Hyperspectral remote sensing pollution gas mass concentration quantitative imaging method and device based on artificial intelligence

Abstract

According to the artificial intelligence-based hyper-spectral remote sensing pollution gas mass concentration quantitative imaging method and device provided by the invention, based on a plurality of basic machine learning sub-modules, the advantages of each sub-module are integrated, the weight of each sub-module is allocated by establishing a weighting and dynamic fusion mechanism, and a scene adaptive fusion machine learning model is constructed, so that the mass concentration of the pollution gas is obtained. Dynamic sensing and self-adaptive processing of smoke plume emission mass concentration of real atmospheric conditions, especially complex and changeable aerosol scenes, are realized, deep interpretability analysis is further provided, the credibility of smoke plume emission mass concentration prediction is enhanced, and the prediction accuracy of smoke plume emission mass concentration prediction is improved. According to the method, the scene self-adaptive design of the model can be better verified, iterative optimization of parameters in the model is guided in a targeted mode, high-precision, high-robustness, high-timeliness and reliable hyper-spectral remote sensing pollution gas mass concentration quantitative imaging prediction is achieved, the method serves for refined monitoring of atmospheric pollution, and the method has important theoretical significance and application value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical measurement, and particularly relates to a hyperspectral remote sensing pollution gas mass concentration quantitative imaging method and device based on artificial intelligence. BACKGROUND

[0002] Hyperspectral remote sensing technology, especially differential optical absorption spectroscopy (DOAS), has been widely used in the monitoring of atmospheric trace gases. DOAS technology can invert the differential slant column density (DSCD) of the gas on the light path by analyzing the narrowband absorption characteristics of the gas in a specific waveband. DSCD is an integral quantity, which reflects the difference in the total number of gas molecules on a complete light path in the field of view relative to the reference background, and its unit is usually molecules / cm². However, in practical applications, we are more concerned about the mass concentration of pollutants (for example, in µg / m³) at a specific spatial point (for example, a specific area of a plume), which is a point quantity concept. There is no simple direct correspondence between the path integral total quantity represented by DSCD and the mass concentration at a specific point, and there is an essential difference between the two. Therefore, how to accurately calculate the mass concentration at a specific location from the light path integral DSCD is a core problem in the application of hyperspectral remote sensing data.

[0003] The traditional conversion method relies on a radiative transfer model (RTM), which estimates the mass concentration by simulating the transmission path of light in the atmosphere and various attenuation processes, combining meteorological parameters, aerosol information, and assumed pollution vertical profiles. However, the radiative transfer model has some inherent limitations: first, the accuracy of RTM is highly dependent on the accuracy of input parameters such as boundary layer height, pollution vertical distribution, etc., which are difficult to accurately obtain and update in real time under actual complex atmospheric conditions, resulting in limited model parameter settings and difficulty in fully reproducing the real and variable atmospheric state. Second, the calculation of RTM is usually complex and time-consuming, making it difficult to meet the needs of rapid and real-time monitoring and imaging of dynamically changing pollution events (such as plume diffusion). Especially in the presence of obvious aerosol emissions, the strong influence of aerosol on light transmission makes the concentration conversion based on RTM even less accurate.

[0004] An artificial intelligence system is a system that uses machine learning and other technologies to simulate human intelligence. Machine learning is a branch of artificial intelligence that aims to enable computer systems to complete tasks by learning from data rather than being explicitly programmed. It also aims to establish a computational theory of learning, construct various learning systems, and apply them to various fields. Machine learning research has three main directions. First, it attempts to establish a learning cognitive physiology model by simulating the human learning process. Second, it develops learning theories suitable for machines, explores all possible learning methods, and compares the similarities and differences between human learning and machine learning. Third, it establishes practical learning systems or knowledge acquisition tools, builds automatic knowledge acquisition systems in the application field of artificial intelligence science, accumulates experience, and perfects knowledge base and control knowledge, so as to make the intelligence level of machines similar to that of humans.

[0005] The most basic approach of machine learning is to discover patterns and learn from them in order to make predictions, classifications, pattern recognitions, or decisions by processing and analyzing a large amount of data. Therefore, based on machine learning, the complex mapping relationship between DSCD and mass concentration can be learned from a large number of observation data, which is expected to overcome some limitations of traditional RTM. However, existing machine learning-based methods still face many challenges: (1) Lack of scene adaptability: Real atmospheric conditions (especially the distribution and optical properties of aerosols) have spatial and temporal heterogeneity and dynamic variability. For example, the radiation transfer characteristics of clean atmosphere without obvious aerosol influence and polluted atmosphere with high aerosol load (such as industrial plume) are very different, resulting in different degrees of nonlinearity in the mapping relationship between DSCD and mass concentration. Most machine learning methods currently use a single model to try to cover all scenarios, or use a simple model integration strategy, which often fails to dynamically and adaptively adjust the internal processing mechanism of the model to the current specific atmospheric scene (especially the aerosol conditions), resulting in limited robustness of the model in complex and variable real environments, and large deviations in some specific scenarios.

[0006] (2) Limitations of fusion strategy: Although ensemble learning is used to improve model performance, most cases use relatively fixed fusion methods (such as simple averaging and fixed weight voting). It fails to fully utilize the unique advantages of different base models in processing specific scene data characteristics, and fails to dynamically optimize the fusion strategy of the model according to the characteristics of real-time input data (such as DSCD (O4), DOAS inversion RMS error, humidity, aerosol scattering index constructed by the patent, etc. which can indicate the level of aerosol), thus limiting the fine response ability of the model to the complexity of the real atmosphere.

[0007] (3) "Black box" dilemma and lack of credibility: Many machine learning models are still essentially "black boxes", with highly opaque internal decision-making processes. This not only makes it difficult for users to understand why the model gives a corresponding mass concentration prediction for a specific input, but also makes it difficult to judge the reliability of the prediction result, hindering the extraction of scientific knowledge consistent with physical and chemical laws from the model. Environmental monitoring requires high credibility, and the unexplainability of the model causes application bottlenecks. It is not enough to know which features are important to the model, but more importantly, to understand how the model responds to different atmospheric scenarios, whether its internal mechanisms conform to physical and chemical knowledge, and how to use this information to verify and optimize the model itself.

