracted, and a concentration anomaly response task flow is constructed;

[0056] Based on the filtered index sequence of concentrations exceeding the threshold, the system extracts the concentration values ​​and sampling times for all corresponding time points. It then uses the vehicle's GPS information to retrieve the synchronously recorded location information to form structured task units. Each task item includes a start time, end time, maximum concentration value, and the corresponding GPS location along the sampling route. If consecutive points form an abnormal segment, the system records the segment's start and end times and calculates the maximum concentration within that segment as a representative value to identify the pollution intensity. To enhance the logical clarity of the task flow, the system assigns each segment a unique number, such as Task 1, Task 2, and so on. Task information is organized into a key-value structure. For example, if a segment starts at 200 seconds and ends at 210 seconds, with a maximum concentration of 1.88 ppm, the task item would read: Start 200 seconds, End 210 seconds, Main value 1.88 ppm, Located at the corresponding location along the monitoring route. All task items are aggregated to form a task data structure that can be called by downstream modules, known as the concentration anomaly response task flow.

[0057] See also Figure 3 ,The specific steps for obtaining the pollution phase change judgment label are:

[0058] S211: Obtain VOCs concentration data for the time segment corresponding to the concentration anomaly response task flow, calculate the concentration time series difference, spatial distribution standard deviation, and local concentration fluctuation anomaly rate, and obtain three change indicators;

[0059] When processing the time segment corresponding to the concentration anomaly response task flow, three key indicator parameters need to be calculated: concentration time series difference, spatial distribution standard deviation and local concentration fluctuation anomaly rate. First, the concentration time series difference is used to reflect the concentration change of a monitoring point between the current time and the previous time. The calculation formula is as follows:

[0060] ΔC i (t) = C i (t)-C i (t-1);

[0061] Where: ΔC i (t): represents the concentration difference value of the i-th monitoring point at time t (unit: ppm); C i (t): represents the concentration value of the i-th monitoring point at the current time t; C i (t-1): represents the concentration value at the previous moment.

[0062] Taking the actual sampling data as an example, the concentration value at t = 121 seconds is [1.30, 1.15, 0.92, 1.00, 0.83] ppm, and the concentration at t = 120 seconds is [1.20, 1.10, 0.95, 1.05, 0.88] ppm. The difference result is: ΔC(121) = [0.10, 0.05, -0.03, -0.05, -0.05] ppm. Secondly, the spatial distribution standard deviation measures the concentration dispersion of different monitoring points in the same time slice. The calculation formula is as follows:

[0063]

[0064] Where: σ(t): spatial distribution standard deviation at time t (unit: ppm); N: number of monitoring points; C i (t): concentration value of the i-th monitoring point; The mean concentration of all monitoring points at time t.

[0065] Taking the 121st second as an example, the concentrations at the monitoring points are [1.30, 1.15, 0.92, 1.00, 0.83] ppm, and the average is:

[0066]

[0067] Substituting into the formula, we obtain: σ(121)≈0.167ppm.

[0068] Third, the local concentration fluctuation anomaly rate is used to indicate the proportion of monitoring points that exceed the set threshold at a certain moment. The calculation formula is: Where: R abn (t): local concentration fluctuation anomaly rate at time t; M t: The number of monitoring points whose concentration value at the current moment is higher than the set threshold T; N: The total number of monitoring points at that moment; T: TVOCs abnormality judgment threshold, which is set to 1.00ppm here.

[0069] At 121 seconds, the concentration values ​​are [1.30, 1.15, 0.92, 1.00, 0.83] ppm, satisfying C i >T has 3 points, so: In summary, at 121 seconds, the concentration time series difference for this time slice is [0.10, 0.05, -0.03, -0.05, -0.05] ppm, the spatial distribution standard deviation is 0.167 ppm, and the local concentration fluctuation anomaly rate is 0.6. These three parameters constitute a complete set of pollution dynamics indicators, which are used in subsequent standardization and trend mutation identification operations.

