Methane inversion algorithm sensitivity evaluation system and method based on airborne hyperspectrum

Abstract

The invention discloses a methane inversion algorithm sensitivity evaluation system based on airborne hyperspectrum, and belongs to the technical field of atmosphere remote sensing and environment monitoring. The method comprises the following steps: firstly, through a hyperspectral data simulation module, fusing a WRF-LES mode and an MODTRAN model to generate hyperspectral data containing different methane emission rates; setting 8 land coverage types and two albedo of 0.1 and 0.4 by an earth surface parameter configuration module, and constructing 32 combined scenes; mF, RWL1MF and DOAS algorithms in an inversion algorithm integration module are called for inversion, performance indexes are calculated through a sensitivity analysis module, and the sensitivity of the algorithms to earth surface parameters is analyzed; and finally, comparing the performance of the DRAMF with that of the existing algorithm by using an improved algorithm verification module. According to the method, the algorithm sensitivity can be systematically evaluated, a scientific basis is provided for algorithm optimization, meanwhile, the precision and efficiency of methane emission monitoring are improved, and powerful support is provided for environmental governance and climate change research.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of atmospheric remote sensing and environmental monitoring, and specifically relates to a method and system for simulating and evaluating the sensitivity of methane concentration inversion algorithms using airborne hyperspectral data. This technology can be widely applied in the performance evaluation of atmospheric methane concentration inversion algorithms, and has important value in the fields of methane emission monitoring, climate change research, and environmental governance. Through this application, it can provide scientific and systematic basis for accurately selecting and optimizing methane inversion algorithms in different scenarios, and help to improve the accuracy and efficiency of methane emission monitoring, providing strong support for global climate regulation and environmental governance decisions.

[0002] In practical applications, this technology can serve specific scenarios such as oil and gas field methane leakage monitoring, landfill emission evaluation, and agricultural source methane emission accounting. For example, in oil and gas field monitoring, by evaluating the sensitivity of different algorithms in complex oil field areas (such as drilling platforms, vegetation-covered areas, and bare soil), the most suitable inversion algorithm can be selected to improve the accuracy of leak detection; in landfill monitoring, by selecting appropriate algorithms for different surface types such as vegetation, soil, and water around the landfill, methane emissions can be more accurately quantified. In addition, this technology provides more reliable methane concentration data input for climate change models, which helps to improve the accuracy of global warming trend prediction. BACKGROUND

[0003] Methane, as a potent greenhouse gas, has contributed more than 20% to global climate warming, and its global warming potential value is 84 times that of carbon dioxide on a 20-year time scale, having a profound impact on global climate change. According to the International Energy Agency (IEA), global methane emissions in 2022 were approximately 590 million tons, with the energy industry, agricultural activities, and waste disposal being the three major sources of emissions. Therefore, accurately monitoring the concentration distribution of methane in the atmosphere and the location and intensity of emission sources is of great significance for developing effective climate regulation strategies and promoting environmental governance.

[0004] Among the many methane monitoring techniques, remote sensing technology has become a key method for methane emission monitoring due to its unique advantages of large-scale and continuous observation. Among them, airborne hyperspectral remote sensing technology can capture the subtle features of methane point source plume because of its high spatial resolution (up to meter level) and spectral resolution (up to nanometer level), and plays an irreplaceable role in methane point source plume detection and concentration inversion. Currently, the main airborne methane inversion algorithms mainly include matching filter (MF), reflectance correction re-weighted matching filter (RWL1MF), and differential optical absorption spectroscopy (DOAS) and so on. These algorithms analyze the characteristic absorption spectrum of methane (mainly in the 2122-2485 nm band) in hyperspectral data to achieve methane concentration inversion.

[0005] However, the inversion accuracy of these algorithms is greatly affected by the heterogeneity of the surface spectrum. The change of surface albedo and the diversity of land cover types will cause complex changes in the surface reflectance spectrum, thereby interfering with the extraction of methane characteristic absorption signals, resulting in deviations in the inversion results, such as false positive detection (misjudging non-methane emission areas as emission sources) or missed detection (failing to identify actual methane emission sources). For example, in the high-albedo cement roof area, the surface reflectivity can reach more than 0.4, and the strong reflection signal will mask the weak absorption characteristics of methane, resulting in a missed detection rate of MF algorithm as high as 35%; while in the vegetation cover area, the high reflection peak of vegetation in the near-infrared band overlaps with the methane absorption band, which will cause the DOAS algorithm to have a false positive rate of about 28%.

[0006] The existing evaluation methods of methane inversion algorithms have obvious limitations. Most evaluations focus on comparison and verification under clear sky conditions and sparse ground monitoring network, which is difficult to systematically control and analyze the impact of surface parameters (such as albedo and land cover type) on algorithm performance. Real airborne hyperspectral data is affected by a variety of factors such as atmospheric conditions (such as clouds and aerosol optical thickness), surface conditions (such as vegetation coverage and soil moisture), and various factors are intertwined, making it difficult to separate the independent effect of surface parameters on algorithm performance. For example, in a flight observation, there are high-albedo surfaces and high-concentration aerosols in a certain area, and the algorithm inversion error may be caused by both, making it difficult to distinguish their respective contributions, resulting in inaccurate evaluation of the applicability of the algorithm in different surface scenarios.

[0007] Although some studies use simulated data to test algorithms, the richness and authenticity of the simulation scenarios are insufficient. For example, the simulated land surface types are single, mostly focusing on vegetation or bare soil, and fail to cover complex types such as roofs and asphalt pavements in urban areas. The albedo settings lack gradient variation and mostly use a single value (e.g., 0.2), which cannot fully reflect the sensitivity of the algorithm to different albedo conditions (e.g., 0.1-0.5). The simulated methane plume distribution is too idealized, mostly using a Gaussian distribution model, ignoring the plume fragmentation and uneven diffusion caused by atmospheric turbulence, which is quite different from the actual diffusion of methane in the atmosphere. These problems make it impossible for simulated data to fully and realistically reflect the sensitivity of the algorithm to complex surface parameters, greatly reducing the reliability and practicality of the evaluation results.

[0008] In addition, existing evaluation methods lack a unified quantitative index system and mostly use qualitative descriptions or a single index (such as root mean square error) for evaluation, making it difficult to fully depict the performance of the algorithm. For example, some studies only evaluate the algorithm by comparing the inversion concentration with the true value, ignoring the detection accuracy of the emission source location, resulting in one-sided evaluation results. At the same time, the evaluation process is mostly manual, lacking automated evaluation tools, which is inefficient and difficult to meet the evaluation needs of a large number of simulation scenarios.

[0009] Therefore, there is an urgent need for a method and system that can evaluate the sensitivity of methane inversion algorithms to surface albedo and land cover types through high-fidelity hyperspectral simulation experiments, to address the problems of insufficient variable control, limited evaluation scenarios, low authenticity of simulation data, and non-uniform index system in existing evaluation methods, and to provide strong support for the optimization and application of methane inversion algorithms. SUMMARY

[0010] (I) Invention purpose

[0011] The core purpose of the present application is to provide a method and system for evaluating the sensitivity of airborne methane inversion algorithms using hyperspectral simulation. Specifically, by constructing high-fidelity hyperspectral simulation scenarios, the sensitivity of mainstream methane inversion algorithms (MF, RWL1MF, DOAS) to parameters such as surface albedo (0.1-0.4 gradient) and land cover types (8 typical types) can be accurately and systematically controlled, and the sensitivity of these parameters can be fully and deeply evaluated. At the same time, in view of the shortcomings of existing algorithms in complex surface scenarios, an improved dynamic regularization adaptive matched filter (DRAMF) algorithm is proposed to improve the accuracy and robustness of methane inversion, ultimately providing scientific and reliable basis for the reasonable selection and effective optimization of algorithms in actual methane emission monitoring.