[0008] Therefore, there is an urgent need for a completely new solution that not only integrates the advantages of different machine learning models, but more importantly, dynamically perceives and adapts to different atmospheric conditions (especially variable aerosol scenarios) and optimally utilizes the strengths of each model through intelligent fusion mechanisms; The solution also needs to provide in-depth explainability analysis, not only to enhance the credibility of model prediction, but also to verify its scene adaptive design, guide the iterative optimization of the model, and promote the extraction of scientific laws from data-driven models. Such a solution has important theoretical significance and application value for realizing high-precision, high-robustness, high-timeliness, and reliable pollutant mass concentration hyperspectral remote sensing imaging to serve fine-grained atmospheric pollution monitoring. SUMMARY

[0009] The purpose of the present application is to provide an artificial intelligence-based hyperspectral remote sensing of pollutant gas mass concentration quantitative imaging method and device, which deeply integrates hyperspectral remote sensing imaging technology and explainable machine learning, and realizes high-precision, high-robustness, high-timeliness, and explainable pollutant mass concentration hyperspectral remote sensing imaging from path-integrated DSCD to pollutant gas plume point mass concentration according to DSCD, DOAS inversion error (RMS), online monitoring mass concentration data, and meteorological information.

[0010] The artificial intelligence-based hyperspectral remote sensing of pollutant gas mass concentration quantitative imaging method provided by the embodiment comprises the following steps: Step 1: Collecting spectral information of the pollutant gas plume, using differential optical absorption spectroscopy based on the spectral information to perform inversion, and constructing a data set D1; At the same time, collecting meteorological data of the location of the pollutant gas plume, and constructing a data set D2; Step 2: Taking the data set D1 and the data set D2 as inputs, based on a plurality of basic machine learning sub-modules, establishing a weighted and dynamic fusion mechanism to dynamically allocate the weights of each sub-module, constructing a scene-adaptive fusion machine learning model N1, using the online detection concentration data of the plume as a label set L1, optimizing the scene-adaptive fusion machine learning model N1, obtaining a model M1, and predicting the final mass concentration; Step 3: Quantify the marginal contribution of the dataset D1 and the dataset D2 to the final mass concentration prediction and its global importance using SHAP analysis, and feedback to optimize the model M1 to obtain the model M2; Step 4: Input the actual plume observation data into the model M2 to obtain the mass concentration image of the plume emission, realizing the hyperspectral remote sensing imaging of the differential slant column concentration to the mass concentration.

[0011] In one embodiment, the dataset D1 includes: the differential slant column concentration of NO2, the differential slant column concentration of O4, and the inversion root mean square error obtained by using the differential optical absorption spectroscopy for inversion; The dataset D2 includes: the wind speed, air pressure, temperature, and humidity at the location of the pollution gas plume.

[0012] In one embodiment, the plurality of basic machine learning sub-modules includes: an XGboost sub-module, a LightGBM sub-module, and a random forest sub-module; The XGboost sub-module adopts a layer-by-layer growing decision tree construction strategy, which takes the dataset D1 and the dataset D2 as input to predict the mass concentration in the scene without obvious or low aerosol influence, wherein the objective function of the XGboost sub-module is composed of a loss function and a regularization term The calculation formula is as follows: , , In the formula, represents the objective function value of the XGBoost sub-module, is the loss function of the th real mass concentration and the th predicted mass concentration , represents the total number of training samples, is the regularization term of the th decision tree, represents the total number of decision trees, represents the leaf node split penalty coefficient, represents the total number of leaf nodes of the th decision tree, represents the leaf weight L2 regularization coefficient, represents the weight value of the th leaf node in the th decision tree; th decision tree; The LightGBM submodule adopts a leaf growth decision tree construction strategy combined with gradient one-sided sampling and mutually exclusive feature bundling methods, and is based on datasets D1 and D2 as inputs to predict the mass concentration in scenarios with obvious or high aerosol influence; The random forest submodule is based on datasets D1 and D2 as inputs, and outputs a benchmark mass concentration for relatively stable hyperspectral remote sensing imaging in any scenario.

[0013] In one embodiment, the establishment of a weighted and dynamic fusion mechanism dynamically allocates the weights of each submodule by constructing a meta-learner for identifying and quantifying aerosol scene characteristics to fuse the prediction results of multiple basic machine learning submodules, including: based on the mass concentration output by each basic machine learning submodule, combining the aerosol scene applicable to each basic machine learning submodule, screening the input feature vector of the meta-learner, predicting the allocation weight of each basic machine learning submodule according to the input feature vector of the meta-learner, and outputting the fusion prediction.

[0014] The input features of the meta-learner are designed to specifically identify and quantify the current atmospheric scene characteristics, especially the aerosol pollution degree, rather than being directly used for predicting the mass concentration, so the input vector of the meta-learner is different from the complete input feature sets D1 and D2 of the three basic machine learning submodules. By dynamically allocating the weights of each submodule through the selected scene indicative feature vector, the prediction of the basic machine learning module that best describes the current atmospheric conditions is realized, and a higher confidence is given.

[0015] In one embodiment, the input feature vector of the meta-learner includes: The differential slant column density of O4 is used to judge the aerosol load and the complexity of the light path; The inversion root mean square error is used to measure the quality of the spectral signal and the degree of aerosol interference; Humidity is used to determine whether the aerosol scattering effect is amplified; The aerosol scattering index is used to reflect the change in nonlinear scattering caused by aerosols relative to the standard path length, and the calculation formula is: , In the formula, denotes the aerosol scattering index, RMS denotes the inversion root mean square error, and DSCD(O4) denotes the differential slant column density of O4.

[0016] Based on the above several types of key indicators, different weights are assigned for different aerosol scenarios. Among them, in the scene with high aerosol concentration, it leads to the abnormal decrease of DSCD(O4) value, so DSCD(O4) is the primary index to judge the aerosol load and the complexity of light path; High concentration of aerosol will introduce nonlinear broadband extinction effect, and may distort the absorption spectrum structure of the gas, leading to increased difficulty in DOAS fitting, thus producing significantly higher RMS value, so RMS is used as a direct indicator to measure the quality of spectral signal and the degree of aerosol interference; High humidity environment will promote the hygroscopic growth of hygroscopic aerosol, change its particle size, morphology and optical properties, thereby greatly enhancing its scattering ability, so humidity is a key meteorological factor to judge whether the scattering effect of aerosol is amplified.