[0070] S212: The three change indicators are standardized using the Z-score method to obtain a standardized indicator sequence;

[0071] After obtaining the three change indicator sequences output by the previous operation—the concentration time-series difference, the spatial distribution standard deviation, and the local concentration fluctuation anomaly rate—the system performs normalization on each sequence to address the issues of inconsistent units and numerical scales across the different indicators, facilitating trend comparison and jump identification under a unified reference. Normalization utilizes a Z-score approach: for each time point in the current indicator sequence, the mean of the sequence is subtracted and then divided by the standard deviation. This process maps the raw data to a dimensionless numerical sequence centered on 0, making each indicator comparable. In this example, if the local concentration fluctuation anomaly rate has a mean of 0.45 and a standard deviation of 0.12 in a set of samples, and the value is 0.60 at 121 seconds, the normalized result is 1.25. Similarly, in the spatial distribution standard deviation sequence, if the value at a certain moment is 0.19, the sequence mean is 0.14, and the standard deviation is 0.03, the normalized result is 1.67. Through the above method, the system obtains three standardized indicator sequences respectively. Their structure remains consistent with the original time series. Each moment corresponds to a set of three standardized values. This structure will serve as the input of the next sliding window trend identification module and is uniformly named as the standardized indicator sequence.

[0072] S213: Setting a double sliding time window based on the standardized indicator sequence, identifying the jump trend of the local concentration fluctuation abnormality rate, marking the location area where the jump occurs, and generating a pollution phase change judgment label;

[0073] Based on the standardized indicator sequence, the system employs a sliding time window strategy to identify and classify trend changes within the sequence, thereby determining sudden changes in pollution concentration over time and space. The system uses two window lengths: a fixed window to obtain the historical mean value for the reference period, and a variable window to detect short-term fluctuations in the indicator. For example, the fixed window length is set to 30 seconds, and the variable window length is set to 10 seconds. The two windows slide synchronously at a step size of one second. At each sliding position, the standardized value of the local concentration fluctuation anomaly rate is averaged in the two windows. The difference between the means is compared. If the difference exceeds a set jump threshold, the current time point is marked as a suspected sudden change. The jump threshold is determined based on experimental statistical analysis and is set to 1.5 in this example. If, at 310 seconds, the fixed window mean is 0.42 and the variable window mean is 2.10, the mean difference is 1.68, exceeding the set threshold of 1.5, indicating a sudden change at that moment. The system also analyzes the trend of the spatial distribution standard deviation at the corresponding time point. If this indicator increases continuously within the current and previous two seconds, it indicates that the pollution is spreading more and more spatially, further strengthening the confirmation of the sudden change at that time point. Once the judgment is confirmed, the system generates a unique pollution phase change identification information based on the time and space location of the pollution point. The information includes fields such as the sudden change time, sudden change coordinates, and the indicator jump value, and outputs the pollution phase change determination label.

[0074] See also Figure 4 , the steps for obtaining the boundary enhancement response map are as follows:

[0075] S311: Based on the pollution phase change determination label, remote sensing image data corresponding to the monitoring area and time period is obtained, the grayscale change of each pixel in the image is extracted, the boundary lines of continuous pixels with prominent gradient changes are identified, and the edge area positions of suspected pollution sources are screened to obtain a set of candidate boundary positions;