[0012] Specific objectives include: building a hyperspectral simulation dataset containing 32 controllable scenarios, achieving independent control of surface parameters and emission intensity; establishing a sensitivity evaluation system for quantitative indicators to achieve a comprehensive characterization of algorithm performance; developing the DRAMF algorithm to improve the model accuracy index (F1 score) by more than 30% on average compared to existing algorithms in complex surface scenarios; and forming an automated evaluation process.

[0013] (B) Technical solutions

[0014] To achieve the above-mentioned objectives and solve the problems in the prior art, the technical solutions of the present application are as follows:

[0015] A system for evaluating the sensitivity of airborne methane inversion algorithms using hyperspectral simulation, which is composed of five core modules: a hyperspectral data simulation module, a surface parameter configuration module, an inversion algorithm integration module, a sensitivity analysis module, and an improved algorithm verification module, as shown in Figure 1 Each module is independent of each other but works cooperatively through data interfaces to complete the evaluation of the sensitivity of airborne methane inversion algorithms.

[0016] The hyperspectral data simulation module is the core data generation unit of the entire system, responsible for simulating the methane-related hyperspectral data acquired by airborne hyperspectral imagers. The quality of the generated data directly determines the reliability of the subsequent evaluation results. This module uses deep fusion technology to organically combine the Weather Research and Forecasting Model (WRF-LES) and the MODTRAN radiation transfer model, enabling the generation of highly realistic methane plume and radiation transfer data.

[0017] In terms of methane plume simulation, the WRF-LES model plays a key role. This model generates precise three-dimensional distribution data of methane plumes with spatiotemporal dynamic characteristics at a horizontal resolution of 30x30m and a grid size of 128x128. To realistically simulate the methane diffusion process under different weather conditions, the model is set to 100W / m 2The uniform heat flux driven buoyancy turbulence is combined with the mechanical turbulence generated by surface drag (set the aerodynamic roughness height to 0.1 m) to make the simulated methane plume truly reflect the influence of atmospheric turbulence on methane diffusion. For example, under the action of buoyancy turbulence, the methane plume will form an upward vortex structure, while the mechanical turbulence will make the plume present irregular diffusion in the horizontal direction. In addition, this module supports flexible setting of different methane emission rates (500 kg / h, 1000 kg / h) simulation scenarios, and the total methane mass in the plume is accurately scaled to achieve precise adjustment of the emission rate, where 500 kg / h corresponds to small and medium-sized emission sources (such as small oil and gas leaks), and 1000 kg / h corresponds to large emission sources (such as landfill concentrated emissions), providing a diversified data basis for subsequent evaluation of the performance of the algorithm under different emission intensities.

[0018] In the radiation transmission simulation, the MODTRAN5.4 radiation transmission model accurately calculates the bidirectional transmittance of the atmosphere layer below the sensor according to the preset atmospheric model (such as model 7, representing mid-latitude summer), atmospheric path type (such as type 1, standard path), sensor height (such as an altitude of 5 km), and other parameters, simulates the reflected solar radiation spectrum. The model takes into account multiple physical processes such as absorption of atmospheric molecules (such as oxygen, water vapor), aerosol scattering, and ground reflection, with a calculation accuracy of 0.01%. The module internally stores a rich library of sensor response functions, covering Gaussian convolution parameters in different wavelength ranges, and can strictly follow the technical specifications of the Airborne Visible / Infrared Imaging Spectrometer Next Generation (AVIRIS-NG) to perform convolution and resampling processing on the simulated spectrum, ultimately generating a hyperspectral data cube that meets the 432-band, approximately 5 nm full-width-at-half-maximum. In order to further approach the characteristics of actual observation data, the radiation spectrum of each pixel includes simulated noise added based on the signal-to-noise ratio (SNR), with a signal-to-noise ratio of 100-300 in the visible light band (400-700 nm) and 50-200 in the near-infrared band (700-2500 nm). By precisely controlling the noise level, the simulated data can truly reflect the noise characteristics of the sensor in different wavelength bands.

[0019] By Beer-Lambert's law, the concentration of different plumes simulated by WRF-LES is converted into transmittance, and multiplied by the reflected radiance generated by MODTRAN5.4 to obtain the reflected radiance data of different gas plume concentrations under different reflectivity backgrounds, which can be used as data support for methane concentration inversion and sensitivity analysis.

[0020] The ground parameter configuration module is a data expansion unit of the entire system and is a key component for implementing ground factor variable control. It is mainly used to configure different ground albedo and land cover type parameters, to provide fine ground feature input for the hyperspectral data simulation module, and to ensure that the simulated scene can fully reflect the influence of different ground conditions on the methane retrieval algorithm. The data generated by this module is directly used in the retrieval algorithm integration module.

[0021] This module contains a systematic arrangement and optimized spectral library storage unit, which stores the reflectance spectral data of 8 typical ground types, including interference (white commercial roof, green sports field), water body (lake), rock, non-photosynthetic vegetation (coniferous deciduous), paved surface (airport asphalt), roof (red tile roof), soil (bare soil), etc. These spectral data are derived from the authoritative South California Natural and Urban Environment Spectral Library (SERC) and are strictly matched with the spectral characteristics of AVIRIS-NG (including convolution and resampling to match the full width at half maximum and band center), ensuring the accuracy and applicability of the spectral data. The spectral data of each ground type contains 432 bands, consistent with the simulated hyperspectral data cube bands, facilitating data fusion.

[0022] This module has a high-precision albedo adjustment function. Its core principle is to achieve precise control of the target albedo value (0.1, 0.4) by globally scaling the reflectance and combining the weighted integral of the solar spectrum. Specifically, it is calculated by the formula , where R(λ) is the reflectance at wavelength λ, E(λ) is the solar irradiance at wavelength λ (using the ASTM G173-03 standard solar spectrum), and the integral range covers 400-2500 nm (visible-near infrared). Through the linear scaling formula "R target (λ) = R orig (λ) x target albedo / original albedo" (R target (λ) is the target reflectance and R orig (λ) is the original reflectance), the reflectance is precisely adjusted, ensuring that the shape characteristics of the reflectance spectrum remain unchanged and the albedo value is precisely controlled. For example, when the original albedo of the soil spectrum is adjusted from 0.2 to 0.4, the reflectance value of each band is multiplied by 2, so that the adjusted spectrum maintains the soil spectral characteristics (such as the water absorption peaks at 1400 nm and 1900 nm) while the overall albedo is doubled.

[0023] The module can be configured with different surface types and albedo in full combination to generate a set of surface scene parameters containing 32 combinations (8 surface types x 2 albedo x 2 emission rates). This comprehensive combination provides a strict variable control basis for system analysis algorithm sensitivity to surface factors, and can accurately identify the independent influence of different surface parameters on algorithm performance. For example, by comparing the algorithm performance in the "white commercial roof + albedo 0.4 + emission rate 500 kg / h" and "white commercial roof + albedo 0.1 + emission rate 500 kg / h" scenarios, the influence of albedo can be analyzed independently; by comparing the "green sports field + albedo 0.4 + emission rate 500 kg / h" and "white commercial roof + albedo 0.4 + emission rate 500 kg / h" scenarios, the influence of land cover type can be analyzed.

[0024] The inversion algorithm integration module is the core execution unit of the algorithm evaluation of the entire system, receives the output results of the hyperspectral data simulation module and the surface parameter configuration module, and is responsible for integrating the matching filter (MF), the albedo correction re-weighted matching filter (RWL1MF), and the differential optical absorption spectrum (DOAS) three mainstream methane inversion algorithms to realize unified and efficient operation and management of these algorithms on simulated hyperspectral data.