[0017] In order to further enhance the robustness of scene recognition, a composite feature, namely aerosol scattering index, is specially constructed for scene recognition, which is used to more sensitively reflect the change of nonlinear scattering caused by aerosol relative to the standard path length, and provide a more stable and explicit scene classification basis for the meta-learner.

[0018] In one embodiment, the allocation weight of each basic machine learning submodule is predicted according to the input feature vector of the meta-learner, including: Based on the input feature vector of the meta-learner, for the low aerosol scene, the predicted mass concentration of the XGboost submodule is mainly trusted, and the allocation weight of each basic machine learning submodule output is represented as , wherein > , and + + =1, represents the weight of the XGboost submodule, represents the weight of the LightGBM submodule, represents the weight of the random forest submodule; Based on the input feature vector of the meta-learner, for the high aerosol scene, the predicted mass concentration of the LightGBM submodule is mainly trusted, and the allocation weight of each basic machine learning submodule output is represented as , wherein > , and + + =1, represents the weight of the XGboost submodule, represents the weight of the LightGBM submodule, represents the weight of the random forest submodule; Based on the input feature vector of the meta-learner, the weights of each sub-module are evenly distributed for transition or mixed scenes.

[0019] In one embodiment, the SHAP analysis is used to quantify the marginal contribution and global importance of dataset D1 and dataset D2 to the final mass concentration prediction, including: According to the scene indicative features including aerosol scattering index, differential slant column density of O4 and inversion root mean square error, dataset D1 and dataset D2 are divided into low aerosol scene subset, high aerosol scene subset and transition or mixed scene subset; the low aerosol scene subset has samples with high differential slant column density of O4, low inversion root mean square error and low aerosol scattering index; the high aerosol scene subset has samples with low differential slant column density of O4, high inversion root mean square error and high aerosol scattering index; the transition or mixed scene subset has samples between the low aerosol scene subset and the high aerosol scene subset; Based on the independent running of SHAP analysis for each scene subset, the effectiveness of each basic machine learning sub-module in the corresponding scene subset is determined.

[0020] In one embodiment, after the independent running of SHAP analysis for each scene subset, the content optimized by each basic machine learning sub-module is located based on the SHAP analysis results, including: In the low aerosol scene subset, macroscopic kinetic features are expected to dominate, but if it is found that the noise indicative features have higher SHAP values, it indicates that the XGboost sub-module has insufficient anti-noise ability in the low aerosol scene, wherein the macroscopic kinetic features include the differential slant column density of NO2 and the wind speed at the location of the pollution gas plume, and the noise indicative features include the inversion root mean square error; In the high aerosol scene subset, the differential slant column density of O4, the inversion root mean square error and the humidity feature and the feature related to the complexity of the light path are expected to have higher SHAP values and complex nonlinear dependence, but if it is found that the SHAP value contribution is lower than expected or the interaction is not significant, it indicates that the LightGBM sub-module has insufficient learning ability in the high aerosol scene.

[0021] In one embodiment, the feedback optimization model M1 includes: When it is located that the XGboost sub-module has insufficient anti-noise ability in the low aerosol scene, the L1 regularization coefficient and the L2 regularization coefficient of the XGboost sub-module are increased to punish excessive weights; When it is located that the LightGBM sub-module has insufficient learning ability in the high aerosol scene, the number of leaf nodes is increased to build a deeper decision number, the minimum data amount is reduced to pay attention to more subdivided difficult samples, and the feature ratio is adjusted to explore more feature combinations; When the contribution of the RF sub-module is suppressed by the XGboost sub-module and the LightGBM sub-module under any scene subset, the number of decision trees or the maximum number of features is increased to improve the diversity and stability of the RF sub-module prediction; The parameters of each basic machine learning sub-module are iteratively optimized until the SHAP analysis results on each scene subset meet the expected physical mechanism, and the prediction error of the model M1 is the lowest.

[0022] In another aspect, the present application also provides an artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging device, comprising a memory and a processor, the memory is used to store a computer program, and the processor is used to realize the artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method when the computer program is executed.

[0023] Compared with the prior art, the present application has at least the following beneficial effects: (1) The present application is based on multiple basic machine learning sub-modules, integrates the advantages of each sub-module, and establishes a lightweight and specially trained meta-learner to establish a weighted and dynamic fusion mechanism, dynamically allocates the weights of each sub-module, and constructs a scene-adaptive fusion machine learning model to realize dynamic perception and adaptive processing of the smoke plume mass concentration under real atmospheric conditions, especially complex and variable aerosol scenes.

[0024] (2) The input feature vector of the meta-learner is determined by the specific aerosol scene. In addition to considering the differential slant column density of oxygen dimers (O4) for judging aerosol load and light path complexity, the root mean square error for measuring spectral signal quality and aerosol interference degree, and humidity for judging whether the aerosol scattering effect is amplified, the aerosol scattering index is specially designed for scene recognition to sensitively reflect the change of nonlinear scattering caused by aerosol relative to the standard path length, providing a more stable and explicit scene classification basis for the meta-learner, so that the weight allocated by the meta-learner can accurately describe the prediction of the basic machine learning sub-module under atmospheric conditions, and give higher confidence.

[0025] (3) Further, it provides in-depth explainability analysis, which can better verify the scene-adaptive design of the model and guide the iterative optimization of the model, and realizes high-precision, high-robustness, high-timeliness and reliable pollution mass concentration hyperspectral remote sensing imaging prediction to serve fine atmospheric pollution monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below.

[0027] Figure 1 A schematic flowchart illustrating the artificial intelligence-based hyperspectral remote sensing method for quantitative imaging of pollutant gas mass concentration provided by this invention; Figure 2 This is a schematic diagram of the structure of the scene adaptive fusion machine learning model provided in the embodiment of the present invention.