[0076] When acquiring remote sensing image data based on the pollution phase change judgment label, the system first determines the monitoring area covered by the label and the corresponding time period, retrieves the multispectral or hyperspectral image frame sequence of the corresponding area within the time period from the remote sensing database, and arranges them in chronological order to ensure that the time axis is aligned with the phase change label. After the image is loaded, the system performs grayscale gradient calculation on all pixels in each frame of the image, calls the grayscale value difference between each pixel and its upper, lower, left and right adjacent pixels, extracts the grayscale change amplitude, and marks the change direction to construct the gradient map of the current frame. Subsequently, the system searches for connected areas composed of pixel points with continuously increasing grayscale change amplitudes in the same image, constructs boundary lines according to the contour extraction logic, retains closed structures with a boundary contour length of more than 10 pixels, and identifies them as continuous boundary lines. After extracting all candidate boundary lines, the system determines whether the gradient amplitude of the boundary grayscale value exceeds the threshold setting, which is set to 20 grayscale levels in this embodiment. If the average gradient value of any section on the continuous boundary line exceeds the threshold, the corresponding boundary area is determined to be the edge area of ​​the suspected pollution source. The boundary coordinates of all regional pixels that meet the conditions are summarized to form a set of candidate boundary locations.

[0077] S312: Extract the position coordinates, gradient values, and gradient point density of the boundary candidate position set in three consecutive frames of images, calculate the position offset, gradient change rate, and density increase or decrease of each point in the time dimension, determine whether there is a continuous upward trend, and obtain a continuously enhanced boundary point set;

[0078] After extracting a set of candidate boundary locations, the system compares them sequentially across three adjacent frames by time index. The spatial position of each boundary point across the three consecutive frames is calibrated as a point coordinate sequence. The system calculates the position offset based on this sequence. If the coordinate position of the point at the three moments shows a unidirectional motion trend and the cumulative offset exceeds 2 pixels, the position change is considered continuous. The system also obtains the gradient value corresponding to the point in each frame and calculates the gradient change rate. Using the gradient value of the first frame as a reference, the system observes the direction of change and the ratio of increase over the next two moments. If the gradient value continuously increases with an increase of more than 20%, it is marked as a strong gradient change point. To further characterize the boundary intensity structure, the system creates a 3×3 pixel sliding window within the neighborhood of each boundary point. The number of pixels in the window marked as boundary points is counted, the local gradient density is calculated, and the magnitude of this density change across the three frames is analyzed. If the density continuously increases with an increment of at least 1 pixel, it indicates that the local structure is strengthening. For boundary points that simultaneously meet the requirements of significant position offset, positive gradient change rate and increasing density, the system determines them as continuously enhanced boundary points, and the collection of all such points constitutes a continuously enhanced boundary point set.

[0079] S313: Input the continuously enhanced boundary point set into the attention mechanism module in the convolutional neural network structure, dynamically assign spatial weights according to the gradient strength and change trend of each point, perform response enhancement processing on the boundary area, and output a boundary enhancement response map;

[0080] The continuously enhanced boundary point set is used as the input feature set and passed into the attention mechanism module in the convolutional neural network structure. The module needs to assign spatial weights to each boundary point to adjust its response strength during the convolution extraction process. The weight assignment is not directly based on a single gradient value, but is calculated based on three dynamic factors: the current frame gradient strength G i , gradient change rate R i , local gradient density increase D i , all three are introduced into the weight function after being processed in a normalized manner, and the normalized processing interval is unified to [0,1]. Among them, the gradient strength G i Indicates the ratio of the grayscale gradient value of the current frame pixel to the maximum gradient value in the frame, the gradient change rate R i Indicates the linear increase degree of the point in three consecutive frames, density increase D i Represents the ratio of the increase in the number of gradient points in the local window at that point to the maximum value. The weighting coefficients α, β, and γ are set based on the contribution of the three indicators to the continuity of the boundary and the stability of mutation detection. Empirical calculations and experimental statistics show that: the gradient strength is directly related to the structural positioning, and the weight should be the first, taking α = 0.5; the gradient change rate can reflect the state of boundary movement and is second in importance for trend detection, setting β = 0.3; the density increase mainly represents the integrity of the local area change. Due to its high volatility, it is set to the lowest weight, taking γ = 0.2. The weight function is as follows:

[0081]