[0025] It contains carefully optimized core calculation libraries for each algorithm, and is modularized and packaged according to the mathematical principles and calculation processes of each algorithm. For example, the MF algorithm library contains sub-modules such as target spectrum generation, covariance matrix calculation, and filter coefficient solving; the DOAS algorithm contains sub-modules such as spectrum preprocessing, reference spectrum fitting, and concentration inversion. This modular design makes the algorithm call more flexible and convenient, allowing quick invocation of the corresponding algorithm for methane concentration inversion of hyperspectral simulation data according to evaluation needs, greatly improving the efficiency of evaluation work.

[0026] To ensure the consistency and comparability of the inversion results, the module designs a unified input and output interface for each algorithm. The input parameters include hyperspectral data cubes, methane absorption characteristic wavelength range (2122-2485 nm), and other key information, and the output results are presented in the form of two-dimensional distribution images of methane column concentration enhancement (spatial resolution 30x30m, consistent with the simulation data), which is convenient for subsequent comparative analysis and sensitivity evaluation.

[0027] The custom setting of key parameters of the algorithm is supported to meet the parameter sensitivity test requirements in different evaluation scenarios. For the matched filter (MF), the target spectral template (such as the standard methane spectrum based on the HTRAN database or the measured spectrum), the covariance matrix calculation window size (5x5 to 21x21 pixels), and the like can be adjusted, wherein the selection of the target spectral template directly affects the identification ability of the algorithm to the methane characteristic signal, and the covariance matrix calculation window size affects the suppression effect of the algorithm on the background noise; for the RWL1MF, the re-weighting iteration number (3-10 times), the albedo correction coefficient (0.5-2.0), and the like can be set, the re-weighting iteration number determines the detection accuracy of the algorithm to the sparse methane plume, and the albedo correction coefficient affects the adaptability of the algorithm to different albedo surfaces; for the DOAS, the fitting window (2122-2200 nm, 2200-2485 nm, or the full waveband), the polynomial order (1-5 orders), the reference spectrum (including a mixed spectrum of methane and oxygen absorption), and the like can be flexibly selected, the size and position of the fitting window affect the extraction accuracy of the methane characteristic absorption signal, the polynomial order is used to eliminate low-frequency interference, and the selection of the reference spectrum has an important influence on the inversion accuracy of the algorithm.

[0028] In addition, the module also has a perfect algorithm running log recording function, which records the parameter setting, calculation time (accurate to milliseconds), intermediate results (such as covariance matrix, filter coefficient), and the like of each inversion in detail. These log information is stored in the SQLite database, which can be queried through keywords such as time stamp and algorithm type, provides valuable data support for subsequent algorithm performance tracing and error analysis, and helps to deeply understand the running mechanism and performance bottleneck of the algorithm.

[0029] The sensitivity analysis module is the quantitative analysis center of the evaluation results of the entire system, mainly receives the results output by the inversion algorithm integration module, is used for system analysis of the sensitivity of the inversion algorithm to the surface albedo and land cover type, and comprehensively reveals the performance and change rule of different algorithms under various surface conditions through multi-dimensional and quantitative analysis.

[0030] The module realizes in-depth evaluation of algorithm sensitivity by comparing algorithm inversion results under different surface parameter configurations. The internally integrated confusion matrix analysis tool can automatically calculate basic indicators such as true positives (TP, correctly detected methane pixels), true negatives (TN, correctly identified non-methane pixels), false positives (FP, misjudged non-methane pixels as methane), and false negatives (FN, missed methane pixels) based on the comparison between inversion results and simulated true values. On this basis, further derivation of comprehensive performance indicators such as true positive rate (TPR = TP / (TP+FN), measuring detection integrity), true negative rate (TNR = TN / (TN+FP), measuring anti-interference ability), precision (PPV = TP / (TP+FP), measuring detection accuracy), recall (REC = TP / (TP+FN), consistent with TPR), and F1 score (F1 = 2×(PPV×REC) / (PPV+REC), comprehensive measure of detection performance) is carried out. These indicators reflect the detection ability, anti-interference ability, and overall performance of the algorithm from different angles, and can comprehensively evaluate the pros and cons of the algorithm.

[0031] With powerful multi-dimensional analysis function, the performance variation law of the algorithm under different albedos (0.1 vs 0.4), different land cover types (8 types compared one by one), and different emission rates (500 kg / h vs 1000 kg / h) can be evaluated respectively. By drawing the trend curve of performance indicators with surface parameters, such as the curve of F1 score with albedo (which can directly show the change trend of algorithm performance with the increase of albedo), the distribution histogram of true positive rate on different land cover types (which can clearly show the adaptability of the algorithm on different surface types), and the scatter plot of false positive rate and emission rate (which can reveal the difference in anti-interference of the algorithm on different intensity emission sources), the sensitivity difference of the algorithm to each surface factor is directly displayed.

[0032] For example, by analyzing the performance of the algorithm under different albedo conditions, it is found that the F1 score of the MF algorithm decreases by an average of 27% when the albedo increases from 0.1 to 0.4, with the most significant decrease (up to 41%) in the white commercial roof scenario, indicating that the algorithm is sensitive to high albedo surfaces; while the RWL1MF algorithm has an albedo correction mechanism, the F1 score only decreases by 12%, and the sensitivity is significantly lower than MF. By comparing the true positive rate under different land cover types, it can be seen that the DOAS algorithm performs best in water scenarios (true positive rate 89%) and worst in green sports field scenarios (true positive rate 53%), mainly due to the interference of vegetation spectra; the RWL1MF algorithm maintains a relatively stable true positive rate (72%-81%) in various surface scenarios, showing strong surface adaptability. Analysis of different emission rates shows that the F1 score of the three algorithms under 1000 kg / h emission rate is higher than that under 500 kg / h scenario, among which the DOAS algorithm has the largest improvement (23%), indicating that it has stronger detection ability for high concentration methane plume.

[0033] The module can also perform cross-factor sensitivity analysis, such as analyzing the influence of the "albedo x land cover type" interaction on algorithm performance. The results show that in the "high albedo + green sports field" combined scenario, the false positive rate of the MF algorithm is as high as 45%, which is 1.8 times that of single factor influence, indicating that the interaction between factors can exacerbate algorithm performance deterioration. Through this multi-dimensional analysis, an "influence factor matrix" of algorithm sensitivity can be constructed, quantifying the contribution of each factor to algorithm performance (such as albedo contribution 38%, land cover type 42%, emission rate 20%), providing clear priority guidance for algorithm optimization.

[0034] Finally, the module integrates all analysis results to generate a comprehensive sensitivity evaluation report containing data statistics table, trend chart, sensitivity level evaluation. The report divides the sensitivity of the algorithm to each factor into "high sensitivity (performance change rate > 30%) "medium sensitivity (10%-30%) "low sensitivity (<10%) three levels, and provides targeted application suggestions such as "MF algorithm is suitable for low albedo water, rock scenarios, and should be used with caution in high albedo urban areas" "RWL1MF algorithm can be preferentially applied to small and medium-sized emission source monitoring in complex surfaces" etc., providing clear guidance for algorithm scene selection.

[0035] The improved algorithm verification module is the verification platform for algorithm innovation and optimization of the whole system, which is used to verify the performance of the dynamic regularization adaptive matched filter (DRAMF) algorithm, evaluate its advantages and limitations compared with existing mainstream algorithms (sensitivity analysis and evaluation module output results comparison), and provide scientific basis for the popularization and application of the algorithm.