[0028] Figure 3 A diagram illustrating the influencing factors for interpretability analysis of a model using the SHAP interpretability method, provided in this embodiment of the invention. Figure 4 The results of plume observation mass concentration are shown in the actual industrial scenario of this invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the accompanying drawings and... The embodiments further illustrate the present invention in detail. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0030] The core problem this invention aims to solve is how to effectively convert differential column concentration (DSCD), which is essentially a path integral, into the more critical mass concentration at a specific spatial location in practical applications. Given that DSCD represents the total number of gas molecules along an optical path, while mass concentration is an indicator of pollutant mass within a specific volume, the conversion between the two is nonlinearly influenced by multiple complex factors such as atmospheric radiative transfer, spatial distribution of pollutants, and meteorological conditions. Therefore, this invention proposes an artificial intelligence-based hyperspectral remote sensing method for quantitative imaging of pollutant gas mass concentration, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect spectral information of the pollutant gas plume, and use differential optical absorption spectroscopy to invert the data based on the spectral information to construct dataset D1; at the same time, collect meteorological data of the location of the pollutant gas plume to construct dataset D2.

[0031] In this embodiment, a hyperspectral device is used to acquire spectral information of the target pollutant gas plume. The acquired raw spectral data is then inverted using differential optical absorption spectroscopy (DOAS). The basic principle of DOAS is based on Beer-Lambert's law, which describes the attenuation of light intensity as it propagates through an absorbing medium. , in, It measures the spectral intensity after passing through the atmosphere; It is the spectral intensity of the initial light source, which is the non-absorbed solar zenith scattering spectrum; is the wavelength; is the differential absorption cross section of the gas with narrowband absorption characteristics; is the slant column density (SCD) of the gas, i.e. the total number of molecules of the gas along the light path. represents the broadband extinction caused by molecular scattering (Rayleigh scattering), aerosol scattering (Mie scattering), and instrument effects, etc.

[0032] The core of the DOAS method is to separate the narrowband absorption structure and broadband extinction and variation in the spectrum. By taking the logarithm of the ratio of the measured spectrum and the reference spectrum, and using polynomial fitting and other methods to remove the broadband structure, the differential optical thickness is obtained: , wherein, is the equivalent initial spectrum containing broadband effects, is the processed differential absorption cross section. By least squares fitting and other methods, the differential slant column density (DSCD) of various target gases (such as NO2 in the present application) and the differential slant column density (DSCD) of O4 are calculated from . The error (RMS) in the inversion process is also obtained. These data together constitute the data set D1.

[0033] At the same time, the latest ERA5 reanalysis product data released by the European Centre for Medium-Range Weather Forecasts (ECMWF) is synchronously acquired. This data set contains various meteorological parameters of the smoke plume and its surroundings, such as wind speed (u10, v10), air pressure (sp), temperature (t2m), humidity (d2m), etc. These meteorological information constitutes the data set D2.

[0034] Step 2: Take the data set D1 and the data set D2 as input, based on multiple basic machine learning sub-modules, establish a weighted and dynamic fusion mechanism to dynamically allocate the weights of each sub-module, construct a scene adaptive fusion machine learning model N1, mark the online detection concentration data of the smoke plume as the label set L1, optimize the scene adaptive fusion machine learning model N1, obtain the model M1, and predict the final mass concentration.

[0035] In the embodiment, as Figure 2As shown, an innovative hierarchical fusion machine learning model N1 with scene recognition and adaptive capabilities was constructed. The core innovation of this model lies in its unique system architecture. This architecture is not a simple superposition of existing machine learning models, but rather combines parallel basic machine learning modules (XGBoost, LightGBM, and RF sub-modules) optimized for different atmospheric conditions with an intelligent fusion mechanism capable of making dynamic decisions based on real-time input data features. This addresses the challenges of accuracy, robustness, and real-time performance in the hyperspectral remote sensing DSCD to mass concentration conversion process caused by the complex and variable conditions of the real atmosphere (especially aerosols). These challenges are difficult to effectively address with a single model or fixed-weight ensemble method. This model aims to overcome the inherent shortcomings of traditional radiative transfer models (RTM), such as limited parameter settings, difficulty in accurately representing complex atmospheric states, low computational efficiency, and inability to meet rapid monitoring needs.

[0036] Data preparation: Pollutant concentration data (in µg / m³) from online monitoring devices at locations corresponding to the plume were collected. This data served as the training label set L1. The specific construction of the scene-adaptive fusion machine learning model N1 is as follows: Based on multiple fundamental machine learning sub-modules, the sub-modules include: (1) XGBoost submodule: Customized for robust benchmark prediction in low aerosol or no significant aerosol scenarios.

[0037] Scene characteristics and the targeted design of this module: In scenarios with no significant aerosol emissions or low aerosol optical thickness (indicated by high DSCD(O4) values ​​and low RMS values), atmospheric radiative transfer is relatively clear. Although the relationship between DSCD(NO2) and the target mass concentration remains nonlinear, it is mainly dominated by macroscopic atmospheric dynamic processes (such as wind field, temperature, and air pressure). XGBoost's level-wise decision tree construction strategy is suitable for capturing global, dominant nonlinear trends under such conditions. Its objective function... By loss function and regularization term composition: , , In the formula, This represents the objective function value of the XGBoost submodule. For the first True mass concentration With the A loss function for predicting mass concentration. This represents the total number of training samples. For the first The regularization term of a decision tree. This represents the total number of decision trees. This represents the leaf node splitting penalty coefficient. Indicates the first The total number of leaf nodes in the decision tree. This represents the L2 regularization coefficient for the leaf weights. Indicates the first The first decision tree The weight values ​​of each leaf node.

[0038] Due to its robust structure and good fitting ability to moderate nonlinear relationships, the XGBoost submodule can efficiently learn from the input DSCD (NO2), DSCD (O4) and RMS values ​​indicating atmospheric clarity, as well as meteorological data such as wind speed (u10, v10), temperature (t2m), and air pressure (sp) provided by ERA5, and provide a reliable mass concentration baseline prediction for low aerosol impact scenarios.

[0039] (2) LightGBM submodule: customized for complex nonlinear strong interaction in high aerosol emission scenarios.