[0082] Taking a continuous enhancement boundary point as an example, the current frame gradient strength is 28, the current image frame maximum gradient is 40, the three-frame gradient is 20→28→35, the corresponding gradient change rate is 0.75 after normalization, the local density increases from 5 to 8, the maximum density is 10, and the normalized parameter is: G i =0.70, R i =0.75, D i =0.30. Substitute into the calculation: The convolutional attention weight assigned to this boundary point is 0.635. The system applies this weight directly to the activation value multiplier of the corresponding pixel in the convolutional layer input tensor. Pixels with higher weights express a stronger response in subsequent feature maps, while those with lower weights are suppressed. After the system traverses all consecutive enhanced boundary points and performs the above process, it completes the generation of a spatial attention map. This map is then convolved bit by bit with the input image. The convolution output retains the response of the enhanced boundary and suppresses the response of the unchanged area, forming a boundary enhanced response map.

[0083] See also Figure 5 , the steps for obtaining component identification results are as follows:

[0084] S411: Obtaining the location of a hotspot region in the boundary enhancement response map where the brightness enhancement exceeds the boundary enhancement response threshold, selecting the corresponding hyperspectral fixed-point pixel and image space position index from the hotspot region location, and obtaining a hotspot pixel location set;

[0085] When obtaining areas where the brightness enhancement exceeds the threshold in the boundary enhancement response map, the system first scans point by point based on the enhanced intensity value corresponding to each pixel in the map, setting the boundary enhancement response threshold to 0.6 (normalized dimensionless value). This value is derived from the weighted setting based on the median of the spectrum intensity value distribution. After sample statistical experiments, it is determined that it can be used to effectively screen areas with strong response to pollution sources. The system judges all pixels. If a pixel value is greater than or equal to the threshold, its corresponding coordinates are marked as hotspot points. Between multiple hotspot points, the system extracts adjacent hotspot pixels in a regional clustering manner, constructs hotspot areas, sets the minimum connected area to be no less than 3×3 pixel blocks, excludes isolated points, determines the boundaries of the hotspot areas, and samples multiple pixel points at the center and edge of the area for spectral data extraction. Based on the pixel position of each hotspot area, the system obtains its spatial index in the image matrix and summarizes it into a list structure. This set is named the hotspot pixel position set, which contains the coordinate index information of all selected fixed-point pixels and the hotspot number to which they belong.

[0086] S412: Extract the reflectance value of each pixel in multiple bands from the hotspot pixel location set, obtain the spectral features in the 3.3μm, 6.2μm and 9.6μm band channels, construct a continuous spectral reflectance sequence and encapsulate it as an input tensor to obtain a spectral feature input sequence;

[0087] Based on the acquired set of hotspot pixel locations, the system sequentially reads the complete band reflectance information for each pixel in the hyperspectral image and selects the reflectance values ​​of the three target band channels, 3.3μm, 6.2μm, and 9.6μm, as feature dimensions. These bands are associated with strong absorption characteristics of volatile organic compounds in the mid-infrared region, as verified by standard band response experiments. In the actual extraction process, the system first obtains the spectral reflectance sequence for each pixel from the original hyperspectral cube image data, totaling N bands. The numerical unit of each band is the reflectance ratio (dimensionless, ranging from 0–1). The system then reads the 45th, 88th, and 132nd band data from the N-dimensional band, corresponding to central wavelengths of 3.3μm, 6.2μm, and 9.6μm, respectively, to construct a three-dimensional reflectance vector. The system then encapsulates this vector and its spatial index position into a single input sample. Ultimately, all samples form an input tensor with an M×3 data structure, where M is the number of hotspot pixels. This resulting spectral feature input sequence serves as input for subsequent model recognition.