[0036] The DRAMF algorithm verified by the module integrates three core functions of dynamic background updating, iterative reweighted regularization and adaptive matched filtering, and realizes performance improvement by integrating the advantages of MF and RWL1MF algorithms. Dynamic background updating adopts hierarchical K-means clustering algorithm (cluster number 5-10 classes), and updates the background model every 50 frames of data, which can adapt to the spatial heterogeneity of the surface spectrum in real time; iterative reweighted regularization introduces L1 norm penalty term (regularization parameter 0.01-0.1), which enhances the detection ability of sparse methane plume; adaptive matched filtering dynamically adjusts the filter coefficient according to the local signal-to-noise ratio (adjustment range 0.5-1.5 times), which improves the weak signal extraction capability. The principle diagram of the algorithm is shown in Figure 6 .

[0037] The module can compare the inversion results of DRAMF algorithm with those of three existing mainstream algorithms (MF, RWL1MF and DOAS) in all aspects. Not only the numerical differences of key performance indicators such as F1 score, true positive rate and true negative rate are compared, but also the spatial distribution consistency of inversion concentration and error distribution characteristics are analyzed in detail. In the average performance comparison of 32 simulation scenarios, the F1 score of DRAMF algorithm is 82%, which is 41%, 55% and 15% higher than that of MF (58%), DOAS (53%) and RWL1MF (71%) respectively; the true positive rate is 83%, which is 32% higher than that of MF; the false positive rate is 9%, which is 68% lower than that of DOAS. In terms of spatial distribution, the coincidence degree of methane plume boundary inverted by DRAMF algorithm and simulation true value is 87%, which is higher than that of RWL1MF (76%), indicating that its spatial positioning accuracy is higher; the error distribution is nearly normal distribution, with mean value of 0.02 ppm·m and standard deviation of 0.15 ppm·m, which is better than the error level of other algorithms.

[0038] Supports verification in both simulation data and actual AVIRIS-NG data scenarios. Simulation data verification precisely evaluates the performance improvement of the algorithm under ideal conditions by strictly controlling variables such as fixed emission rate and meteorological conditions, and only changing the surface parameters. For example, in the "high albedo + white commercial roof + 500 kg / h" simulation scenario, the F1 score of DRAMF algorithm is 79%, which is 27% higher than that of RWL1MF (62%), mainly because dynamic background updating effectively eliminates the interference of high reflection background. Actual data verification selects AVIRIS-NG flight data containing known methane emission sources (such as flight route ang20160211t075004 covering Los Angeles oil and gas field area and ang20170909t210217 covering Santa Barbara landfill), and evaluates the application potential of the algorithm in real complex environment. The results show that the recognition accuracy of DRAMF algorithm for known emission sources is 80%, and the correlation coefficient between the inverted concentration and the ground monitoring station data is 0.86, which meets the actual monitoring requirements.

[0039] The module includes a professional result visualization unit, which can generate inversion concentration distribution contrast maps (superimposed on Google Earth images), performance index radar charts (multi-index comprehensive comparison), error frequency histograms (showing error distribution characteristics), and various other visualized results. For example, in the visualization results of the Los Angeles oil and gas field data, the DRAMF algorithm can clearly identify 3 small leakage sources (emission rate of about 400 kg / h).

[0040] Meanwhile, the application also provides a method for evaluating the sensitivity of airborne methane inversion algorithm by using hyperspectral simulation, which specifically comprises the following steps:

[0041] Step 1: Generate hyperspectral data containing different methane emission rates through a hyperspectral data simulation module. First, simulate the three-dimensional distribution of the methane plume by using the WRF-LES model, and set the emission rates to be 500 kg / h and 1000 kg / h, respectively. The simulation time step is dynamically adjusted according to the characteristics of the turbulent flow. When the turbulent flow is intense (such as the initial 0-2 hours of methane emission, the turbulent intensity is greater than 0.8 m / s), a smaller time step (1-5 seconds) is used to capture the rapid changes of the plume. When the turbulent flow is stable (after 2 hours, the turbulent intensity is less than 0.3 m / s), a larger time step (10-30 seconds) is used to obtain the methane volume mixing ratio enhancement data at different times. The horizontal range of the simulation area is 3.84 km x 3.84 km (128 x 128 grids), and the vertical direction is divided into 20 layers (0-2 km), ensuring the integrity of the three-dimensional structure of the plume. 2 2 2 2

[0042] Secondly, input the plume data into the MODTRAN radiative transfer model, set the atmospheric model (model 7, mid-latitude summer), the sensor height (elevation 5 km), the methane background concentration 1.8 ppm, and other parameters, simulate the reflected solar radiation spectrum. When calculating the model, consider the atmospheric molecular absorption (H2O absorption at 1400 nm and 1900 nm, O2 absorption at 760 nm), aerosol scattering (visibility 23 km, rural aerosol model), and ground reflection, output the spectrum range 400-2500 nm, and the spectral resolution 0.1 nm.

[0043] Then, based on the sensor response function (Gaussian function, half-peak full width 5 nm) of AVIRIS-NG, the simulated spectrum is convolved and resampled to generate a hyperspectral data cube of 432 bands, with a band center wavelength interval of 5-10 nm, consistent with the band setting of AVIRIS-NG.

[0044] ​​​​Finally, the simulated noise based on signal-to-noise ratio is added, the signal-to-noise ratio of the visible band (400-700 nm) is set to 100-300 (mean 200), the near-infrared band (700-2500 nm) is set to 50-200 (mean 120), the noise type is Gaussian white noise, the noise value is calculated and superimposed into the simulated spectrum, and the hyperspectral data generation is completed.

[0045] Step 2, set multiple surface albedos and land cover types through the surface parameter configuration module. Select 8 typical land cover type reflectance spectra from the spectral library, including interference (white commercial roof, green sports field), water body (lake), rock, non-photosynthetic vegetation (coniferous deciduous), paved surface (airport asphalt), roof (red tile roof), soil (bare soil). Each spectral data is preprocessed: remove noise bands (such as 1350-1450 nm band severely affected by atmospheric absorption), normalize to 0-1 range, and match AVIRIS-NG spectrum (convolution and resampling).

[0046] Using the solar spectrum weighted integral method, the reflectance spectrum of each surface type is adjusted to two target values of albedo 0.1 and 0.4 respectively. When calculating the original albedo, the solar irradiance uses the ASTM G173-03 standard spectrum, and the integral range is 400-2500 nm; when adjusting, through the linear scaling factor k = target albedo / original albedo, the reflectivity R(λ) of each band is modified (R target (λ))=R orug (λ)×k), to ensure that the shape of the adjusted spectrum remains unchanged. After generating 16 kinds of surface spectral data, the complete hyperspectral data set containing different surface scenes is generated by fusing the hyperspectral data generated in step 1 through the formula "radiation spectrum = surface reflectivity × atmospheric downward radiation + atmospheric path radiation".

[0047] Step 3, call the MF, RWL1MF, DOAS algorithms in the inversion algorithm integration module to perform methane concentration inversion. For each surface scene data set generated in step 2, run the three algorithms respectively:

[0048] For the MF algorithm, the target spectral template is set to the methane absorption spectrum based on the HITRAN database (2122-2485 nm), the covariance matrix calculation window is set to 11x11 pixels, and the filter value (s is the target spectrum, μb, Cb are the background mean and covariance matrix respectively, r is the pixel spectrum) is calculated through the formula "MF = (s-μb) T Cb -1 (r-μb)", and the filter value is converted to the methane column concentration enhancement.

[0049] For RWL1MF algorithm, the number of reweighting iterations is set to 5, the albedo correction factor is set to 1.2, the initial weight is set to 1, and the weight is updated by "weight = 1 / (|concentration|+e)"(e = 1e-6) in each iteration. The target spectrum is adjusted by the albedo correction factor(based on the surface albedo) to output the sparse concentration distribution finally.