[0040] Scene characteristics and targeted design of this module: In high aerosol load scenarios (industrial plumes containing large amounts of particulate matter and water vapor, indicated by abnormally low DSCD(O4) values, significantly increased RMS values, and high humidity d2m), the strong forward and backward scattering and absorption of aerosols make the optical path extremely complex, strongly affecting the DSCD(NO2) signal. Its relationship with the target mass concentration exhibits a highly nonlinearity and is strongly influenced by multiple factors (such as aerosol optical properties, particle size distribution, interaction with humidity, and the mixing state of NO2 in the plume). LightGBM's leaf-wise strategy allows it to prioritize and deepen the decision paths that contribute most to reducing overall loss, thereby capturing these strong interactive effects introduced by high aerosol levels. Its gradient-based one-sided sampling (GOSS) technology, by focusing on "hard-to-learn" sample points with larger gradients, performs better on samples in the core region of the plume or in areas with high aerosol density and low signal-to-noise ratio and complex information. Meanwhile, mutually exclusive feature bundling (EFB) technology helps to address high-dimensional problems caused by numerous potentially relevant features introduced by aerosols.

[0041] LightGBM, due to its efficient learning mechanism for complex interactions, can mine and fit the nonlinear coupling relationship dominated by high-concentration aerosols from DSCD(NO2), DSCD(O4), RMS signals, and meteorological parameters related to aerosols (especially humidity d2m, and u10, v10 affecting the diffusion and evolution of pollutants and aerosols), providing a high-precision prediction source for the fusion module in high-aerosol scenarios.

[0042] (3) Random Forest (RF) submodule: customized to improve the generalization ability of the overall model and cope with scenario diversity.

[0043] Scenario characteristics and the pertinence design of this module: Real atmospheric scenarios often exist in the transition state between high and low aerosols, or rare meteorological and emission combinations that training data cannot completely cover. Although XGBoost and LightGBM are powerful gradient boosting models, their serial enhancement mechanism may make them sensitive to specific training patterns, and there is a risk of overfitting when the model structure is too complex. Random Forest builds a large number of independent decision trees and averages their predictions. This Bagging idea combined with feature random sampling naturally has excellent variance reduction ability and insensitivity to noisy data and outliers, thereby providing cross-scene stability. Its prediction result is the average of the prediction results of all D decision trees : The averaging effect can significantly reduce the variance and thus reduce the risk of overfitting.

[0044] The RF submodule in this invention focuses on stability and generalization. By providing predictions that are not dependent on specific scenarios (high or low aerosols) and have good overall generalization ability, it ensures that the entire fusion model N1 can maintain stability and reliability in the face of various unexpected atmospheric condition combinations. At the same time, as an integrated regularization method, it prevents the XGBoost and LightGBM submodules from over-relying on certain scene characteristics by introducing diversity, thereby improving the performance of the final model M1 on unknown real data.

[0045] Next, a weighted and dynamic fusion mechanism is established, one of the core innovations of this invention is embodied in the design and implementation of this mechanism, which enables model N1 to surpass the limitations of traditional static ensemble learning and gives it true scene adaptability. Collect independent quality concentration predictions of XGBoost, LightGBM, and RF three parallel basic machine learning, respectively denoted as , , .

[0046] The established weighting and dynamic fusion mechanism is different from the fixed weighted average, and a dynamic fusion mechanism based on the selected "scene indicative feature vector" in the current input data is realized. This mechanism is realized through a lightweight, specially trained meta-learner (Meta-Learner). The input features of the meta-learner are specially designed to specifically identify and quantify the current atmospheric scene characteristics (especially the aerosol pollution level), rather than being directly used for prediction of the concentration, so its input vector is different from the complete input feature set (D1 and D2) of the three basic models. Specifically as follows: The input feature vector of the meta-learner is specifically composed of the following key indicator factors, different sub-modules are assigned different weights, and the sub-modules are aimed at different aerosol scenes, so they focus on indicators that have a greater impact in specific aerosol scenes: DSCD(O4): Difference slant column density of oxygen dimer (O4). Since the concentration of O4 is relatively constant in the atmosphere, the change in its DSCD mainly reflects the actual transmission path length of photons in the atmosphere. In the scene with high aerosol concentration, the DSCD(O4) value abnormally decreases, which is the primary indicator for judging the aerosol load and the complexity of the light path.

[0047] RMS (root mean square error of inversion): Fitting residual in the DOAS inversion process. High concentration of aerosol will introduce nonlinear broadband extinction effect and may distort the absorption spectrum structure of the gas, resulting in increased difficulty in DOAS fitting, thus producing significantly higher RMS value. Therefore, RMS is a direct indicator to measure the quality of the spectral signal and the degree of aerosol interference.

[0048] d2m (humidity): High humidity environment will promote the hygroscopic growth of hygroscopic aerosols, change their particle size, morphology and optical properties, and thus greatly enhance their scattering ability. Therefore, d2m is a key meteorological factor for judging whether the scattering effect of aerosols is amplified.

[0049] Aerosol scattering index (ASI): In order to further enhance the robustness of scene recognition, a composite feature is constructed, which is specially proposed by the present patent and is the first of its kind, and the calculation formula is: , This ratio can more sensitively reflect the change in nonlinear scattering caused by aerosols relative to the standard path length, providing a more stable and explicit scene classification basis for the meta-learner.

[0050] The task of the meta-learner is to predict an optimal weight combination These weights are dynamically changing, aiming to give higher confidence to the prediction of the basic machine learning module that best describes the current atmospheric conditions.

[0051] For example, in the low-aerosol scenario (represented by the DSCD(O4) value in the normal range, the RMS value is low, and the d2m value is relatively low), the meta-learner mainly trusts the predicted mass concentration of the XGboost submodule after training, and the distribution weight of each basic machine learning submodule output is represented as , , , , , ,

[0052] In the high-aerosol scenario (represented by the abnormally low DSCD(O4) value, the significantly increased RMS value, and the high d2m value), the meta-learner shifts its focus to the LightGBM submodule that can efficiently handle complex interactions, i.e., in the high-aerosol scenario, mainly trusts the prediction results of the LightGBM submodule optimized for this scenario, while retaining the weight of the RF submodule to ensure generalization, and the distribution weight of each basic machine learning submodule output is represented as , , , , , ,

[0053] In the transition or mixed scenario, the weight distribution will be more balanced, reflecting the intelligent processing capability for uncertainty.