[0088] S413: Input the spectral feature input sequence into the convolutional neural network classification model, call the convolution extraction layer and the fully connected discriminant layer to identify and compare the spectral angle structure, output the corresponding VOCs compound name label, and obtain the component identification result;

[0089] The system passes the spectral feature input sequence into the convolutional neural network classification model. Each row of the input tensor represents the spectral reflectance vector of a hotspot pixel in the selected band, which is set as a three-dimensional vector x=

[0090] [x a ,x b ,x c ], where: x a :Indicates that the pixel is in band λ a (i.e., reflectivity at 3.3 μm); x b :Indicates that the pixel is in band λ b (i.e., reflectivity at 6.2 μm); x c :Indicates that the pixel is in band λ c (i.e., the reflectivity at 9.6 μm); the first layer of the convolutional network is a one-dimensional convolutional layer, and its convolution kernel parameter is recorded as w = [w a ,w b ,w c ], corresponding to the weight parameter of each input dimension, where: w a : Convolution kernel band λ a The response weight of w b : Convolution kernel band λ b The response weight of w c : Convolution kernel band λ c The response weight of the convolutional layer is:

[0091] y=σ(w a ·x a +w b ·x b +w c ·x c +b);

[0092] Where: y: is the activation value output by the convolutional layer; σ(·): is the activation function, using the ReLU function σ(z) = max(0,z).

[0093] Set the input value to x a =0.52, x b =0.65, x c =0.70, and the convolution kernel parameter is w a =0.4, w b =0.3, w c =0.2, bias term b = -0.1, substituting into:

[0094] y=σ(0.4·0.52+0.3·0.65+0.2·0.70-0.1)=σ(0.443)=0.443.

[0095] The system then passes the activation value as input to the fully connected discriminant layer for classification and recognition, and uses cosine similarity calculation to compare the input vector with the standard reflectance vector of known VOCs components. The formula is as follows:

[0096]

[0097] Where: x = [x a ,x b ,x c ]: is the reflectivity vector of the input pixel; t=[t a ,t b ,t c ]: standard VOCs component reflectivity vector; t a , t b , t c : are the standard species in band λ a ,λ b ,λ c Reflectivity under; S(x,t): is the cosine similarity between the input and the standard vector.

[0098] Assume that the input pixel vector is the above value and the reflectance vector of the standard species ethylbenzene is t a =0.50, t b =0.66, t c =0.68, then we can calculate:

[0099]

[0100] The system sets the similarity threshold to θ = 0.92. When S(x, t) ≥ θ, the sample is determined to belong to the ethylbenzene component. a ,w b ,w c The system refers to the response ability of each band in the VOCs spectrum to the component identification contribution. Through statistical analysis of a large number of labeled samples, the following setting basis is obtained: Band λ a (3.3 μm) is the main absorption region of most aromatic structures in VOCs, and the signal is stable and has strong discrimination, so we set w a The largest weight; band λ b (6.2μm) is the reflection characteristic of some ketones and alkanes, with medium signal. b Next is band λ c (9.6μm) is more sensitive to heterocyclic rings, and the signal-to-noise ratio is low. c In summary, all input samples complete convolution calculation and classification recognition according to the above process, and finally form the VOCs compound component identification results corresponding to the hotspot pixel location set for subsequent pollution source inversion.

[0101] See also Figure 6 , the specific steps for obtaining the pollution source tracing list are:

[0102] S511: Based on the component identification results, the types and relative abundance information of the VOCs contained are extracted, and similar items are merged according to the standard format of component names to obtain the VOCs component feature set;

[0103] Based on the component identification results, the type and relative abundance information of VOCs contained are extracted. First, the characteristic absorption peaks of the atmospheric samples in the target area are despected according to the real-time spectral data collected by the detection instrument. The despectral process sequentially performs spectral background correction, spectral line denoising, baseline flattening and peak recognition processing. The background correction is achieved by introducing the environmental reference signal and the point-by-point difference between the original signal. The denoising link removes the weak disturbance band according to the set amplitude threshold, and sets the baseline stable band reference average as the normalization benchmark. Then, the positions of the multiple characteristic peaks identified are compared one by one with the characteristic absorption positions included in the standard VOCs database. In the comparison process, the wavelength difference is no more than ±0.5nm as the matching basis, and the peak height is checked. The relative difference is calculated, and the match is confirmed to be successful if the relative error is less than 10%. Target components such as benzene, toluene, and isopropylbenzene are matched in turn in a group of samples, and then the relative abundance of each component is estimated according to the proportion. The abundance estimation adopts the normalization of the total peak area to obtain the proportion value. The corresponding proportion of each component is calculated by dividing the integral intensity by the total integral intensity. On this basis, components with the same name or similar structure are classified into a unified category. For example, toluene and p-methylbenzene are merged into "toluenes", benzene is retained separately, and isopropylbenzene is classified into alkylbenzenes according to its structure. The preliminary component merging is completed, and the compound index standard is uniformly used for naming. Finally, the VOCs component feature set is obtained, which contains the standardized component category names and their relative abundance values.