[0050] For DOAS algorithm, the inversion window is selected as 2122-2485nm, the background spectrum is fitted by a 3rd order polynomial, and the reference spectrum contains methane(CH4) and oxygen(O2) absorption. The methane column concentration enhancement X is solved by least square fitting "r(λ) = I0(λ)exp(-σ_CH4(λ)×X-σ_O2(λ)×X_O2)+P(λ)+noise"(I0 is incident light, σ is absorption cross section, X is column concentration, P is polynomial).

[0051] The inversion results of the three algorithms are stored as a two-dimensional concentration distribution image(128x128 pixels, spatial resolution 30x30m), and the concentration value unit is ppm·m.

[0052] Step 4, Calculate the performance indicators of each algorithm and analyze the sensitivity through the sensitivity analysis module. Take the simulated methane plume distribution as the true value(pixels with concentration>0 are methane area, otherwise non-methane area), and perform threshold segmentation(MF, RWL1MF take filtered value>3σ, DOAS take concentration>0.5ppm·m) on the inversion results to determine the detection area.

[0053] Calculate the basic indicators TP(True Positive, true positive, the number of samples that are actually positive and are correctly classified as positive), TN(True Negative, true negative, the number of samples that are actually negative and are correctly classified as negative), FP(False Positive, false positive, the number of samples that are actually negative and are incorrectly classified as positive(also known as Type I error)), FN(False Negative, false negative, the number of samples that are actually positive and are incorrectly classified as negative) and derived indicators(F1, etc.) of each algorithm under different ground surface scenarios, and store them in CSV files, including scene ID, algorithm type, ground type, albedo, emission rate, 12 indicator values, etc.

[0054] Analyze the variation of the indicators with the surface albedo, calculate the "index difference between high albedo(0.4) and low albedo(0.1) scenes / low albedo scene index value" as the sensitivity coefficient, and the larger the absolute value of the coefficient, the higher the sensitivity. The results show that the albedo sensitivity coefficient of MF is -0.38(F1 score), that of RWL1MF is -0.15, and that of DOAS is -0.22, indicating that MF is most sensitive to albedo.

[0055] The distribution characteristics of the analysis indicators on the eight land cover types were analyzed, and the standard deviation of each algorithm indicator was calculated (reflecting stability). The F1 score standard deviation of DOAS was 0.18, MF was 0.15, and RWL1MF was 0.09, indicating that RWL1MF had the best stability on different land surface types.

[0056] The sensitivity changes under different emission rates were analyzed, and the "F1 score difference between 1000kg / h and 500kg / h scenarios" was calculated. The DOAS difference was 0.17, the MF was 0.12, and the RWL1MF was 0.08, indicating that DOAS had the highest sensitivity to emission rates.

[0057] Based on the above analysis, a sensitivity evaluation matrix of the algorithm was generated to clearly identify the strengths and weaknesses of each algorithm in different scenarios.

[0058] Step 5, test the performance of DRAMF algorithm and conduct comparative evaluation using improved algorithm verification module. Run DRAMF algorithm on all land surface scenario data sets generated in step 2, set dynamic background update interval to 50 frames, cluster number to 8, iteration reweighting times to 8, regularization parameter to 0.05, and adaptive filtering coefficient range to 0.8-1.5.

[0059] Calculate the F1 score, true positive rate, and true negative rate of DRAMF algorithm, and compare it with MF, RWL1MF, and DOAS algorithm. Calculate the performance improvement percentage by "(DRAMF indicator - comparison algorithm indicator) / comparison algorithm indicator x 100%". In 64 scenarios, the F1 score of DRAMF algorithm improved by an average of 30%, with the largest improvement (58%) in the "high albedo + green sports field + 500kg / h" scenario.

[0060] Verification on actual AVIRIS-NG data, select flight routes ang20160211t075004 (Los Angeles oil field) and ang20170909t210217 (Santa Barbara landfill), which contain known methane emission sources (verified by ground). After running the four algorithms, compare the inversion results with ground monitoring data (such as Los Angeles Basin ground station concentration data).

[0061] Generate inversion concentration distribution comparison chart (superimposed on Google Earth image), performance index radar chart (including F1, TPR, TNR, and other 6 indicators), error frequency histogram, and other visual results. Comprehensive simulation data and actual data verification results, evaluate the advantages (such as complex surface adaptability, low concentration detection ability) and limitations (such as about 20% increase in calculation time) of DRAMF algorithm in different scenarios.

[0062] (Three) Effective benefits

[0063] The present application realizes the system quantitative analysis of the sensitivity of the airborne methane inversion algorithm by constructing a hyperspectral-based methane inversion algorithm sensitivity evaluation system and method. Compared with the existing evaluation method, the present application has the following remarkable effects:

[0064] (1) Improve evaluation accuracy: through the full-factor design of 32 controllable scenes, the influence of surface albedo, land cover type and emission rate on the algorithm is separated, and the sensitivity analysis accuracy is 1%, solving the problem of mixed factors in real data.

[0065] (2) Enhance algorithm optimization pertinence: through quantitative indicators (such as F1 score, true positive rate) to locate algorithm bottlenecks, for example, it is found that the false negative rate of MF in high albedo scene is more than 30%, which provides a clear direction for improvement, and the comprehensive performance of DRAMF algorithm is improved by 30%-55% compared with traditional algorithm.

[0066] (3) Improve monitoring reliability: based on the evaluation results, the "scene-algorithm" matching strategy is formulated, which makes the missed detection rate of actual methane emission monitoring reduced to below 5%, and the concentration inversion error is reduced to ±12%, which is significantly better than the prior art.

[0067] (4) Expand application universality: the system supports customizing surface types, algorithm parameters and evaluation indicators, and can be extended to the inversion algorithm evaluation of other trace gases such as carbon dioxide and nitrogen oxides, and is compatible with AVIRIS-NG, PRISMA and other types of hyperspectral sensor data. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 It is a schematic diagram of the module composition of the present application system, which shows the connection relationship and data flow of the hyperspectral data simulation module, the surface parameter configuration module, the inversion algorithm integration module, the sensitivity analysis module and the improved algorithm verification module.

[0069] Figure 2 It is a flowchart of the method of the present application, which clearly presents the complete steps from hyperspectral data generation, surface parameter configuration, algorithm inversion, sensitivity analysis to improved algorithm verification.

[0070] Figure 3 It is a schematic diagram of image synthesis process, which shows the whole process of data simulation.

[0071] Figure 4 It is a comparison chart of inversion results of different algorithms under simulated plume samples, including MF, RWL1MF, DOAS, etc.

[0072] Figure 5The F1 scores of different algorithms are compared in the box plot of different surface albedos to analyze the F1 scores of the algorithms at different albedo levels.

[0073] Figure 6 The schematic diagram of the DRAMF algorithm. DETAILED DESCRIPTION

[0074] The application will be further described in detail below in combination with the drawings and specific embodiments.

[0075] This embodiment takes the evaluation of the sensitivity of three mainstream algorithms in urban area methane monitoring as an example to show the specific application process of the application.

[0076] (I) Data preparation

[0077] 1. Hyperspectral data simulation: Simulate the methane plume by using the WRF-LES model, set the emission rate to 500 kg / h (simulate small oil and gas leakage), simulate for 2 hours, time step 5 seconds, generate 128x128x20 (horizontalxvertical) three-dimensional plume data. Input the data into the MODTRAN model, set the atmospheric model to 7, the sensor height to 5 km, generate the reflectance spectrum, resample to 432 bands, add noise (SNR = 120 in the near-infrared band), and obtain the hyperspectral data cube.

[0078] 2. Surface parameter configuration: Select typical urban surface types (white commercial roof, red tile roof, airport asphalt), adjust their reflectance spectra to albedo 0.1 and 0.4, and fuse them with the hyperspectral data to generate 6 scene data sets (3 surfaces x 2 albedos).