[0054] The fusion prediction after weight distribution is represented as: , wherein, , and is the weight function allocated by the meta-learner, reflecting dynamic adaptability.

[0055] ​​​​​​​Finally, through the fusion strategy of dynamic scene perception, the advantages of each basic machine learning submodule are utilized to construct a scene-adaptive fusion machine learning model N1, ensuring that the model N1 can optimally utilize the unique advantages of each basic machine learning module under different aerosol conditions and data characteristics, thereby improving the prediction accuracy and robustness of the conversion from DSCD to mass concentration under real atmospheric conditions.

[0056] In the embodiment, the data sets D1 and D2 (meteorological information) are further taken as input features of the scene-adaptive fusion machine learning model N1, and the label set L1 (online monitoring mass concentration data) is taken as the output label. The data set is used to independently train and optimize the hyperparameters of the three basic machine learning submodules XGBoost, LightGBM and RF, and the goal is to minimize the mean square error of each submodule to the label set L1. After the training of the basic machine learning module is fixed, the meta-learner is trained. The training goal of this stage is to minimize the mean square error (MSE) between the final fusion prediction and the real label of the validation set. This step is the key to realizing the scene-adaptive capability of the present application, and the predictions of each basic machine learning module are optimally combined by learning the real-time scene indicating features of the input. Finally, a high-precision and universal mass concentration prediction model M1 is obtained.

[0057] Step 3: SHAP analysis is used to quantify the marginal contribution of the data set D1 and the data set D2 to the final mass concentration prediction and its global importance, and the model M1 is fed back to optimize the model M2.

[0058] In the embodiment, in order to understand how the model M1 of the dynamic scene-adaptive fusion mechanism proposed by the present application makes the final mass concentration prediction based on multiple source inputs and improves the credibility of the model, the present application innovatively integrates the SHAP (SHapley Additive exPlanations) analysis process as an indispensable part of model development, verification and application: 1. Revealing the internal decision-making mechanism and scene response of the fusion model: the present application applies SHAP analysis to reveal the internal decision-making logic of the aforementioned scene-adaptive fusion model M1 containing XGBoost, LightGBM, RF and intelligent weighting and dynamic fusion modules. This includes: quantifying the marginal contribution (SHAP value) of each original input feature, i.e. the data set D1 and the data set D2, to the final mass concentration prediction and its global importance, as shown in Figure 3 An example of SHAP feature importance is shown. SHAP is based on the Shapley value in game theory, which decomposes the model's prediction output for a specific sample into the contribution value of each feature. For a given prediction , the SHAP value represents the​ Each feature contributes to the prediction. For a specific prediction, the sum of its SHAP values ​​plus the baseline value (the average of all training sample predictions) equals the predicted value: , in, It is the number of features; It is the expected output (baseline) of the model across the entire training dataset. Indicates the first The contribution of each feature to the prediction is used to represent the feature. The presence of this factor pushes the predicted value away from the baseline by how much. The SHAP value calculates the marginal contribution of each feature by considering all possible subsets of features and then weights them as an average: , in: It is the set of all features; yes The middle does not contain features A subset of features; When the feature subset The eigenvalues ​​in the given eigenvalues ​​are known, and the eigenvalues ​​are known. When the value of is also known, the model's predicted output; When the feature subset The eigenvalues ​​in the model are known, but the eigenvalues ​​are unknown. The model's predicted output when the value of is unknown.

[0059] By comparing and analyzing the distribution and relative importance of the SHAP values ​​of various input features under different aerosol levels characterized by parameters such as DSCD(O4) and RMS, this study verifies and understands how the contributions of different basic machine learning modules (such as LightGBM and XGBoost) in the fusion model M1 are reflected through the fusion mechanism in different scenarios. For example, it can be verified whether features related to aerosol properties and complex optical paths (such as DSCD(O4), RMS, and d2m) exhibit a greater SHAP influence as expected in high aerosol scenarios.

[0060] 2. SHAP-based Physical Mechanism Verification and Proactive Model Iterative Optimization: The innovation of this invention lies not only in interpreting the model's prediction results, but also in establishing a dynamic feedback loop from "interpretability insight" to "backward optimization of model structure and parameters." This process transforms SHAP analysis from a passive verification tool into a proactive and refined model tuning guide, ensuring that the scenario adaptation mechanism of model M1 is not only logically sound but also optimally implemented at the parameter level, better fitting the model optimization for specific application scenarios. The specific optimization path is as follows: (1) Differentiated SHAP analysis of scene slices: First, according to the scene indicative features such as aerosol scattering index (ASI), DSCD(O4), RMS, the validation dataset is divided into different physical scene subsets. For example: low aerosol scene subset: samples with high DSCD(O4), low RMS and low ASI value; high aerosol scene subset: samples with low DSCD(O4), high RMS and high ASI value; transition scene subset: samples between the two. Then, instead of a general SHAP analysis of the model M1, SHAP is independently run on each scene subset to obtain feature contribution graphs and feature interaction graphs for specific physical situations.

[0061] (2) From "expected validation" to "mismatch judgment": Through the comparison of SHAP analysis results of different scene subsets, the "expected validation" is deepened to the "mismatch judgment", so as to accurately locate the link that needs to be optimized by the model. Specifically: judge the effectiveness of the LightGBM submodule in the high aerosol scene: in the high aerosol scene subset, features such as DSCD(O4), RMS and humidity (d2m) related to light path complexity are expected to show very high SHAP values and complex nonlinear dependence. If it is found that the SHAP value contribution of these features is lower than expected, or its interaction is not significant, it indicates that the LightGBM submodule customized for this scene has failed to learn the strong interaction effect dominated by aerosol. Judge the robustness of the XGBoost submodule in the low aerosol scene: in the low aerosol scene subset, macrodynamic features such as DSCD(NO2) and wind field (u10, v10) are expected to dominate. If it is found that noise indicating features such as RMS still have high SHAP values, it indicates that the XGBoost submodule may be too sensitive to noise in the data, and its regularization is insufficient.