[0104] S512: Based on the VOCs component feature set, the spatial range marked in the pollution phase change label is obtained, and the species information and coordinate data of all emission points in the target area are extracted from the spatial emission database to construct a regional emission source list;

[0105] Based on the VOCs component feature set, the spatial region boundary coordinates marked in the pollution phase change label are first parsed. The target analysis rectangle is constructed based on the four obtained boundary points. Spatial screening logic is then set based on this information and used as the screening criteria for emission point data in the spatial emission database. A bidirectional interval judgment method is used to determine whether the latitude and longitude of each emission point simultaneously fall within the set coordinate interval. If the conditions are met, the emission point is considered a valid emission point in the target area. The emission species information corresponding to the emission point is then extracted. The extraction process is based on the species name and emission concentration or intensity values ​​recorded in the emission database. All registered emission component information for the point is recorded and summarized by category. The summary process maintains consistency with the component merging logic in the feature set, that is, classification by "toluene", "benzene", "alkylbenzene", etc., and the original emission values ​​and relative proportions are recorded. This operation is performed sequentially for all filtered emission points to construct a regional emission source list containing geographical location, emission species type, and emission amount. This ensures that each emission point in the emission source list has a traceable location and corresponding VOC species and emission information.

[0106] S513: Based on the regional emission source list and the VOCs component feature set, the Bayesian inversion method is used to calculate the posterior probability value of the component corresponding to each emission point, the probability results are sorted, and a pollution source tracing list is generated;

[0107] Based on the regional emission source list and VOCs component feature set, the emission information of each emission point and the feature set must first be processed at the element level. Specifically, the emission species list of each emission point is used as the basic data set, and a cross-search is performed with the standardized and merged component list in the feature set. For the component items that appear in both sets, their emission intensity values ​​in the emission point are extracted, and their relative proportions are calculated. Let the emission intensity of the i1th component in the emission point be The total emission intensity is The standardized emission intensity of this component is Simultaneously extract the relative abundance of the corresponding components in the feature set Then the core parameter pair involved in the calculation of the posterior probability is formed Then, the posterior probability value is calculated for each emission point. The posterior probability value indicates the likelihood that the emission point is a pollution source. The calculation formula is as follows:

[0108]

[0109] in, The posterior probability value of emission point j1 indicates the relative possibility that this point is a pollution source; The emission intensity of the i1th component in the emission point j1; The sum of the emission intensities of all components in emission point j1; The relative abundance of the i1th component in the feature set; Weight coefficient, used to strengthen the role of certain key components in the overall matching; Z: normalization constant, so that the sum of the posterior probabilities of all emission points is 1.

[0110] The weight coefficient The basis for setting the value is as follows: Based on the pollution incident data of national key monitoring points within ten years, the frequency of each type of VOCs components identified as the dominant species in major pollution incidents is counted, and the components with an occurrence frequency higher than 20% are assigned a weight of 1.2, the frequency between 10% and 20% is assigned a value of 1.1, and the frequency below 10% is set to 1.0. The toluene component is given a weight of 1.2 due to its frequency of 26.3%, the benzene is set to 1.1 due to its frequency of 18.7%, and the alkylbenzene is set to 1.0 at 9.1%. This setting ensures that the weight source has historical statistical support and avoids subjective conjecture.