[0079] (II) Algorithm inversion

[0080] Call the inversion algorithm integration module, run MF, RWL1MF, and DOAS algorithms on the 6 scene data sets, set the uniform inversion band range (2122-2485 nm), and record the key parameters and running results of each algorithm.

[0081] For the MF algorithm, the target spectrum template uses the methane absorption cross section based on the HITRAN database version 2020 (temperature 296 K, pressure 1 atm), and the covariance matrix calculation window is set to 11x11 pixels to balance noise suppression and spatial resolution. In the white commercial roof (albedo 0.4) scene, the algorithm runs for about 2.3 seconds per scene, and the output methane column concentration enhancement ranges from 0 to 5.2 ppm·m, with high values concentrated in the center of the plume, but there is obvious blur phenomenon in the edge.

[0082] The RWL1MF algorithm sets the number of reweighting iterations to 5 times, and the albedo correction coefficient is dynamically adjusted according to the ground surface albedo (1.3 for high albedo scenes and 1.0 for low albedo scenes). In the red tile roof (albedo 0.1) scene, the weight matrix gradually focuses on the plume area during the iteration process, and the final output concentration distribution is significantly sparse, with the concentration value in the non-plume area being mostly below 0.3 ppm·m. The running time is about 4.1 seconds per scene, which is 78% higher than the MF algorithm, but the edge positioning accuracy is improved by about 15%.

[0083] The DOAS algorithm selects a 3rd order polynomial to fit the background spectrum, and the reference spectrum contains the absorption cross sections of methane (CH4), oxygen (O2) and water vapor (H2O). The fitting window is divided into two sub-windows of 2122-2250 nm and 2250-2485 nm to reduce spectral interference. In the airport asphalt (albedo 0.1) scene, the root mean square of the fitting residual of the background spectrum is 0.008, the inversion concentration range is 0-4.8 ppm·m, the plume center concentration deviation is less than 5% compared with the MF algorithm result, but there is a small amount of oscillation at the edge.

[0084] The inversion results of all algorithms are stored as concentration files in ENVI format, including geographic coordinates (UTM projection, WGS84 datum), concentration value matrix and metadata (algorithm type, parameter setting, running time, etc.), which provides standardized input for subsequent sensitivity analysis.

[0085] (Three) Sensitivity analysis implementation

[0086] The inversion results of the 6 scenes are input into the sensitivity analysis module, and the methane plume distribution simulated by WRF-LES is taken as the true value (pixels with concentration >0.1 ppm·m are defined as methane area). The threshold of the inversion result is set: MF and RWL1MF take the filtering value >3 times the background standard deviation (based on the non-plume area calculation), DOAS takes the concentration >0.5 ppm·m, and the detection area and non-detection area are divided.

[0087] The performance indicators of each algorithm in the 6 scenes are calculated, and the results show that in the high albedo (0.4) scene, the F1 score of the MF algorithm is 52% on average, which is 16 percentage points lower than that of RWL1MF (68%), and the false negative rate reaches 31% (mainly concentrated in the plume edge); the false positive rate of the DOAS algorithm in the green sports field scene is as high as 29%, which is significantly higher than that of other ground types. Through the sensitivity coefficient calculation (the difference in F1 score between high albedo and low albedo scenes / F1 score in low albedo scene), the albedo sensitivity coefficient of MF is -0.35, that of RWL1MF is -0.14, and that of DOAS is -0.21, indicating that MF is most sensitive to albedo change.

[0088] A "surface type-albedo-F1 score" three-dimensional thermal map is generated to intuitively show the change rule of the algorithm performance: the RWL1MF keeps high stability in various scenes (F1 score 62%-73%), the DOAS performs well in water and rock scenes (F1 score more than 70%), but the performance drops sharply in vegetation coverage scenes. Based on this, targeted suggestions are put forward: the RWL1MF algorithm is preferred in urban high-albedo areas, and the DOAS algorithm can be selected in natural water areas to improve the detection efficiency.

[0089] (IV) Improved algorithm verification

[0090] The DRAMF algorithm is run in 6 scene data, the dynamic background update interval is set to 50 frames, the cluster number is 8 classes (adapted to the diversity of urban surfaces), the iteration reweighting number is 8 times, and the regularization parameter is 0.05. Compared with existing algorithms, the F1 score of DRAMF in the white commercial roof (albedo 0.4) scene reaches 79%, which is 11 percentage points higher than RWL1MF, and the false negative rate is reduced to 15%, mainly because the dynamic background update effectively separates the high-reflective surface from the methane signal.

[0091] The actual AVIRIS-NG data (ang20160211t075004) of the Los Angeles oil and gas field is selected for verification, which contains 3 known oil and gas leakage sources (confirmed by ground flux monitoring, emission rate 450-600 kg / h). The DRAMF algorithm successfully identifies all 3 leakage sources, and the correlation coefficient of the inversion concentration and the ground monitoring data is 0.83. The visualization result shows that the plume shape of DRAMF inversion has higher consistency with the actual diffusion direction (affected by wind direction) and clearer edge details.

[0092] Compared with the prior art, the present application has the following obvious advantages:

[0093] 1. Systematicity of evaluation system: the prior art mainly uses single scene or real data to evaluate the algorithm, which cannot separate the independent influence of surface parameters. The present application realizes the full-factor experimental design of "surface type x albedo x emission rate" by constructing 64 controllable simulation scenes, which can quantify the contribution of each parameter to the algorithm performance (such as the contribution of land cover type reaching 42%), and provides a clear direction for algorithm optimization.

[0094] 2. High fidelity of simulation data: the existing simulation technology mainly ignores the influence of atmospheric turbulence on plume, or simplifies the spectral characteristics of the surface. The present application combines WRF-LES and MODTRAN models, and the simulated methane plume contains non-uniform structure of turbulent diffusion (such as eddy, vortex), and the surface spectrum is strictly corrected and matched with the albedo, with a correlation of 0.92 with the actual surface spectrum, which ensures the authenticity of the evaluation results.

[0095] 3. Comprehensive algorithm verification: Existing technologies only verify improved algorithms in simulation data, while the present application verifies the DRAMF algorithm in both simulation data (64 scenarios) and actual AVIRIS-NG data (2 flight lines), comprehensively evaluating its performance from theory to real environment, with verification indicators covering F1 score, spatial positioning accuracy, concentration error, etc. 8 items, and the results are more convincing.

[0096] 4. Application guidance pertinence: The evaluation reports of existing technologies are mostly qualitative descriptions, while the present application provides sensitivity level division (high / medium / low sensitivity) and "scene-algorithm" matching suggestions, such as "high albedo urban areas prefer DRAMF or RWL1MF", directly guiding actual monitoring work, and making the efficiency of algorithm selection improve by more than 60%.

[0097] Application prospect of the invention

[0098] The present application has broad application prospects in the fields of atmospheric environment monitoring, climate change research, energy industry supervision, etc.:

[0099] In the field of environmental monitoring, it can be used as a "standard evaluation tool" for methane inversion algorithms, providing scientific basis for environmental protection departments to select appropriate monitoring algorithms. For example, for urban landfills (mostly high albedo bare soil and vegetation mixed ground), the DRAMF algorithm is recommended, which can improve the emission source identification accuracy from 65% in existing technologies to more than 80%.

[0100] In climate change research, by quantifying the sensitivity of different algorithms in global typical land surface types (such as tropical rainforest, desert, city), the error of existing methane emission inventory can be corrected (estimated to be ±30% due to algorithm bias), and the accuracy of climate model simulation of methane warming effect can be improved.

[0101] In the energy industry supervision, it can guide the design of airborne monitoring schemes for oil and gas fields, coal mines, etc. For example, for high albedo oilfield facility areas, using the DRAMF algorithm combined with an 11x11 pixel inversion window can improve the leak detection efficiency by 40%, and reduce the false negative rate to less than 5%, helping enterprises achieve precise emission reduction.