[0062] (3) Convert SHAP analysis into precise model parameter optimization instructions, which is the most innovative step of the present application. The above judgment results are directly mapped to specific parameter adjustment instructions for different components of the model M1, forming a precise iterative optimization closed loop: (a) Guide LightGBM submodule parameter adjustment: If it is judged that the LightGBM submodule has insufficient learning ability in the high aerosol scene, the model complexity is increased to capture strong interaction. Specific operations include: increasing num_leaves to allow the construction of deeper decision trees, reducing min_child_samples to allow the model to pay attention to more subdivided "difficult to learn" samples, and adjusting feature_fraction (feature ratio) to explore more feature combinations.

[0063] (b) Guidance for XGBoost submodule parameter adjustment: if it is judged that the XGBoost submodule has insufficient noise resistance in the low-aerosol scenario, the regularization strength is specifically enhanced to improve the robustness of the model. The specific operation includes: increasing the regularization coefficient lambda (L2 regularization) and alpha (L1 regularization), which will punish excessive weights and force the model to focus on more universal features.

[0064] (c) Guidance for RF submodule parameter adjustment: as a guarantee of generalization ability, if it is found that the contribution of the RF submodule is suppressed by the other two models in all scenarios, n_estimators (the number of decision trees) or max_features (the maximum number of features) can be appropriately increased to enhance the diversity and stability of its prediction.

[0065] Through the above iterative cycle of "scenario slicing analysis → mismatch judgment → precise parameter tuning", the process is repeatedly performed until the SHAP analysis results on each scenario subset meet the expected physical mechanism, and the model M1 achieves the lowest prediction error on the independent test set. This iterative optimization process ensures that the scene adaptation ability of the model is real and efficient, greatly enhancing the credibility and accuracy of the final prediction results in real and complex environments, and more closely fitting the real, complex and variable atmospheric scenarios.

[0066] 3. Enhance the credibility of model application: by clearly showing how model M1 dynamically adjusts its internal judgment basis according to the input real-time data features, and giving feature contributions that conform to the intuition of atmospheric physical and chemical processes, the user's trust in the stability and accuracy of the prediction results of the method in complex and variable environments is significantly enhanced, overcoming the application barriers of traditional "black box" models.

[0067] In the embodiment, an artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging device is also provided, which comprises a memory and a processor, the memory is used to store a computer program, and the processor is used to realize the artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method when the computer program is executed.

[0068] Step 4: input the actual plume observation data into the model M2 to obtain the mass concentration image of the plume emission, and realize the hyperspectral remote sensing imaging of the difference oblique column concentration to the mass concentration.

[0069] In the embodiment, the trained final model M2 is applied to the actual plume observation data. The model M2 can quickly extract feature information directly from the input differential slant column concentration data and meteorological data, and output the corresponding mass concentration value, without relying on a complex radiative transfer model (RTM) to reconstruct the real atmospheric state. By predicting each pixel covered by the plume, a mass concentration spatial distribution image of the plume emission is generated, as shown in FIG. 6. Figure 4

[0070] It should be noted that the above-mentioned embodiment provides an artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method and device. When performing hyperspectral remote sensing of pollution gas mass concentration quantitative imaging prediction, the above-mentioned each functional module should be divided for example, and the above-mentioned functions can be completed by different functional modules according to the needs, that is, the internal structure of the terminal or server is divided into different functional modules to complete all or part of the above-described functions. In addition, the artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method and device provided in the above embodiment belong to the same concept, and the specific implementation process is described in detail in the artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method embodiment. Here, it is not repeated.

[0071] In summary, the present application is based on multiple basic machine learning sub-modules, integrates the advantages of each sub-module, distributes the weights of each sub-module by establishing a weighted and dynamic fusion mechanism, constructs a scene-adaptive fusion machine learning model, realizes dynamic perception and adaptive processing of plume emission mass concentration under real atmospheric conditions, especially complex and variable aerosol scenes, and further provides in-depth explainable analysis, not only enhances the credibility of plume emission mass concentration prediction, but also verifies the scene-adaptive design of the model and guides the iterative optimization of the model, realizes high-precision, high-robustness, high-timeliness and reliable pollution mass concentration hyperspectral remote sensing imaging prediction, and serves fine atmospheric pollution monitoring, which has important theoretical significance and application value.

[0072] The specific embodiments described above detail the technical solutions and beneficial effects of the present application. It should be understood that the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modifications, supplements, and equivalent replacements made within the principles of the present application should be included in the protection scope of the present application.​

Claims

1. An artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method, characterized in that, The method comprises the following steps: Step 1: collecting spectral information of the pollution gas plume, performing inversion based on the spectral information using the differential optical absorption spectroscopy method, and constructing a data set D1; at the same time, collecting meteorological data of the position where the pollution gas plume is located, and constructing a data set D2; Step 2: taking the data set D1 and the data set D2 as inputs, based on multiple basic machine learning sub-modules, establishing a weighted and dynamic fusion mechanism to dynamically distribute the weights of the sub-modules, constructing a scene-adaptive fusion machine learning model N1, taking the online detection concentration data of the plume as a label set L1, optimizing the scene-adaptive fusion machine learning model N1, obtaining a model M1, and predicting the final mass concentration; Step 3: using SHAP analysis to quantify the marginal contribution and global importance of the data set D1 and the data set D2 to the prediction of the final mass concentration, and feeding back the optimization model M1 to obtain a model M2; Step 4: inputting the actual plume observation data into the model M2 to obtain a mass concentration image of the plume emission, and realizing hyperspectral remote sensing imaging from differential slant column density to mass concentration.