[0111] Assume that the relative abundances of toluene, benzene, and alkylbenzenes in the feature set are 0.324, 0.287, and 0.195, respectively, and the corresponding weights are set to 1.2, 1.1, and 1.0, respectively. The emission intensities of these three components at emission point j1 = B are 17.9 mg / s, 21.3 mg / s, and 14.2 mg / s, respectively, and the total emission intensity is 53.4 mg / s. Then:

[0112] Toluene standardized strength is The error is |0.335-0.324|=0.011, and the weighted error is 0.011×1.2=0.0132;

[0113] The normalized strength of benzene is The error is |0.399-0.287|=0.112, and the weighted error is 0.112×1.1=0.1232;

[0114] The normalized strength of alkylbenzenes is The error is |0.266-0.195|=0.071, and the weighted error is 0.071×1.0=0.071.

[0115] Then the sum of the exponential terms is:

[0116]

[0117] Substituting this into the formula, we can get the unnormalized posterior probability of the emission point j1=B:

[0118]

[0119] After completing the above operations for all emission points, Take the sum and normalize. For example, there are three emission points A, B, and C, and the unnormalized values ​​are 0.813, 0.785, and 0.642 respectively. Then the normalization constant Z = 0.813 + 0.785 + 0.642 = 2.240, and the final posterior probability of emission point B is:

[0120]

[0121] The posterior probability values ​​calculated for all emission points are arranged in descending order to form a pollution source tracing list. The list is sorted from high to low according to the posterior probability of each emission point, indicating the possibility level of it as the source of the target pollution event. The result is used for the subsequent tracing lock and key emission source identification link in the emergency response strategy, completing the tracing closed loop from identifying characteristic components to the corresponding emission points.

[0122] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. The hyperspectral-VOCs navigation combined tracing method based on deep learning is characterized by: The following steps are involved: S1: Use a mobile monitoring vehicle to obtain the VOCs concentration response sequence in the monitoring area, screen out abnormal VOCs concentrations, and build a concentration abnormality response task flow; S2: Obtaining the concentration time series difference parameter, spatial distribution standard deviation, and local concentration fluctuation anomaly rate of the concentration anomaly response task flow and performing standardization processing, determining the pollution phase change point based on the three sets of indicators, and obtaining a pollution phase change determination label; S3: Obtain a remote sensing image of the pollution phase change determination label monitoring area for a corresponding period of time, select the boundary position of the suspected pollution source area from the remote sensing image, perform gradient change judgment on the position, and output a boundary enhancement response map; S4: Obtaining the hyperspectral fixed-point pixels of the suspected pollution source area in the boundary enhancement response map, and inputting them into the convolutional neural network model to perform VOCs compound component identification to obtain component identification results; S5: Based on the component identification results, the posterior probability of the correspondence between the emission sources and the VOCs components is calculated by the Bayesian inversion method, and the pollution source tracing list is obtained by sorting and screening according to the probability values.

2. The deep learning-based hyperspectral-VOCs navigation combined tracing method according to claim 1 is characterized in that: The concentration anomaly response task flow includes the concentration anomaly moment, abnormal concentration amplitude, and abnormal duration; the pollution phase change judgment label includes the phase change occurrence time, phase change spatial position, and phase change type; the boundary enhancement response map specifically includes the boundary position coordinates, gradient change direction, and boundary continuity index; the component identification result includes the VOCs species type, reflectivity characteristic vector, and relative abundance value; the pollution source tracing list specifically refers to the emission source coordinates, emission source type, and posterior probability value.