[0102] In the future, by expanding the land surface type library (such as glaciers, wetlands, etc. Special topography), increasing the sensitivity evaluation of atmospheric parameters (such as aerosol types, cloud cover), the universality of the system can be further improved, and core technical support can be provided for the construction of global methane monitoring network.

[0103] The above is further detailed description of the present application in combination with specific embodiments, and cannot be deemed as limitation of the specific embodiments of the present application. For those skilled in the art of the present application, several simple optimizations or changes can be made without departing from the concept of the present application, and all of them shall be deemed as falling within the protection scope of the present application.

Claims

1. A sensitivity assessment system for methane inversion algorithm based on airborne hyperspectral imaging, characterized in that, It includes a hyperspectral data simulation module, a surface parameter configuration module, an inversion algorithm integration module, a sensitivity analysis module, and an improved algorithm verification module; The hyperspectral data simulation module is used to simulate methane-related hyperspectral data acquired by an airborne hyperspectral imager. It deeply integrates large eddy simulation (LES) models from weather research and forecasting models, generating three-dimensional methane plume distribution data with spatiotemporal dynamics at a horizontal resolution of 30×30m and a 128×128 grid. This can be achieved by setting a value of 100W / m³. 2 The uniform sensible heat flux drives buoyancy turbulence, which, combined with mechanical turbulence generated by surface drag, accurately simulates the methane diffusion process under different meteorological conditions. Simultaneously, it integrates the MODTRAN 5.4 radiative transfer model, calculating the bidirectional transmittance of the atmosphere below the sensor according to preset atmospheric models, atmospheric path types, and sensor height parameters, simulating the reflected solar radiation spectrum. It internally stores a rich library of sensor response functions, covering Gaussian convolution parameters across different spectral bands. It can rigorously perform convolution and resampling processing on the simulated spectrum according to the next-generation technical specifications of airborne visible / infrared imaging spectrometers, ultimately generating a hyperspectral data cube with 432 bands and a full width at half maximum (FWHM) of approximately 5 nm. Furthermore, the radiation spectrum of each pixel includes simulated noise added based on the signal-to-noise ratio, more closely resembling the characteristics of actual observational data. It supports flexible setting of simulation scenarios with different methane emission rates, adjusting the emission rate by scaling the total methane mass in the plume. The surface parameter configuration module is used to configure different surface albedo and land cover type parameters. It is a key component for controlling surface factor variables and provides refined surface feature input for the hyperspectral data simulation module. It includes a systematically organized spectral library storage unit, storing reflectance spectral data for eight typical surface types, specifically covering disturbances, water bodies, rocks, non-photosynthetic vegetation, paved surfaces, roofs, and soil types. This spectral data originates from the Southern California Natural and Urban Environmental Spectral Library and has been matched with AVIRIS-NG spectral characteristics. It features high-precision albedo adjustment capabilities. Its core principle is to achieve precise control of target albedo values ​​(0.1, 0.4) by globally scaling reflectance and combining it with the weighted integral of the solar spectrum. Specifically, this is achieved through the formula "Albedo = The calculation is performed, where R(λ) is the reflectivity at band λ and E(λ) is the solar irradiance at band λ. The integration range covers 400-2500 nm, and is expressed through "R". target (λ)=R orig The linear scaling formula (λ) × target albedo / original albedo enables precise adjustment of reflectivity, R. orig (λ) is the linear scaling formula for the original reflectance; the surface parameter configuration module can perform full combination configuration of different surface types and albedo to generate a surface scene parameter set containing 32 combinations; The inversion algorithm integration module integrates three mainstream methane inversion algorithms: matched filter, albedo-corrected reweighted matched filter, and differential optical absorption spectroscopy. Each algorithm has a unified input / output interface. The input parameters include a hyperspectral data cube and the characteristic wavelength range of methane absorption. The output results are presented as a two-dimensional distribution image of the methane column concentration enhancement. Custom settings for key algorithm parameters are supported. For the matched filter, the target spectral template and covariance matrix calculation window size can be adjusted. For the albedo-corrected reweighted matched filter, the number of reweighting iterations and albedo correction coefficients can be set. For differential optical absorption spectroscopy, the fitting window, polynomial order, and reference spectrum can be flexibly selected to meet the parameter sensitivity testing needs of different evaluation scenarios. The module also features an algorithm operation log recording function, which records detailed information such as parameter settings, calculation time, and intermediate results for each inversion, providing a basis for subsequent algorithm performance tracking and error analysis. The sensitivity analysis module is used to systematically analyze the sensitivity of inversion algorithms to surface albedo and land cover type. By comparing the algorithm inversion results under different surface parameter configuration modules, it achieves a multi-dimensional and quantitative assessment of algorithm sensitivity. The sensitivity analysis module integrates a fully functional confusion matrix analysis tool, which can automatically calculate the basic indicators of true positive, true negative, false positive, and false negative based on the comparison between the inversion results and the simulated true values, and further derive the true positive rate, true negative rate, precision, recall rate, and F1 score. It can evaluate the performance variation patterns of the algorithm under different albedo, different land cover types, and different emission rates. It can generate trend curves of performance indicators as a function of surface parameters for each algorithm. It integrates all analysis results to generate a comprehensive sensitivity assessment report that includes data statistics tables, trend charts, and sensitivity level assessments, providing clear guidance for the selection of applicable scenarios for the algorithm. The improved algorithm verification module is used to verify the performance of the dynamic regularized adaptive matched filter algorithm, serving as a verification platform for algorithm innovation and optimization. The verified dynamic regularized adaptive matched filter algorithm integrates three core functions: dynamic background update, iterative reweighted regularization, and adaptive matched filtering. Performance improvement is achieved by fusing the advantages of matched filters and RWL1 matched filters. A comprehensive comparison is made between the dynamic regularized adaptive matched filter algorithm and the inversion results of matched filters, albedo-corrected reweighted matched filters, and differential optical absorption spectra. This comparison not only examines the numerical differences in key performance indicators such as F1 score, true positive rate, and true negative rate, but also analyzes details such as the spatial distribution consistency of inverted concentrations and error distribution characteristics. Verification is supported in both simulated and actual AVIRIS-NG data scenarios to evaluate the performance improvement under ideal conditions. Actual data verification uses AVIRIS-NG flight data containing known methane emission sources to assess the algorithm's application potential in real, complex environments. A professional results visualization unit is included, capable of generating various visualization results such as inversion concentration distribution comparison charts, performance indicator radar charts, and error frequency histograms, clearly and intuitively demonstrating the advantages and limitations of the improved algorithm in different scenarios.