2. The artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method according to claim 1, characterized in that, The data set D1 comprises: the differential slant column density of NO2, the differential slant column density of O4 and the inversion root mean square error obtained by using the differential optical absorption spectroscopy method for inversion; The data set D2 comprises: the wind speed, air pressure, temperature and humidity at the position where the pollution gas plume is located. 3.The method of claim 1, wherein, The multiple basic machine learning sub-modules comprise: an XGboost sub-module, a LightGBM sub-module and a random forest sub-module; The XGboost submodule adopts a layer-by-layer growing decision tree construction strategy, takes the data set D1 and the data set D2 as inputs, and predicts the mass concentration in a scene without obvious or low aerosol influence, wherein the objective function of the XGboost submodule is composed of a loss function and a regularization term , and the calculation formula is as follows: ， ， In the formula, This represents the objective function value of the XGBoost submodule. For the first True mass concentration With the Predicted mass concentration loss function, This represents the total number of training samples. For the first The regularization term of a decision tree. This represents the total number of decision trees. This represents the leaf node splitting penalty coefficient. Indicates the first The total number of leaf nodes in the decision tree. This represents the L2 regularization coefficient for the leaf weights. Indicates the first The first decision tree The weight values ​​of each leaf node; The LightGBM sub-module adopts a leaf growth decision tree construction strategy and combines gradient one-side sampling and mutual exclusion feature bundling methods, takes the data set D1 and the data set D2 as inputs, and predicts the mass concentration in the scene with obvious or high aerosol influence; The random forest sub-module takes the data set D1 and the data set D2 as inputs, and outputs a benchmark mass concentration, which is used for relatively stable hyperspectral remote sensing imaging in any scene.

4. The artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method according to claim 2, characterized in that, The dynamic distribution of the weights of the sub-modules by the weighted and dynamic fusion mechanism is performed by constructing a meta-learner for identifying and quantifying the aerosol scene characteristics and fusing the prediction results of the multiple basic machine learning sub-modules, comprising: based on the mass concentration output by each basic machine learning sub-module, combining the aerosol scene applicable to each basic machine learning sub-module, screening the input feature vector of the meta-learner, predicting the distribution weight of each basic machine learning sub-module according to the input feature vector of the meta-learner, and outputting the fusion prediction.

5. The artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method according to claim 4, characterized in that, The input feature vector of the meta-learner comprises: The differential slant column density of O4 is used to judge the aerosol load and light path complexity; The inversion root mean square error is used to measure the spectral signal quality and aerosol interference degree; The humidity is used to judge whether the aerosol scattering effect is amplified; The aerosol scattering index is used to reflect the change of the nonlinear scattering caused by the aerosol relative to the standard path length, and the calculation formula is: ， wherein denotes the aerosol scattering index, RMS denotes the inverse root mean square error, and DSCD(O4) denotes the differential slant column density of O4.

6. The artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method according to claim 5, characterized in that, The prediction of the distribution weight of each basic machine learning sub-module according to the input feature vector of the meta-learner comprises: Based on the input feature vector of the meta-learner, for the low aerosol scene, the predicted mass concentration of the XGboost submodule is mainly trusted, and the distribution weight of each basic machine learning submodule output is expressed as wherein and =1, represents the XGboost submodule weight, represents the LightGBM submodule weight, represents the random forest submodule weight;​​​ Based on the input feature vector of the meta-learner, for the high aerosol scene, the distribution weight of each basic machine learning submodule output is represented as wherein and + =1, represents the XGboost submodule weight, represents the LightGBM submodule weight, represents the random forest submodule weight;​​ Based on the input feature vector of the meta-learner, the weights of each sub-module are evenly distributed for the transition or mixed scene.

7. The artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method according to claim 6, characterized in that, The SHAP analysis quantifies the marginal contribution and global importance of the dataset D1 and the dataset D2 to the final mass concentration prediction, including: According to the scene indicative features including the aerosol scattering index, the differential slant column density of O4 and the inversion root mean square error, the dataset D1 and the dataset D2 are divided into a low aerosol scene subset, a high aerosol scene subset and a transition or mixed scene subset; the low aerosol scene subset has samples with high differential slant column density of O4, low inversion root mean square error and low aerosol scattering index; the high aerosol scene subset has samples with low differential slant column density of O4, high inversion root mean square error and high aerosol scattering index; the transition or mixed scene subset has samples between the low aerosol scene subset and the high aerosol scene subset; Based on the independent running of the SHAP analysis for each scene subset, the effectiveness of each basic machine learning sub-module in the corresponding scene subset is determined.

8. The artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method according to claim 7, characterized in that, After the independent running of the SHAP analysis for each scene subset, the content optimized by each basic machine learning sub-module is located based on the SHAP analysis results, including: In the low aerosol scene subset, it is expected that macroscopic kinetic features should dominate, but if it is found that the noise indicative features have higher SHAP values, it indicates that the XGboost sub-module has insufficient anti-noise ability in the low aerosol scene, wherein the macroscopic kinetic features include the differential slant column density of NO2 and the wind speed at the position of the pollution gas plume, and the noise indicative features include the inversion root mean square error; In the high aerosol scene subset, it is expected that the differential slant column density of O4, the inversion root mean square error and the humidity features and the features related to the complexity of the light path show higher SHAP values and complex nonlinear dependence, but if it is found that the SHAP value contribution is lower than expected or the interaction is not significant, it indicates that the LightGBM sub-module has insufficient learning ability in the high aerosol scene.

9. The artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging method according to claim 8, characterized in that, The feedback optimization model M1 includes: When it is located that the XGboost sub-module has insufficient anti-noise ability in the low aerosol scene, the L1 regularization coefficient and the L2 regularization coefficient of the XGboost sub-module are increased to punish excessive weights; When it is located that the LightGBM sub-module has insufficient learning ability in the high aerosol scene, the number of leaf nodes is increased to construct deeper decision numbers, the minimum data amount is reduced to pay attention to more subdivided difficult samples, and the feature ratio is adjusted to explore more feature combinations; When it is located that the contribution of the RF sub-module is suppressed by the XGboost sub-module and the LightGBM sub-module in any scene subset, the number of decision trees or the maximum number of features is adjusted to improve the diversity and stability of the prediction of the RF sub-module; The parameters of each basic machine learning sub-module are iteratively optimized until the SHAP analysis results in each scene subset meet the expected physical mechanism, and the prediction error of the model M1 is the lowest.

10. An artificial intelligence-based hyperspectral remote sensing of pollution gas mass concentration quantitative imaging device, comprising a memory and a processor, the memory is used to store a computer program, characterized in that, The processor is configured to implement the artificial intelligence-based hyperspectral remote sensing pollution gas mass concentration quantitative imaging method of any one of claims 1-9 when executing the computer program.

Images