3. The deep learning-based hyperspectral-VOCs navigation-based tracing method according to claim 1 is characterized in that: The specific steps for obtaining the concentration anomaly response task flow are as follows: S111: Obtain the VOCs concentration response sequence in the monitoring area through the mobile monitoring vehicle, collect the TVOCs continuous concentration data stream at a frequency of one second from the single mass spectrometer, and arrange the continuously collected TVOCs concentration values ​​in chronological order to generate a concentration time series; S112: Based on the TVOCs value at each time point in the concentration time series, using a concentration threshold of 1 ppm as a judgment criterion, recording all time indexes that match the judgment criterion, and obtaining an over-threshold concentration index sequence; S113: According to the time points recorded in the above-threshold concentration index sequence, the corresponding TVOCs concentration values ​​and occurrence times are extracted to construct a concentration anomaly response task flow.

4. The deep learning-based hyperspectral-VOCs navigation-based tracing method according to claim 3 is characterized in that: The steps for obtaining the pollution phase change determination label are specifically as follows: S211: Acquire VOCs concentration data for a time segment corresponding to the concentration anomaly response task flow, calculate concentration time series difference, spatial distribution standard deviation, and local concentration fluctuation anomaly rate, and obtain three change indicators; S212: normalizing the three change indicators using the Z-score method to obtain a normalized indicator sequence; S213: Setting a double sliding time window based on the standardized indicator sequence, identifying the jump trend of the local concentration fluctuation abnormality rate, marking the location area where the jump occurs, and generating a pollution phase change judgment label.

5. The deep learning-based hyperspectral-VOCs navigation-based tracing method according to claim 4 is characterized in that: The steps for obtaining the boundary enhancement response map are specifically as follows: S311: Acquire remote sensing image data corresponding to the monitoring area and time period based on the pollution phase change determination label, extract the grayscale change of each pixel in the image, identify the boundary lines of continuous pixels with prominent gradient changes, screen the edge area positions of suspected pollution sources, and obtain a set of candidate boundary positions; S312: Extracting the position coordinates, gradient values, and gradient point density of the boundary candidate position set in three consecutive frames of images, calculating the position offset, gradient change rate, and density increase or decrease of each point in the time dimension, determining whether there is a continuous upward trend, and obtaining a continuously enhanced boundary point set; S313: Input the continuously enhanced boundary point set into the attention mechanism module in the convolutional neural network structure, dynamically assign spatial weights according to the gradient strength and change trend of each point, perform response enhancement processing on the boundary area, and output a boundary enhancement response map.

6. The deep learning-based hyperspectral-VOCs navigation-based tracing method according to claim 5 is characterized in that: The steps for obtaining the component identification results are specifically as follows: S411: Obtaining the location of a hotspot region in the boundary enhancement response map where the brightness enhancement exceeds the boundary enhancement response threshold, selecting corresponding hyperspectral fixed-point pixels and image space position indexes from the hotspot region locations to obtain a hotspot pixel location set; S412: Extracting the reflectance value of each pixel in multiple bands from the hotspot pixel position set, obtaining spectral features in the 3.3μm, 6.2μm, and 9.6μm band channels, constructing a continuous spectral reflectance sequence and encapsulating it as an input tensor to obtain a spectral feature input sequence; S413: Input the spectral feature input sequence into the convolutional neural network classification model, call the convolution extraction layer and the fully connected discriminant layer to identify and compare the spectral angle structure, output the corresponding VOCs compound name label, and obtain the component identification result.

7. The deep learning-based hyperspectral-VOCs navigation-based tracing method according to claim 6 is characterized in that: The specific steps for obtaining the pollution source tracing list are as follows: S511: extracting the VOCs type and relative abundance information based on the component identification results, and merging similar items according to the standard format of component names to obtain a VOCs component feature set; S512: Based on the VOCs component feature set, the spatial range marked in the pollution phase change label is obtained, and the species information and coordinate data of all emission points in the target area are extracted from the spatial emission database to construct a regional emission source list; S513: Based on the regional emission source list and the VOCs component feature set, the Bayesian inversion method is called to calculate the posterior probability value of the component corresponding to each emission point, the probability results are sorted, and a pollution source tracing list is generated.

Images