2. A sensitivity assessment method for an airborne methane inversion algorithm based on hyperspectral simulation, characterized in that, The evaluation system described in claim 1 is used to implement this, specifically including the following steps: Step 1: Generate hyperspectral data containing different methane emission rates using the hyperspectral data simulation module; First, the three-dimensional distribution of methane plume was simulated using the WRF-LES model, with emission rates set at 500 kg / h and 1000 kg / h. The simulation time step was dynamically adjusted according to the turbulence variation characteristics to obtain enhanced methane volume mixing ratio data at different times. Secondly, the plume data is input into the MODTRAN radiative transfer model, and parameters such as atmospheric model and sensor height are set to simulate the reflected solar radiation spectrum; Then, based on the sensor response function of AVIRIS-NG, the simulated spectrum is convolved and resampled to generate a hyperspectral data cube with 432 bands; Finally, simulated noise based on the signal-to-noise ratio is added to complete the generation of hyperspectral data; Step 2: Configure various surface albedo and land cover types using the surface parameter configuration module; The reflectance spectra of eight typical land cover types were selected from the spectral library, including disturbances, water bodies, rocks, non-photosynthetic vegetation, paved surfaces, roofs, and soil. The reflectance spectra of each land cover type were adjusted to two target values ​​of albedo of 0.1 and 0.4 using the solar spectral weighted integral method, generating 16 types of land cover spectral data. The above reflectance spectral data is fused with the hyperspectral data generated in step 1 to generate a complete hyperspectral dataset containing different surface scenes; Step 3: Use the matched filter, RWL1 matched filter, and differential optical absorption spectroscopy algorithm in the inversion algorithm integration module to perform methane concentration inversion; For each type of terrain scene dataset generated in step 2, run three algorithms respectively: For the matched filter algorithm, the target spectrum is set as the absorption characteristic spectrum of methane in the 2122-2485nm band. The covariance matrix is ​​calculated and the least squares solution is obtained. For the albedo correction reweighted matched filter algorithm, the number of reweighting iterations is set to 5, and the target spectrum is corrected by the pixel albedo factor; For the differential optical absorption spectroscopy algorithm, 2122-2485nm was selected as the inversion window. A third-order polynomial fitting was used to eliminate low-frequency interference, and the methane concentration was calculated by least-squares fitting. The inversion results of the three algorithms were uniformly stored as a two-dimensional concentration distribution image. Step 4: Calculate the performance indicators of each algorithm and analyze its sensitivity using the sensitivity analysis module; based on the inversion results and simulated true values, calculate the true positive rate, true negative rate, F1 score, and other indicators for each algorithm under different land surface scenarios; analyze the variation of indicators with land surface albedo, and compare the performance differences under high albedo (0.4) and low albedo (0.1) scenarios; analyze the distribution characteristics of indicators on 8 land cover types, and identify the land surface types with the best and worst algorithm performance; analyze the changing trends of algorithm sensitivity under different emission rates (500 kg / h and 1000 kg / h); based on the above analysis, summarize the sensitivity characteristics of each algorithm to land surface parameters. Step 5: Test and compare the performance of the albedo-corrected reweighted matched filter algorithm using the improved algorithm verification module. Run the albedo-corrected reweighted matched filter algorithm on all surface scene datasets generated in Step 2. This algorithm dynamically updates the background model through hierarchical background estimation, sets the number of iterations to 10, and dynamically adjusts the regularization parameter during the iteration process. Calculate the F1 score, true positive rate, and true negative rate of the albedo-corrected reweighted matched filter algorithm and compare them with the corresponding indicators of the matched filter, RWL1 matched filter, and differential optical absorption spectroscopy algorithm. Verify the performance of the albedo-corrected reweighted matched filter algorithm on actual AVIRIS-NG data and compare its consistency with other algorithms in the detection of real emission sources. Based on the verification results of simulated data and actual data, evaluate the improvement effect and application value of the albedo-corrected reweighted matched filter algorithm.

3. The evaluation method according to claim 2, characterized in that, In the hyperspectral data simulation process, the calculation of the optical thickness of the methane plume is a crucial step in ensuring the accuracy of the simulation data; the calculation is strictly based on the HITRAN absorption cross-section database and is achieved through the following steps: First, the absorption cross section data (σ) of methane in the 2122-2485nm band were extracted from the HITRAN database; Secondly, for each grid point of the WRF-LES simulation, i = 1-72, the vertical column density of dry air below 5 km in the methane volume mixing ratio enhanced ΔVMR and MERRA-2 meteorological reanalysis data were obtained; Then, the optical thickness at each wavelength λ is calculated using the formula τ(λ)=∑ΔVMR×VCD×σ; Finally, the plume transmittance was calculated based on Beer's Law.

4. The evaluation method according to claim 2, characterized in that, The surface albedo adjustment employs a scientific spectral scaling method to ensure precise control of the albedo value; The core formula of this method is Where R(λ) is the reflectivity at band λ, and E(λ) is the solar irradiance at band λ. The integration range covers the visible-near infrared region of 400-2500nm. The specific adjustment process is as follows: First, calculate the original reflectance spectrum R. orig The original albedo value (λ) is used; then, the global scaling factor is determined based on the ratio of the target albedo value (0.1 or 0.4) to the original albedo value; finally, the scaling factor is calculated using the formula "R". target (λ)=R orig The original reflectance spectrum is linearly scaled by "(λ)×target albedo / original albedo" to obtain the reflectance spectrum corresponding to the target albedo.

5. The evaluation method according to claim 2, characterized in that, The matched filter algorithm in the inversion algorithm integration module is based on the principle of statistical pattern recognition and achieves concentration inversion by enhancing the methane spectral signal. The core process of this algorithm is to model the radiation spectrum measured by the sensor as a function of methane concentration and target spectrum. Specifically, this includes: First, assuming that the surface albedo has spectral smoothness, the radiation measured by the sensor L(α,s) is expressed as a function of the unenhanced ambient radiation L0, the methane concentration enhancement α, and the target spectrum t, i.e., L(α,s) = L0e^(-α / t). -αs ; Then, the function is linearized using a first-order Taylor series expansion to obtain L0e. -αs =L0-αt s Approximate relationship of (L0); Finally, a least-squares optimization problem is constructed to estimate the increase in methane column concentration. Where L i For the observed spectrum, μ is the average sensor radiation approximation L0, and C is the covariance matrix of the background spectrum; this algorithm utilizes spectral covariance information to enhance the methane signal.

6. The evaluation method according to claim 2, characterized in that, The albedo-corrected reweighted matched filter algorithm improves the performance of traditional matched filter algorithms on complex surfaces by introducing albedo correction and sparse regularization mechanisms. Specifically, this includes: First, calculating the albedo factor r for each pixel. i Through formula Implementation, where L i Let be the radiation spectrum of the i-th pixel, and μ be the spectral mean. Then, the target spectrum is scaled using the albedo factor to compensate for the weakening of methane absorption signals under low albedo surfaces; Next, a reweighted L1 regularization method is adopted to enhance the algorithm's ability to detect sparse methane plumes by iteratively solving the L1 regularization optimization problem. Finally, by combining the original matched filter results, the sparse solution, and the albedo-corrected solution, the optimal methane concentration enhancement solution is obtained.

7. The evaluation method according to claim 2, characterized in that, The differential optical absorption spectroscopy algorithm achieves methane concentration inversion by separating high-frequency absorption features and low-frequency interference signals in the spectrum, and performs calculations within the methane inversion window of 2122-2485 nm; the core principle is to use the gas absorption cross section σ j (λ) is decomposed into a low-frequency component mainly caused by Rayleigh scattering and Mie scattering. and the rapidly changing portion σ′ mainly caused by molecular absorption characteristics j (λ); Then, the low-frequency influence part of the radiative transfer equation is approximated by a polynomial as follows: Where I0 is the incident radiation, L is the optical path length, and c j Let ε be the gas concentration. R and ε M These represent the extinction effects of Rayleigh scattering and Mie scattering, respectively, and A(λ) represents the system and turbulence effects. Finally, the methane concentration was calculated by extracting high-frequency absorption features from the measured spectrum using the least squares fitting method.

8. The evaluation method according to claim 2, characterized in that, The performance metrics of the sensitivity analysis module are calculated based on the confusion matrix theory, and the algorithm performance is comprehensively evaluated through multi-dimensional metrics. First, let's clarify the definitions of the four basic indicators: a true positive refers to a pixel that is actually a methane plume and is correctly detected; True negative refers to pixels that are actually non-methane plumes and are correctly identified; A false positive refers to a pixel that is actually a non-methane plume but is mistakenly identified as a plume. A false negative refers to a pixel that is actually a methane plume but was not detected. Based on these four basic metrics, a series of derived metrics are calculated: precision and recall.

Images